Hi friends 👋,

Happy Friday and welcome to our 188th Weekly Dose of Optimism!

I keep saying this, and then it gets cold again, but I think this is the one: it’s a beautiful morning here in New York City, spring is springing, and our fellow humans keep pumping out optimism-worthy work.

This week, for example, we have new funding for AI-assisted Alzheimer’s research and, in Science Breakthroughs, a peptide that degrades the amyloid-β linked with the disease. That, and much much more, in this week’s Dose.

Let’s get to it.

Today’s Not Boring is brought to you by… Silicon Valley Bank

There is a lot of noise out there about how AI is impacting companies. Of course, big AI lab CEOs are predicting massive job disruption and productivity increases. But to find out what’s really going on, ask the people who control where the dollars flow: the CFOs.

SVB surveyed 200+ startup finance leaders for its new State of the VC-Backed CFO report, and a few things jumped out:

• AI adoption is now the #1 issue for startups, and 63% of CFOs rank it top-two

• Companies are doubling AI spend this year (to ~$50K median)

• Over half of CFOs surveyed are already seeing real ROI on AI spend

• The biggest impact to staffing isn’t layoffs, it’s hiring fewer junior hires

In short, startups are getting leaner, faster, and more experimental. If you want to understand where company-building is going next from the people who control the spend, grab your free copy today.

Read SVB’s State of the VC-Backed CFO

(1) OpenAI Foundation Commits $100M to Cure Alzheimer’s

My grandmother had Alzheimer’s. One of my biggest fears in life is that me and the people I love will get it; it’s one of the reasons I started writing, to help me remember. Unfortunately, since my grandmother had it, we’ve made no progress on defeating Alzheimer’s. Measured by age-standardized mortality rate, while we’ve made strides against cardiovascular disease, cancer, and infectious diseases, Alzeheimer’s has only gotten worse. It is a complex disease with hundreds of environmental and genetic risk factors interacting across cell types over decades, and as such, treating it has been a sad game of whack-a-mole.

Maybe AI can help. This week, the OpenAI Foundation, the nonprofit parent of OpenAI, now sitting on a $25 billion life sciences and resilience commitment, announced it’s putting $100M+ into six research institutions to throw modern AI at the problem. Jacob Trefethen, who ran half a billion dollars of science grantmaking at Coefficient Giving before joining OpenAI, is leading the effort. The grants span AI-assisted drug design, biomarker discovery, mapping disease pathways, and personalizing treatment.

The most interesting grantee is Arc Institute, whose progress we’ve covered many times in the dose, which announced a parallel partnership focused on what co-founder Patrick Hsu calls an “AI lab-in-the-loop” approach. The idea is to systematically perturb human brain organoids guided by patient data, measure what happens, feed the results back into AI models, and iteratively build a causal map of the disease, the actual chain of cause and effect from genetics to protein misfolding to synaptic collapse. Arc executive director Silvana Konermann describes the goal as finding “perturbations that can click and drag a cell from a diseased state back into a healthy one.” The other grantees include David Baker’s Institute for Protein Design and EvE Bio.

NIH funds Alzheimer’s, and has for decades, and we are where we are. What Arc and the OpenAI Foundation are doing is what private capital and motivated foundations can do that institutional science usually can’t: pick a hard problem, fund a full-stack experimental-and-AI engine, and run the loop fast enough that we might actually get somewhere by the time it matters to my family.

I would very much like to forget about Alzheimer’s once and for all.

(2) Information to Atoms

Substrate x Google DeepMind

One big lab I have been consistently impressed with is Google DeepMind, and those wacky geniuses are at it again. This week, the American semiconductor foundry startup Substrate, whose launched we kicked off with in Dose #168, shared that GDM’s AlphaEvolve rewrote a chunk of its lithography software and made it nearly seven times faster while cutting compute costs by 97%.

Substrate is the James Proud-founded startup that came out of stealth last October with $100M and a crazily ambitious mission: build a U.S. foundry that uses X-ray lithography to compete with ASML's EUV machines by patterning features tens of atoms wide, on American soil, at lower cost than the Dutch monopoly. AlphaEvolve is DeepMind’s evolutionary coding agent, the latest entry in the Alpha lineage that runs from AlphaGo to AlphaGoZero to AlphaZero to MuZero to AlphaFold to AlphaChip.

This week, Substrate pointed AlphaEvolve at its computational lithography stack, the software that simulates how trillions of photons interact with photoresist to print a chip layer. In just a few weeks, AlphaEvolve explored thousands of algorithmic variations and landed on lossless compression tricks and lower-precision representations that cut memory by 74%, sped up runtime 6.8x, and dropped the Google Cloud TPU bill by 97%. Now, Substrate’s tool can print metal-one layers (the hardest, most defect-prone layer in a chip) at a 24nm pitch in a single exposure, with bidirectional 2D patterns and sharp 90-degree corners. That’s 2nm-node territory, the bleeding edge. Without the speedup, the same patterns would have required multi-patterning: two or three exposures, which means more defects, more cost, and more time.

This is fun because: a) we are rooting for American ASML and b) it’s a combination of the stuff we’ve talked about in the last couple of co-written essays: World Models talked about the evolution of the Alpha line of models, which learn causality instead of rules, and Electromagnetism Secretly Runs the World talked about using AI to design better chips. It’s also an example of America’s best hope in the electric competition with China: using new technologies to find new, better, and cheaper ways to build the things we don’t build today. I suspect we’ll be hearing more from Substrate - it would be insane to compete with ASML at their own game, so they’re creating a new one.

(2a) Arena Physica Releases Heaviside

Speaking of EM modeling… A couple of weeks ago, in Electromagnetism Secretly Runs the World, Arena Physica CEO Pratap Ranade and I mentioned the new models and blog posts they would be releasing soon. Soon has come. Its new model, Heaviside, is its first EM foundation model, and you can play with it in the RF Studio.

(3) Syncere Unveils Lume Lamp That Folds Your Laundry

I have spent a lot of time thinking about robotics recently, and I’ve got to admit, I did not think of a lamp that folds your laundry. This is the future that Beauty and the Beast and The Brave Little Toaster promised.

Lume is a pair of six-foot bedside floor lamps with hidden six-axis robotic arms, cameras, and on-device AI processors. Toss your clean laundry on the bed, the lamps quietly extend their arms and fold it in under two hours for a full load. You don’t have to tell it to do anything, it just watches the bed and figures out when to get to work. (There’s a Lume sitting in the factory somewhere just praying Bryan Johnson orders it.)

The graveyard of laundry-folding robots is packed. Foldimate made you feed clothes in one at a time and quickly… folded. Laundroid was a $16,000 closet-sized monster that bankrupted its parent company in 2019. Laundry is so hard that every robotics company makes demos of their bots figuring out how to do it in a controlled environment, and people go wild, because folding laundry is decidedly a thing we want the robots to do for us.

The why now? is the perception stack. Vision-language models (VLMs) are now good enough that a small startup can train a thing to look at a balled-up t-shirt on a comforter, figure out what it is, grab it, and fold it.

It’s also a constrained problem. It doesn’t promise to do everything; it gets paid to do one thing well, and learn. Maybe it, like Standard Bots, will be able to take small steps across the spectrum. And you gotta buy a lamp anyway.

I want to wait to see the reviews before I drop $2,499 on two lamps, but if you want to be an early adopter, be our guest.

(4) Humans Travel Farther From Earth Than Ever Before

Jared Isaacman for NASA

Last week, we covered the launch of the Artemis II mission, which planned to take North Americans back to the Moon for the first time in more than half a century, and farther from the home planet than we’ve ever traveled (that we know of). As the crew rounded the Moon on Monday, NASA Administrator Jared Isaacman tweeted:

On the far side of the Moon, 252,756 miles away, Reid, Victor, Christina, and Jeremy have now traveled farther from Earth than any humans in history and now begin their journey home. Before they left, they said they hoped this mission would be forgotten, but it will be remembered as the moment people started to believe that America can once again do the near-impossible and change the world.

It’s rare for us to do back-to-back coverage of the same story, but unprecedented accomplishments call for unprecedented measures. Instead of writing about it, I’d just ask you to watch Astronaut Victor Glover’s Easter message back home:

(5) Primer Launches New Brand with The Orator

Primer is one of the not boring capital portfolio companies I’m most proud to be involved with. I’ve written about them here and talked to the CEO Ryan Delk on moving from software into running schools here. They take kids seriously, and exist to give more kids access to excellent education.

This week, they released a video narrated by one of its students and featuring others that highlights both its approach to education, and an optimistic vision for how we can all meet the future: “The next 200 years will look nothing like the last. We will face problems we can’t yet imagine, and possibilities we can barely dream of. We won’t just need answers. We’ll need resilience, flexibility, courage, mastery.”

Amen.

EXTRA DOSE: Science Breakthroughs, Hermeus, Pace Fellowship

Read more

Hey friends 👋,

Happy Friday and Happy Easter Weekend!

We’re going back to the Moon. OpenAI bought TBPN for a lot of money. Somos is going to blanket the globe in fast internet. And a whole lot more good news where that came from.

Let’s get to it.

Today’s Weekly Dose is brought to you by… Guru

Your company has invested in AI tools, which is a great first step. But your AI is only as good as the knowledge behind it. And for most companies, that knowledge is scattered, outdated, and nobody’s job to fix. So your AI gives confidently and sneakily wrong answers to all of your employees.

Guru is the governed knowledge layer that fixes this at the foundation. Companies like Spotify and Brex use Guru to structure their company knowledge, keep it accurate automatically, and make it available to every AI tool and every person, all from one trusted source.

With Guru, when an expert corrects something once, it’s right everywhere you work, in every AI tool, workflow, and answer.

No more garbage in, garbage out. Just cited, verified, always-current knowledge your AI can actually rely on.

Try Guru Today

(1) NASA’s Artemis II Crew Launches to the Moon

NASA

For the first time in 53 years, 5 months, and 20 days, human beings are on their way to the Moon.

Last night at 6:35 p.m. EDT, NASA’s Space Launch System (SLS), standing 322 feet tall, weighing 5.7 million pounds, and generating 8.8 million pounds of thrust (17% more than the Saturn V), lifted off from Pad 39B at Kennedy Space Center carrying four astronauts aboard the Orion spacecraft (which the crew named Integrity). Commander Reid Wiseman, Pilot Victor Glover, and Mission Specialists Christina Koch and Jeremy Hansen are now on an approximately 10-day trip around the Moon and back. No one has done this since Apollo 17 splashed down in December 1972.

Integrity will spend about a day in a high Earth orbit stretching 46,000 miles out, the crew will manually fly the capsule to test its handling, and then a six-minute translunar injection burn will send them on a free-return trajectory using the Moon’s gravity to slingshot them around the far side and back home. On Monday, April 6, the four of them will become the first humans to lay eyes on parts of the lunar far side, reaching a record distance of 252,000 miles from Earth. They’ll splash down in the Pacific on April 10.

Is getting to the Moon more expensive than it should be? Hell yeah, roughly $4.2 billion per SLS flight by the Inspector General’s math. Did it take way longer than planned? Also hell yeah. But four people are about to travel further from our home planet than any human has traveled before, in the first of a series of missions designed to get us back to the Moon to stay.

I can’t really sum up how I feel any better than this kid did:

(2) OpenAI Acquires TBPN

Katie Deighton for WSJ

Speaking of going to the f*cking moon… OpenAI, which just closed a $122 billion round at an $852 billion post-money valuation, has spent some of its treasure to acquire tech media’s crown jewel, TBPN. The FT reported that the deal is in the “low hundreds of millions of dollars.”

Huge congrats to Jordi, John, Dylan, Ben, and the whole TBPN crew.

This is an incredibly smart acquisition for OpenAI. I’ve been saying that the frontier AI labs have completely botched their narrative, and this is about as good of a way to get to work fixing it as I could imagine.

A lot of the commentary so far has been about OpenAI acquiring the media property, which will certainly be helpful, but the big get is Jordi & John. They’re simultaneously highly generative creative geniuses and normal guys. I expect that we’ll see their impact on the way the company communicates to non-San Franciscans pretty quickly, and I hope they’ll have a hand in shaping where AI goes more broadly.

In the meantime, it’s just great to see the good guys get a W.

(3) Google DeepMind Releases Gemma 4 Open Source Model

Completely unrelatedly, I’ve told you guys how awesome I think Google DeepMind is, right? Been saying it for years. What a great lab. Could obviously use a little help telling their story, but wow. What a family of models. AlphaFold! Gemini! AlphaProof! AlphaGo! AlphaZero! Nano Banana! Genie! SIMA! DolphinGemma! I could go on and on.

Yesterday, the good folks at GDM just released a family of AI models that can see images, watch video, understand speech, reason through competition-level math, and write code, all open source. You can download them for free, modify them, and ship them in a commercial product under the Apache 2.0 license.

GDM released Gemma 4 in four sizes: a 31B dense model, a 26B mixture-of-experts model, and two edge models (E2B and E4B) designed to run on phones and laptops. None of them is quite up to par with the best frontier models, including GDM’s Gemini, but since they can run locally on your hardware, it means you get the tokens they produce basically for free.

Yesterday, in Bad Analogies, I wrote that OpenAI and Anthropic were burning a ton of money to compete in an incredibly competitive field, and that Google DeepMind was Zeus in this battle of the gods. Gemma is another lightning bolt. The more things people can do with free tokens, the less they’ll spend on tokens, even if the most advanced capabilities will always come at a price. Of course, closed model customers are paying for more than tokens: reliability, ease-of-use, etc… but it’s fun to see the competition drive down prices for consumers.

Jokes aside, I continue to be very impressed by GDM. The lab came up more than any other in my research for both World Models and Many Small Steps for Robots; they’ve done more to push the frontier in more directions than any lab in the world. With Gemma 4, Demis the Menace strikes again.

(4) Solar cells just did the “impossible” with this 130% breakthrough

Kyushu University via Science Daily

Three weeks back, we led off the Dose with a story on Swift Solar buying a gigawatt-scale solar factory to produce 30-40% efficiency silicon-perovskite tandem solar cells.

This week, researchers at Kyushu University in Japan announced cells with a seemingly impossible 130% efficiency.

There’s a number in solar energy called the Shockley-Queisser limit. It says that a single-junction silicon solar cell can convert, at most, about 33% of the sunlight hitting it into electricity. Silicon-perovskite tandems can up that to 45%. In either case, the rest, especially the high-energy photons, gets wasted as heat. Shockley-Queisser been the ceiling since 1961.

Until now! A team led by Associate Professor Yoichi Sasaki at Kyushu University in Japan, collaborating with Katja Heinze’s group at Johannes Gutenberg University Mainz in Germany, just published a paper in the Journal of the American Chemical Society showing that they paired tetracene-based molecules with a molybdenum-based metal complex (a “spin-flip” emitter) and achieved a quantum yield of roughly 130%. That means for every photon absorbed, the system produced 1.3 energy carriers. More out than in, which seems like it shouldn’t happen.

The trick is a process called singlet fission. When a high-energy photon hits the tetracene, instead of creating one excited electron and dumping the rest as heat, the molecule splits that energy into two lower-energy “triplet” excitons, a little bit like nuclear fission (don’t yell at me real physicists). The problem has always been that a competing process called Förster resonance energy transfer tends to steal the energy back before fission can finish.

Sasaki’s team solved this with the molybdenum complex, which selectively accepts only the multiplied triplet excitons by flipping the spin of its electrons during near-infrared emission. If molybdenum sounds familiar, it’s because it’s the hard-to-pronounce mineral Earth AI found in its first discovery, and because it’s come up a couple of times in recent American Alchemy episodes.

There is an asterisk, for now: the process was demonstrated in solution, not in a working solar panel. It’s a proof of concept. The team’s next step is integrating the tetracene and molybdenum materials together in a solid-state system.

But the result proved that the 100% quantum yield barrier isn’t a physical law, which means the engineers can get to work sorting out the details. If singlet fission can be made to work in commercial cells, we’re going to start putting a lot more photons to good use.

(5) Somos Internet Raises $40M Series B

Alex Konrad for Upstarts Media

In November, I released my longest company Deep Dive ever: Cable Caballero, the story of Forrest Heath III and the vertically integrated internet company he and his band of pirates are building down in Medellín, Colombia.

Explaining how he built what is now the best-performing internet in the region, and maybe the world, with its own fiber network, switches, routers, and ISP, Forrest said:

It’s been this never-ending game of doing something janky, getting credibility, doing crazier stuff, getting more resources, getting smarter people so that we can fix the things that were messed up in the janky past iteration, then gaining credibility to get more resources to get cooler people to do crazier stuff. It’s like this self-sustaining fission process.

Now, Somos has a lot more resources to hire more smart people and do even crazier stuff! This week, the company announced that they raised a $40 million Series B co-led by Bracket Capital and Ribbit Capital. Following on to not boring capital’s original investment was a no-brainer.

Somos is growing fast in Colombia, and it announced that it’s getting ready to expand into Mexico. On TBPN, Forrest even hinted at Somos Caracas…

I think this company is going to build a ton of better, faster, cheaper infrastructure in the coming decades, and they have both the fuel and the backers to do it. If you’ve been looking for a sign to put your stuff in a bag and move to Colombia, this is it.

Vamos Somos!

EXTRA DOSE: Science Breakthroughs, Electric Delivery, Bangalter

Read more

Subscribe now

Hi friends 👋,

Happy Thursday! I wasn’t planning on sending an essay this week. I wanted to give you a little time to breathe after World Models and Electromagnetism.

But then I saw someone make a bad point on the internet, and it was a particular bad point that has always bugged me, so I whipped up a short and sweet rebuttal. In short: not every company that burns a lot of money is the next Amazon. (a bit of a strawman, I admit, except this analogy keeps popping up!)

Let’s get to it.

Today’s Not Boring is brought to you by… Framer

Framer gives designers superpowers.

Framer is the design-first, no-code website builder that lets anyone ship a production-ready site in minutes. Whether you’re starting with a template or a blank canvas, Framer gives you total creative control with no coding required. Add animations, localize with one click, and collaborate in real-time with your whole team. You can even A/B test and track clicks with built-in analytics.

Launch for free at Framer dot com. Use code NOTBORING for a free month on Framer Pro.

Just Publish it With Framer

Bad Analogies

I saw an interaction on twitter the other day that I’m not going to share here because these specific people don’t deserve to be singled out because almost everyone is guilty of something similar.

Basically, one person said, “It’s amazing that OpenAI was able to raise $122 billion when they don’t have a single business that works,” and the other person replied, “Yeah they do they have a bunch of different business lines doing billions of dollars in revenue,” and the OP responded, “Yes but none of them is profitable, they’re losing a lot of money,” and the replier replied, “You could have said the same thing about Amazon!”

Amazon’s success has done a great deal of harm to a great number of companies.

Obviously, it’s done more good. AWS is a miracle. But go with me.

Jeff Bezos is a generational entrepreneur who came from a hedge fund and made a very calculated decision to lose money in the short term if it meant making more of it in the long term. In his 1997 Letter to Shareholders, his first as a public company CEO, he makes the trade-off he’s willing to make clear: “When forced to choose between optimizing the appearance of our GAAP accounting and maximizing the present value of future cash flows, we’ll take the cash flows.”

He spends his entire 2004 Letter to Shareholders walking investors through the logic.

Earnings don’t directly translate into cash flows, and shares are worth only the present value of their future cash flows. Working capital and capital expenditures matter too, and a company can actually destroy shareholder value by growing earnings if the capital investments required exceed the present value of the cash flow derived from them.

Bezos had his hands firmly on the wheel, and a negative working capital engine at his back. There are a million little details that made Bezos’ plan work, all of which, and the connections between them, were taken into consideration. Get to scale category-by-category, starting with the ones best suited to the internet, build an unmatched distribution and logistics network, drive down prices and delivery times, get more scale, grow the distribution network, drive down prices, make the “divinely discontent” customers discontent with anything slower or more expensive than what they’d come to expect with Amazon, get them hooked on Prime, paying a subscription and ordering more, keep pressing the advantage, grow FCF per share. All the while, thanks to negative working capital, growth generated cash, cash funded more infrastructure, infrastructure enabled lower prices and faster delivery, which drove more growth.

We all know the Amazon story to some extent, right? I don’t need to go into more detail than that for now. We could tell a similar story about Facebook: they had no revenue! they didn’t even have a business model!

You could, and many have, use this to justify all manner of sins in your business, in a business in which you’ve invested, or, in the case of the AI labs, in a business that a bunch of weirdos on the internet seem to feel is their duty to defend.

I am perhaps a bit sensitive to this after spending six years pre-not boring in the same industry as WeWork. WeWork lost a lot of money, and so people compared WeWork to Amazon, and maybe there’s an argument to be made that of course you need to spend a lot of money in real estate because you need to do construction and furnish spaces and buy a lot of beer and tequila upfront but then you have this whole network of long-term leases that will generate cashflow for years to come. Bezos wasn’t afraid of spending money today for cashflow tomorrow, why should Neumann be? But we often put in offers on the same spaces as WeWork, and we had these super finely tuned underwriting models, and looking at our underwriting models versus the prices they paid to outbid us on certain spaces it was clear that no matter how optimistic your monthly revenue projections, there was just no way they were going to make money on each space. That was before you took into account the custom builds they were fronting for customers who signed two-year deals (builds that would potentially need to be redone for the next one) or the 100% commissions they were paying brokers for an entire year’s worth of revenue on any customer they brought in. Adam Neumann went on Rick Rubin recently, and it was incredible and made me like him more and even made me bullish on Flow, but even Adam Neumann admitted that he let the economics get away from him in the halcyon days of 2018 and 2019.

If you had justified a WeWork investment to yourself by analogizing to Amazon and repeating “It takes money to make money,” you would have lost it all in the 2023 bankruptcy. Maybe the cleaned up company makes a lot more sense today than the bloated one did then, I haven’t looked at it so I don’t know, but if it does, it’s certainly not because it’s like Amazon.

Of course, analogies can be a useful starting point. “This thing that I might have otherwise dismissed out of hand is probably worth a closer look, because things that have looked this obviously wrong in the past have turned out to look pretty good” is a good way to think.

Uber, too, lost a shitload of money. Uber lost way more money than WeWork. WeWork lost like $2-3 billion in its worst year. Uber burned $9.1 billion in 2022. Uber’s losses made WeWork’s look like rookie numbers.

And yet… Uber is still a publicly traded company, with a market cap hovering around $150 billion.

If you’d looked at Uber and said, “Hey, Amazon lost money too, let me take a closer look,” and then realized that Uber was burning money in actually kind of the same way that Amazon was, to get its product as close to customers as possible in order to improve the customer experience, particularly if you’d looked during the depths of 2022 when it was burning so much cash, you could have made a smart investment.

Then again, using Uber as an analogy – literally, as in Uber-for-X – has kept the startup graveyard operators in business. The one successful company that could be called an Uber-for-X is DoorDash, and it worked explicitly because the team looked at Uber, said “there’s something interesting here,” and then tuned it to fit the situation at hand. Instead of competing in dense urban markets, they took the model to underserved suburbs, where competition was lower, CACs were lower, average order values were higher, retention was higher (fewer alternatives), pickups and dropoffs were easier (stores and homes have driveways and parking lots), and then, per Dan Hockenmaier, they executed relentlessly: “They made deliveries just a little faster and more reliable every week. They scrutinized the quality of every one of their restaurants and dashers. They optimized cost out of the system and gave it back to customers in the form of Dashpass (which launched in 2018 at the start of this chart).”

So are frontier AI labs good or bad businesses? I don’t know. Smarter minds than mine are doing a lot of non-analogical work trying to answer this question. The economy depends on it.

I do know that they’re not definitely bad businesses because they’re losing money.

They are certainly generating a lot of revenue very quickly, and growing at rates that are unprecedented at their scale. Anthropic grew from a $9 billion run rate at the end of 2025 to a $19 billion run rate in February (and rumors are it went much higher in March). In announcing its $122 billion fundraise, OpenAI said, “Within a year of launching ChatGPT, we reached $1B in revenue. By the end of 2024 we were generating $1B per quarter. We are now generating $2B in revenue per month. At this stage, we are growing revenue four times faster than the companies who defined the Internet and mobile eras, including Alphabet and Meta.”

To dismiss those numbers because of a little cash burn would be just as dumb as saying that cash burn is the same as Amazon’s.

On the other hand, it’s worth looking more closely at the burn to understand what you get for it.

Amazon was a high fixed-cost business with variable revenue. The more customers it served, the cheaper each marginal customer got. Each time you ask Claude to build you an app that tells you when to do your laundry, it costs Anthropic money. Each time you ask(ed) Sora to make you a video of Sam Altman twerking, it cost OpenAI money (and scarce compute that could be used for other things), which is why they shut it down. That said, margins at both Anthropic and OpenAI are improving, and with every point in margin, that variable cost problem gets less painful.

One thing to consider is that API tokens are likely a much better business because Anthropic charges based on tokens consumed. Funnily, when I asked Claude about this, it linked me to a public artifact of research someone else had done with Claude, which said that average users of Pro and Max make Anthropic money, Heavy users of Pro and Max lose Anthropic money, and the API business (which is 70-75% of revenue) produces “50–65% gross margins on Sonnet workloads and 35–50% on Opus workloads.” Caveat emptor. This is a monster business, but I wonder how much it benefits from another bad analogy: linesofcodemaxxxxxing.

In the olden days, lines of code (LOC) were expensive to produce. Only a select few even knew how. There were more problems to solve than people or LOCs to solve them, so companies would create a backlog and their engineers would crank through them by producing more LOCs. The lines of code themselves were not the valuable thing; they were valuable insofar as they were used to solve problems the company needed to solve. You could look at a successful software company, see that they’d written a lot of code, and lazily analogize: LOCs are correlated with success, lines of code are themselves a goal. Goodhart’s Law strikes again.

“You gotta be tokenmaxxxxxing,” they’ll tell you. I do wonder if this part of the business slows down once the psychosis breaks and we have maxxxxed all that we can maxxxx. At some point, when you’ve gone too deep, the ghost of Jeff Bezos appears and asks you what any of these tokens you’re maxxxxing has done for your customers.

Anyway, another thing about Amazon’s business is that while they had a ton of direct competitors - Barnes & Noble, eBay, Walmart, a graveyard of dot com competitors - they had no real competition in their strategy. If Bezos was right that getting to scale category-by-category, starting with the ones best suited to the internet, building an unmatched distribution and logistics network, driving down prices and delivery times, getting more scale, growing the distribution network, driving down prices, making the “divinely discontent” customers discontent with anything slower or more expensive than what they’d come to expect with Amazon, getting them hooked on Prime, paying a subscription and ordering more, continuing to press the advantage, and growing FCF per share proved to be the winning strategy, Amazon would emerge on the other side of all that spend as the only company with all of that in place, and with the pricing power and control over the levers of profitability that come with domination.

Uber had direct competition, but it was weaker, and Uber was making a very explicit bet that scale and network effects were all that would matter. They were right, too, and emerged on the other side of all that spend as the only company with the right network in place, and with the pricing power and control over the levers of profitability that come with domination.

Could you say the same for any of the AI labs?

We don’t have to spend a lot of time answering this. There is a lot of direct competition. The strategies seem to be very similar from the outside. Sure, different companies play with targeting different end users, and maybe OpenAI will turn ads into a big business while Anthropic owns the enterprise (although it looks like OpenAI is pushing all of its resources towards competing directly for coding and enterprise), but the core strategy is spend a lot on CapEx now to build increasingly strong models and maybe God is on the other side of all of that spend.

Ask any of the polytheistic religions whether competition among the gods is any less intense. And if we’re playing gods, Google is probably Zeus. Meta is Hermes - god of commerce, god of thieves, trickster god making everyone else’s shit more expensive. Should we do DeepSeek? Sun Wukong, the Monkey God who stole and devoured the peaches of immortality until the Jade Emperor begged Buddha to intervene, and Buddha trapped him under a mountain.

Our competition is multidenominational, and everyone is fighting with basically the same weapons. Every major lab, it seems, prays at the same altar, the Altar of Scaling Laws.

But! You say. But! Recursive self-improvement! Fast takeoff! Those tokens aren’t just producing code — they’re creating a proprietary data flywheel! There will be One True God.

Maybe!

Another reason I am allergic to some of this analogizing and extrapolating is because I have done it myself. I, too, have sinned. I have analogized crypto protocols to companies that didn’t fit just right. I have written I, Exponential and Compounding Crazy. As much as you think you love a good exponential, I can assure you I love them more. So when I ever dare question whether AI might hit a ceiling before it makes us irrelevant meatbags, and am met with the inevitable, “You don’t get exponentials! This is exactly like X! Have you even seen the Tim Urban chart?!” just know that I know these tricks. I have learned their painful lessons.

Not all exponentials are created equal. Does God think forever or does he know all in an instant?

I have been pushing back against Leopold’s assertion that believing ASI will come soon after 2027 “doesn’t require believing in sci-fi; it just requires believing in straight lines on a graph” since he first tweeted it. (to be fair, he’s made billions more dollars than me in the interim.)

Just this week, METR’s Joel Becker wrote a post attempting to provide an intuition for the “Straight lines on graphs” argument. The point is that even if it doesn’t make sense, the line will continue to go up. We’ll figure out ways. But my point is that even if the models can think longer, it doesn’t mean that we will birth God. I think we’re going to end up with really excellent software, which is great, instead of an all-knowing and all-loving AI.

AND, more relevant to the Amazon discussion, if each lab’s model can think longer and get smarter at roughly the same rate, without any clear runaway winner, then the situation is not analogous to Amazon, or even Uber, who ate short-term pain for long-term strategic solitude.

I don’t mean to pick on the AI labs, I don’t even think they themselves are the ones perpetrating this malanalogous thinking, it’s just that they’re the only thing anyone is talking about so they are the richest source of bad analogies. And I really don’t know how it’s going to play out. There is a chance we wake up in a decade and OpenAI and Anthropic are the most valuable companies in the world. There is also a chance that they are the AOLs of this whole thing. There’s a chance that the biggest business in a decade is the one that bought all the cheap tokens after the crash, the way Rockefeller bought all the excess rail capacity worth buying.

AOL and Rockefeller are analogies, too, huh? Avoiding them is hard, because they truly are useful starting points. They’re so easy to reach for, in fact, that on their own, they are at best consensus and at worst dangerous.

It’s the work you do from those starting points that will decide whether you end up the next Hummer Winblad or the next J.P. Morgan. (damnit)

That’s all for today. If you liked what you read, join us in not boring world:

Subscribe now

We’ll be back in your inbox tomorrow with a Weekly Dose.

Thanks for reading,

Packy

Hey friends 👋,

Happy Friday! The sun is shining here in New York City, Duke plays in the Sweet 16 tonight, and there are so many incredible stories this week that I kept having to change what made the cut as new news came out.

Grab a cup of coffee and catch up on all of it in one place. And if you have a second cup, take an hour this weekend to read Electromagnetism Secretly Runs the World.

Let’s get to it.

Today’s Weekly Dose is brought to you by… Quince

Last week, it was cold and rainy in New York. Today, it’s warm and sunny. Aside from the renewed joie de vivre, that means it’s time to switch out the wardrobe.

More and more, when people are thinking about refreshing (and upgrading) their … anything, they ask, “Does Quince make this?” More and more, the answer is yes. Quince is the default starting point for quality, across cashmere, luggage, furniture, fragrance, fine jewelry, and new spring wardrobes.

I have been lucky to be a Quince investor for years. They recently raised at a $10.1 billion valuation. The reason is that they’re able to deliver high-quality across a growing number of categories at surprisingly affordable prices. 100% Grade A Mongolian cashmere. 100% mulberry silk. 100% European linen. 0% brand tax.

They can pull it off because Quince built an M2C (Manufacturer-to-Consumer) operating system with AI demand forecasting, real-time production planning, and direct factory partnerships that cut out distributors, wholesalers, and the retail markup entirely. This week, they even hit #1 in the App Store for shopping. The machine is humming.

Look, I’m biased, but I’m also writing this while wearing Quince jeans and a Quince q-zip. We dress our kids in Quince. And I’m about to do a big spring haul. If you haven’t started shopping at Quince, or haven’t checked out everything they have to offer in a while, join me.

Shop Quince

(1) NASA Goes Nuclear (and to the Moon)

Jared Isaacman’s NASA is shaping up to be as cool as we expected.

On Monday, Isaacman went onstage at the Hill & Valley Forum in Washington for a conversation with Founders Fund / Varda’s Delian Asparouhov and laid out his plans to build a permanent base on the Moon and send a nuclear-powered spacecraft to Mars.

In a normal week, either one would get its own top billing in the Dose, but since JI dropped both at once, let’s Start with the Moon Base. NASA is committing $20 billion over seven years to construct humanity's first permanent outpost near the lunar south pole. “The moon base will not appear overnight,” Isaacman said. “We will invest approximately $20 billion over the next seven years and build it through dozens of missions.” The plan has three phases, with Phase 1 starting now: up to 24 launches and 20+ landings in the next 35 months. That's roughly a mission every six weeks, which sounds like SpaceX cadence.

The first wave will MoonFall surveillance drones to map terrain around the south pole, fission reactors and RTGs for power, and enhanced rovers for scouting. Phase 2 adds pressurized rovers built in partnership with Japan's aerospace agency and semi-habitable modules after Artemis IV. Phase 3 is the full base: permanent infrastructure for sustained human presence. We’re going, and we’re staying. Moon will be a State.

Between NASA and SpaceX, the Moon has been getting a lot of love recently, and Mars has been pushed out. But Isaacman also announced some Mars plans.

SR-1 Freedom will be the first spacecraft to use a nuclear fission reactor for propulsion beyond Earth orbit. A HALEU-fueled reactor delivers 20+ kilowatts of electrical power, driving xenon ion thrusters via a closed Brayton cycle. It launches in December 2028 (likely on a Falcon Heavy) and should arrive at Mars about a year later. Once there, it will deploy three Ingenuity-class helicopters carrying cameras, ground-penetrating radar, and radios to scout human landing sites and hunt for subsurface water. To summarize, we’re sending a nuclear-powered space craft to Mars to release scout helicopters.

SR-1 repurposes the already-built Power and Propulsion Element from the scrapped Gateway program. Rather than shelve expensive hardware when priorities shifted, Isaacman's team pointed it at Mars. The reactor will activate within 48 hours of launch, then it’s off to the Red Planet.

Nuclear is going to be critical to our space ambitions, whether going or staying. We wrote about it in our Deep Dive on Radiant. Fission reactors will power the Moon Base and a fission reactor will propel SR-1 to Mars. Eventually, we’ll have nuclear-powered civilizations as far as the eye can see.

Looks like putting Isaacman in charge of NASA set off a … nuclear chain reaction.

(2) Terraform Industries Breaks Ground on Synth Hydrocarbon Site

Casey Handmer on TBPN

We’re going to need a lot of energy here on earth, too, and even as we add a ton of clean nuclear, solar, and even geothermal, we’re going to need those sweet, sweet hydrocarbons. We just want them to be clean, too.

Good news. Casey Handmer and the Terraform Industries team have been hard at work figuring out how to turn sunlight and air into pipeline-grade natural gas for less than the cost of drilling it. In March 2024, his team in Burbank proved the chemistry works by producing synthetic methane end-to-end for the first time. Now, two years later, Terraform Industries has broken ground on a manufacturing site in Kern County, California, to scale up production.

Kern County is a lucky place: it’s one of California's biggest oil-producing regions and it's also absolutely sun drenched. Those rays are what Terraform wants. The Terraformer system is straightforward in concept: solar electricity powers water electrolysis for hydrogen, direct air capture pulls CO2 from the atmosphere, and a reactor combines them into carbon-neutral methane. What's hard is making it cheap enough to compete with drilling a hole in the ground. Terraform's bet is that plummeting solar costs will get them there, and their qualified electrolyzer stack (under $100/kW) and full-scale reactor suggest they’re heading in the right direction.

Last week, we wrote about one Australian, Chris Power, launching a huge new manufacturing site in Alabama. This week, another Aussie, Handmer, is launching a hydrocarbon manufacturing site in the California desert. Chris and Casey are two of the smartest people I’ve met. With them working on the country’s behalf, it’s a very g’day in America.

(3) Arbor Energy Lands Billion-Dollar Turbine Deal

TechCrunch

But wait… how are we going to turn all of our natural gas into power to feed the data centers? I’m glad you asked.

Arbor Energy, founded by former SpaceX engineers Brad Hartwig and Andres Garcia Clark, just landed a deal with GridMarket to supply up to five gigawatts of its Halcyon turbines, or roughly 200 units, valued in the single-digit billions. They expect the first turbine to connect to the grid in 2028 to be manufacturing 100+ per year by 2030.

Turbines are so hot right now. They’re such a big bottleneck to the data center buildout that their shortage is a big driver behind the rush towards orbital data centers, where they can be powered by the sun. Traditional gas turbines from the big OEMs are backordered until 2032. Hartwig puts it simply: “Everyone wants more power. They wanted it yesterday. The time frames are compressing and the scale is getting larger.”

Arbor takes rocket turbomachinery designed for spaceflight, 3D-prints it into 25-megawatt modular turbines, and sells them to the data centers that needed power yesterday. The turbines run a supercritical CO2 cycle with oxy-combustion, which means zero operating emissions on natural gas, or carbon-negative operation if you feed them biomass waste. They're fuel-agnostic, pre-assembled, and 3D-printed, which is important because it decouples manufacturing from the supply chain bottlenecks (specialized blades, vanes, castings) that plague conventional turbine production or even alternatives proposed by companies like Boom Supersonic.

Our friends at Cantos can’t stop winning.

(4) Centivax raises $37M to advance universal flu vaccine

Katherine Davis for Axios

All of this progress in rockets and energy is sick… but you might not have to be.

Last month, we covered research on a single vaccine to protect against all colds and flus. The company behind that research, Centivax, whose homepage reads “smash the mutants,” just announced that they’ve raised $37 million in follow-on funding led by Structure Fund, with Meiji Seika Pharma, Sigmas Group, Kendall Capital Partners, and Stripe founders Patrick and John Collison joining. The company has now raised $133 million, including from our friend at Amplify Bio.

When we get the flu vaccine every year, we line up for a shot that's been reformulated to match whichever strains scientists think will circulate that season. Sometimes they guess right and sometimes they don’t. Centivax wants to change that with a single dose that protects against all flu strains, with a booster needed only every two to four years instead of annually. The company's universal flu vaccine is in a Phase 1 trial, with results expected by end of year.

“If our data looks good by the end of this year,” CEO Jake Glanville said, “effectively, the pandemic era for influenza is over.”

Centivax is applying its approach, targeting conserved regions across strains rather than trying to keep up with seasonal mutations, beyond flu, too: universal snake venom, Alzheimer's, malaria, and cancer vaccines are in the pipeline. As Glanville puts it, “You want to treat all the flus, not just one strain. You want to treat all the snakes, not just one snake.”

Of course, it wouldn’t be a medical miracle entry without some FDA hair. Last month, the FDA about-faced on Moderna’s mRNA flu vaccine proposal and put it back under review. Just to be safe, Centivax is hedging by running trials in Europe and Australia alongside the US. “If something unexpected happens with the FDA, we would just proceed in Europe,” Glanville says.

The possibility that Europeans get a universal flu vaccine before Americans because of regulatory dysfunction is both absurd and completely believable. Hopefully, if the FDA blocks the universal everything vaccine, they will also watch my kids during sick season.

(5) How to Turn a Chicken Egg Into a Drug Factory

Carl Zimmer for The New York Times

I am so clucking eggcited that Neion Bio is finally coming out of its shell.

Today, pharma uses Chinese hamster ovary (CHO) cells to produce biologics like Keytruda and Humira in huge stainless steel bioreactors. Merck recently spent $1B on a single Keytruda facility.

In early 2024, Elliot wrote that chicken eggs are much more efficient bioreactors. They run on grain and water, produce six grams of protein per unit, and we already farm them at massive scale.

As the (excellent) Neion website says, chicken egg drug manufacturing can be “extremely low COGS, hyper resilient, CapEx avoidant, ultra scalable, very high reliability, and lower environmental impact.” Nature evolved the chicken egg to be that way. The hard part has been that the science hasn’t been able to modify what’s inside the egg fast and reliably enough to actually manufacture drugs.

Sam Levin, who we previously backed at Melonfrost, and Dimi Kellari explored the frontiers of chicken egg research, which they could get their arms around because the cutting edge stuff is happening in a very small number of labs around the world. They went to all of them, and they realized that the time was right to build a company that uses nature's bioreactors to produce drugs at a fraction of the cost.

In today's NYT article, Sam predicts that the cost can be 1/10th or even 1/100th of the current cost, and that just 3,900 hens could meet global Humira demand.

I'm proud to back Sam, Dimi, and the Neion Bio team out of not boring capital as they work to hatch the balk of the world's biologics and dramatically lower the cost to produce critical drugs, right here in NYC. And I'm sure they're happy that I can share my chicken puns with all of you instead of just replying to investor updates with them.

Extra Dose: Science Breakthroughs, GPS, Memory Without Brains, Astro Mechanica

Join not boring world to get an even bigger Dose of Optimism.

Read more

Subscribe now

Hi friends 👋 ,

Happy Tuesday! Welcome to our newest installment in what has become an unintentional two-part series on non-LLM models that can do things that humans can’t, things that will give us superhuman abilities in the physical world. They’re also both co-written with founders you’d expect to find in SF but are building right here in the greatest city in the world, NYC.

The first was last week’s essay on World Models with Pim de Witte.

Today’s is about machines that can intuit electromagnetic fields in a way almost no humans can that will help us design and build better electromagnetic (EM) systems.

As you know, I’m very bullish on the growing role of electromagnetic systems in the economy. After Sam and I wrote The Electric Slide, Arena Physica CEO Pratap Ranade and I traded emails. In one of them, he wrote:

The electrical and electromagnetic components are the “nervous system” of modern hardware and contribute to 40-50% of failures. Our ability as a nation to test and build it has declined, but –– imo even bigger –– as a species, we’re still unable to wield electromagnetism to its full potential.

Over the past seven months, we’ve developed a friendship, and Pratap has broken my brain many times. One of the things that’s most fascinated me is the idea he’s betting his company on, the one he emailed me about: humans can’t intuit EM, and it’s a bottleneck to the electric progress we both want to see. There’s no reason machines can’t be taught to understand them much better than we can, though.

For the past few years, Arena has been building AI tools and deploying expert electrical and RF engineers to help companies design, develop, and debug electromagnetic hardware. They’re working with companies including AMD, Anduril, and Sivers Semiconductors. They are backed by investors including Founders Fund, Peter Thiel, Initialized (Garry Tan), Shield Capital, and 137 Ventures.

Today, they’re re-branding as Arena Physica with the expanded mission to develop “Electromagnetic Superintelligence.”

This is an essay about how to teach machines to see the fields that we can’t, and what the world might look like if we can.

Let’s get to it.

Today’s Not Boring is brought to you by… The Pitch by Deel

Normally, our sponsors would like me to tell you why you should give them your hard-earned money. This time, Deel has asked me to tell you how they can give YOU money in The Pitch.

Deel recently launched a global tournament with $15M in prizes for startups: the top 10 get a $1M investment each, and 100 regional winners will get $50k. You don’t need warm intros and you don’t need to pay to apply. It’s just a pure competition for the best entrepreneurs.

Pitch in just two minutes for the chance to win a $1M investment, access to a global ecosystem designed to help you scale (partners and sponsors include Stripe, Google, AWS, and a16z), and to get your startup in front of global leaders. This is your shot.

Apply Now

Electromagnetism Secretly Runs the World

Electromagnetism secretly runs our world. “Secretly,” because only a few people on this planet can intuit how it works.

Your phone’s GPS is powered by satellites that broadcast electromagnetic (EM) waves with timestamps. The wi-fi in your apartment is created by EM waves bouncing around the walls. Air traffic control is radar, as in EM waves that pulse out and listen for echoes off aircraft. When Maverick locks onto a bogey in Top Gun, he’s using a phased array radar steering EM beams electronically. Contactless payment? EM. Microwave oven? EM. The fiber optic cables carrying the internet across the ocean floor and through Somos’ network? That’s light… which is also EM.

Every single wireless signal, medical image, radar sweep, every chip talking to another chip inside a data center. All of it is electromagnetic waves, shaped and directed by physical structures designed to manipulate these waves. And, as electricity and intelligence race to define our era, EM’s presence is only growing more pronounced. In our data centers, chips communicate with each other via short-range EM waves. If Elon successfully moves the data centers to space, AI will be beamed down from satellites to your device via EM waves.

As Packy and Sam wrote in The Electric Slide, everything that can economically go electric will. Cars, trucks, buses, drones, boats, stoves, heat pumps, batteries, bikes, even planes, anything that moves, heats, lights, computes, or converts energy is moving from mechanical to electric. All of those newly electric things will be full of EM components.

In 1970, electronics accounted for five percent of a new car’s cost, on average. By 2020, that number reached forty percent. By 2030, it’s anticipated that the cost of the electronics of a consumer automotive vehicle will reach fifty percent of the vehicle cost.

Electronics comprise 35% of the cost of the F35 Lightning II, more than the cost of the engine itself, and 15% of the Pratt & Whitney F135 engine, which costs $20 million. By the 2030s, when it’s projected that defense contractors will be building the F-47, they’ll be spending over 40% of the $300 million airframe on electronics.

This is good. We want to see this electrification continue. Electric machines perform better with less impact on the environment, give us capabilities combustion engines can’t, are better-suited for autonomy, and are riding cost/performance curves that will continue to widen the advantage.

But among a number of challenges addressed in The Electric Slide particularly as it relates to production, there’s an equally large one looming in the research and development of new and better electromagnetic machines: our electromagnetic capabilities are bottlenecked by the very small number of humans who actually understand how any of this works.

There is a reason radiofrequency (RF) engineering — the practice of designing hardware that shapes and directs EM waves — is often referred to as black magic. There are maybe ten people in the world who can deeply intuit electromagnetism, who can see which shapes will create which EM fields in their mind’s eye1. I am not one of those people, but I’ve met them. I’m hiring them at my company, Arena Physica, and I went to school with many of them.

There was a guy in my physics program about whom our professor asked, “You know what’s special about this guy?” We all said no. “This guy thinks like an electron.”

What he meant was that, electrons, if they were sentient, would feel all of these different fields pulling at them. Electrons would probably have an intuition for this feeling, the same way we have an intuitive feel for gravity, how we just know that when we let go of a ball, it will drop to the ground. Those friends of our ancestors’ who couldn’t intuit gravity did not live long enough to reproduce.

Some people – a vanishingly small number – have spent enough time studying, testing, designing, and simulating electromagnetic systems to be able to intuit them like gravity. But for the rest of us, electromagnetism is mostly invisible.

For the vast majority of human history, we haven’t needed to see beyond the visible spectrum in order to survive. And so we haven’t. Those of our ancestors’ friends who wasted precious resources on seeing the full spectrum of EM waves wouldn’t have lived to pass on these traits, either.

Humans can see a small part of the electromagnetic spectrum, the “visible light” portion with wavelengths between 400 nanometers (violet) and 700 nanometers (red). We don’t see shorter wavelengths (ultraviolet light, X-rays, or gamma rays) or longer wavelengths (infrared, microwave, or radio waves).

This has served us just fine. Until electromagnetism came to run the world.

We have a fundamental force that we rely deeply on, but one that very few of us can work with naturally. This slows technological progress and limits what we can make.

Fortunately, AI doesn’t share our blind spots. It is particularly good at seeing patterns, at making connections and understanding dependencies that are not necessarily intuitive to humans.

Because of this, we believe that computers will be much better at grasping electromagnetism than we are. We should be able to build a Large Field Model (LFM) — like an LLM that generalizes across language, except ours generalizes across EM. We should be able to use this LFM to understand EM waves and shape them to do what we’d like them to do.

That’s the big bet we’re making at Arena Physica. To understand why we’re making it, I want to first make sure you understand electromagnetism.

A Brief Primer on Electromagnetism

Packy and Sam gave A Brief History of Electromagnetism in The Electric Slide.

I’m going to add to that with A Brief Primer on Electromagnetism. I’ll sprinkle in relevant history, but my goal is to make sure we have a working understanding of electromagnetism.

There are four fundamental forces that govern how everything in our universe works:

1. Strong force

2. Weak force

3. Gravity

4. Electromagnetism

The strong force and weak force operate at subatomic scales. The strong force binds protons and neutrons together in atomic nuclei. The weak force enables radioactive decay and nuclear fusion.

Gravity is the weakest of the four forces by an enormous margin (roughly 10³⁶ times weaker than electromagnetism). Yet, it dominates at cosmic scales. It only attracts, never repels, which means its forces keep adding up. And it acts on every single particle with mass or energy. It’s also mysterious at a fundamental level: gravity and quantum theory are incredibly powerful theories for how our world works, but they are fundamentally incompatible. This remains one of the deepest unsolved problems in physics. But we have a strong, intuitive relationship with gravity. In our day-to-day lives, we can feel the force.

Electromagnetism is the force we interact with most directly in everyday life. It’s also the one we’ve industrialized most aggressively. It governs light, electricity, magnetism, and chemistry, essentially everything about how matter behaves above the nuclear scale. It’s why matter has structure, why chemistry works, and why technology works. It’s responsible for the structure of atoms (electrons bound to nuclei), the bonds between molecules, the rigidity of solid objects, and all electronics and communication technology. Unlike gravity, electromagnetism has both positive and negative charges, which means it can attract or repel, and large accumulations tend to neutralize themselves. Our mathematical understanding of electromagnetism is extremely accurate: it is described by quantum electrodynamics (QED), the most precisely tested theory in all of science.

And yet… despite this precision, electromagnetic systems can be deeply counterintuitive. RF engineering, for instance, has a reputation as black magic. The wave-like, distributed nature of fields at certain frequencies produces effects that violate the intuitions built from simple circuit theory.

But let’s try to build up our intuition as best we can.

Every force carries energy. Electromagnetic energy comes in quanta we call photons, particles of light, but for most of what we build — antennas, radars, communication systems, phased arrays — it is easier to think about it as waves, in terms of frequency, wavelength, and phase. A photon has different amounts of energy based on its frequency, which can be seen on the electromagnetic spectrum. A super-high energy photon would be something like an ASML machine making chips using EUV (extreme ultraviolet). EUV is very, very high frequency. Therefore, it is very high energy, and in turn, very short wavelength. As you go through the visible spectrum and get to the other side, you reach infrared, and electromagnetic energy becomes heat. Then you get to RF (radiofrequency). With RF you have very low-energy photons. But it’s really just all electromagnetic waves that follow the same principle. High energy = high frequency = short wavelength and vice versa. (Explore the EM spectrum here.)

Now, think about a prism. A prism is an object or a material that treats incident EM waves — or different photons — differently. So if you’re a red photon, your refractive index, or how much you bend, is a certain number. If you’re a blue photon, it’s different. Refract the whole beam and you end up with that beautiful rainbow on the other side of the prism.

The prism is an early example of humans manipulating electromagnetic fields. People noticed that light, when it passed through glass or a crystal shaped in a certain way, would form a rainbow. We’ve come up with many ways to manipulate electromagnetic fields since, the most consequential of which is also the simplest.

If a caveman discovered he could manipulate electrons, the very first thing he might do is the simplest possible thing: electron on, electron off. On, off. One, zero.

The prism is actually much more sophisticated than the on, off switch. It’s more subtle, more expressive, and its implications are more powerful. But computing is based on the caveman’s idea of flipping the switch.

Early computers were literally built on this simple idea: they used mechanical switches known as relays to compute. When the metal touches, current flows (1/on). When it separates, current stops (0/off).

One problem with mechanical switches was that, when you switch often in air, the air ionizes and creates tiny bolts of lightning that can jump across the gap and create an arc. These arcs can make your “bit” unreliable. Sometimes it’s supposed to switch, and it doesn’t.

The other problem was that they are very slow. It was the advent of the electronic switch that put us on a path to be able to now build processors that can work at Ghz speeds (10⁹ cycles per second).

So we invented vacuum tubes2. They removed the air. Without atmosphere, there was no arcing. But vacuum tubes were fragile, power-hungry, and couldn’t scale. The great breakthrough was the semiconductor, materials like silicon that can have their conduction controlled. They can be made to conduct or not conduct (hence semi-conductor) based on applied voltage. Semiconductors enabled the transition from mechanical to digital, and it gave us the transistor, which are tiny silicon devices that switch on or off using voltage. This gave us Moore’s Law, which gave us Boolean logic, which gave us everything in modern computing. That singular innovation, the transistor, has produced most of our technological advancement over the past seventy years.

But if you read Gordon Moore’s 1965 paper, the one in which he described what would come to be known as Moore’s Law, you’ll find that only the first half is about digital silicon; the second half is about analog silicon.

Nobody paid attention to the analog part, but I think it’s even more fascinating today than the digital one.

Digital silicon is about switching: transistor on or off, conducting or not, one or zero. All the gates, all the logic, all of computation follows from that binary foundation. It’s powerful, but it’s also, as we’ve discussed, the simplest possible thing you can do with an electron. It’s caveman math.

Analog silicon is about shaping. Instead of just on/off, you’re asking: what if I could bend the electromagnetic wave? What if I could guide it, direct it, absorb it at specific frequencies and reflect it at others? In practice, this is RF front-ends, antennas, packages, and printed circuit boards (PCBs) behaving like a single, unified electromagnetic object.

This is how the world works too. The world is analog. The world does not work in 0s and 1s, but rather in the continuum in between them. Even if all computation is done digitally, you’ll need to deal with analog signals and shape waves the moment you need to interact with the real world (for example, capture sound in a microphone, produce sound in a speaker, send wireless signals over the air, send light on optical fibre)

Remember the prism? That’s what analog silicon does, but for all electromagnetic frequencies, not just visible light. Instead of glass bending light, you can use carefully shaped conductors printed on silicon to bend, direct, and shape EM waves.

This is where we leave the realm of deterministic computing and enter a world of black magic.

Try This At Home

Here’s an experiment. You can try this at home.

Materials List: COPPER WIRE, COMPASS, BATTERY.

Take your copper wire, connect it to a battery, and run current straight through it. The magnetic field it produces will wrap around the wire in a helix. You can confirm it’s doing this by holding a compass near it and watching the needle deflect perpendicular to the wire.

Now, coil that same wire into a spring shape (a solenoid) by wrapping it around a pencil or screw 10-15 times. Run current through it. The magnetic field is completely different: instead of wrapping around the wire, it shoots straight through the center of the coil. Same wire, same current. But a different shape = radically different field.

This is the fundamental game of electromagnetism: geometry determines behavior. Every antenna, radar, or phased array tile is just a more sophisticated version of this principle. Find the right shape, and you can make electromagnetic fields do almost anything.

To understand why shapes matter so much, consider what happens when an electromagnetic wave hits a conductor.

A conductor is special because it has free electrons. Free electrons are not locked into a lattice like in an insulator, but instead swim around in what physicists call an “electron sea.” When a photon (an EM wave) hits this electron sea, those electrons start to move in response. They oscillate with the wave.

This is fundamentally how an antenna works. The old bent TV antenna on your grandparent’s roof was shaped specifically to receive UHF frequencies broadcast from a distant TV station. The EM waves traveling through the atmosphere would hit the antenna, excite the electrons in the metal, and those oscillating electrons would travel down the wire into your TV as a signal.

That signal carried information, like encoded images of I Love Lucy, compressed into patterns of electromagnetic oscillation, broadcasted through the air, absorbed by your antenna, decoded by your TV. If you step back and think about it, this entire chain is completely absurd. We transmit moving pictures through the air using invisible waves. And turning those waves back into pictures all comes down to the shape of a wire.

Radar works basically the same way, except it’s more high-powered and moves in reverse.

World War II accelerated radar. It also showed us how badly we needed it. The Allies were being pummeled and needed to track incoming threats. They turned to radar, which had developed quickly thanks to the war. In the late 19th and early 20th centuries, Heinrich Hertz (of Hz fame) showed that radio waves could reflect off objects. Several physicists also noticed that radio signals behaved strangely when ships or other objects were nearby. Through the 1920s and early 1930s, scientists in the U.S., U.K., Germany, France, the Soviet Union, Italy, and Japan all experimented with using radio echoes to detect objects.

In 1935, a Brit named Robert Watson-Watt (no relation to the steam engine Watts) proposed and then demonstrated a practical aircraft-detection system using pulsed radio waves. This led to the Chain Home early-warning network along the English coast. Chain Home was operational at the start of WWII and gave the Royal Air Force so much of an early warning in the Battle of Britain that it’s often credited as a key factor in preventing a German invasion. The United States picked up development a bit later, with the benefit of British tech transfer, and scaled up the technology’s capabilities and manufacturing. In the U.S., Alfred Loomis led research efforts at Tuxedo Park3 and helped establish MIT’s Rad Lab, which developed fire-control radar, airborne radar, and navigation radar. Germany built parallel systems that pushed the state-of-the-art in different directions.

Radar, instead of receiving a broadcast signal, transmits a beam on multiple wavelengths, waits for it to reflect off something (like a bomber), and then listens for the echo. If the object is big enough and close enough, you can detect it.

But to scan the sky, you need to point your beam in different directions. In the 1940s, that meant literally spinning a large dish antenna. You needed mechanical motors to do it. A mechanical gimbal rotating a giant antenna.

This worked, but it had obvious limitations. Moving parts break, for one. For another, the dish can only spin so fast. Today, Starlink satellites need to update their pointing multiple times a second, since they are moving at 7.6km/second. Try doing that mechanically and for 5,000 simultaneous beams.

This is where the second half of Moore’s 1965 paper becomes relevant. Moore realized you could use transistors to solve the spinning-dish problem. Replace the mechanical movement with electronic steering.

The key insight is constructive and destructive interference, the same phenomenon you see when ripples on a pond meet and either amplify or cancel each other.

Imagine you have a grid of small antenna tiles instead of a single big dish, like a checkerboard where each square is a tiny antenna. Each tile can emit a signal. Each tile is the whole RF front-end and antenna structures across chip/package/PCB that acts like a single, holistic EM object. Now, if you start the signal from the leftmost tile first, then the next tile a tiny bit later, then the next, and so on and so forth, the wave fronts from each tile will interfere with each other. Get the timing right and they’ll constructively interfere in one specific direction, creating the effect of a single, focused beam pointing in the direction you want.

Change the timing pattern, and the beam will point somewhere else. You can replace moving parts with analog and digital logic that controls when each tile fires.

This is called a phased array. And it’s how modern radar works. If you want to develop a better intuition by playing with it, we built a little simulator here.

The radar on an F-35 is called an AESA (Active Electronically Scanned Array). Nothing on it moves. It’s just a grid of semiconductor tiles, and the “beam” sweeps across the sky purely through timing. It’s also how Starlink works. Each Starlink terminal has 1,280 of these beam-forming silicon tiles. That’s why you can buy a flat panel for $300 that does what used to require a million-dollar spinning dish.

What’s happening on those tiles?

Remember: digital silicon is about transistors switching on and off. But the tiles in a phased array shape electromagnetic fields through their physical geometry.

Think back to the TV antenna, a bent wire specifically shaped to receive certain frequencies. Imagine you could print that shape onto a silicon chip, layer by layer, laying down copper traces in specific geometries, creating structures that interact with EM waves in precise ways.

On one layer, you might have a spiral. On the next, a grid. On the next, something that looks like a QR code. Stack them up with tiny connections between layers called vias, and you’ve created a three-dimensional structure that can emit, receive, absorb, and reflect electromagnetic waves at specific frequencies and in specific directions, which you have full control over.

This is what a phased array tile actually is: a 3D sculpture of copper and silicon, designed so that electrons create exactly the EM field you want when electrons flow through it.

Why This is So Hard to Build

Creating an EM field of your liking is all about geometry, which means it’s all about shapes. But how do you know which shapes to make?

With digital silicon, the rules are relatively simple. Transistors are either on or off. You can simulate billions of them with perfect accuracy. The design problem is about routing and timing, but the physics is well-behaved.

Analog silicon is different. The physics is wave physics, and waves do things that violate our intuitions.

At optical (high) frequencies, we often get away with “ray optics” intuition; light mostly travels in straight-ish lines, reflecting and refracting, and you can treat its components as fairly local.

At RF, the wavelength is big enough that your whole device becomes part of the circuit. Fields couple into enclosures, PCBs, screws, nearby tiles… Everything talks to everything. That’s why RF feels like black magic, and why you have to simulate the whole object to predict how it’ll perform.

When you’re designing a Starlink tile, for example, you can’t just model the tile in isolation. The EM waves emanating from it will interact with the entire Starlink unit: the metal casing, the other tiles, the mounting bracket, the support structure. You have to simulate the whole system at once.

This is why there are no automated tools for analog circuit design. Digital circuits can be “synthesized” from code; a digital designer can write “RTL” code that describes how his digital circuit works. Then, there are tools that can read the code, and “compile” it to a chip. But no such tools exist for analog design. There are no “standard cells” for analog, no standard analog designs, no standard building blocks. Everything interacts with everything.

Which is why there’s no ARM for analog silicon4. There are no companies that can sell “IP,” standardized circuit designs, to a multitude of customers in a highly profitable fashion — no such standard circuit exists. Every new system is different, and as a result, each new customer has different needs.

ARM can design a chip that works in any phone because digital chips are self-contained. But an analog phased array tile designed for a Starlink terminal won’t work in a different satellite. The interference patterns will be completely different!

And the simulation tools are slow. The equations governing electromagnetic fields are called Maxwell’s equations, four partial differential equations that are notoriously difficult to solve.

They’re just equations. What’s the problem?

If an EM wave is at a higher frequency, it’s more “particle-like”— intuitively, it’s like a ball — you know where it is, and it bounces off stuff cleanly. If the ball is in one corner, it doesn’t affect anything in the other corner. As EM wavelengths get longer (into RF), they become more wave-like, and the particle is sort of “spread out.” The waves start to interfere with each other a lot, like ripples on a pond. They can either strengthen or cancel each other out.

So, if you’re NVIDIA and selling high frequency chips in boxes, you can sell a single product. You can design one GPU and sell it to everyone. There is no difference between the Navy putting your chip on a ship, or Sony putting it in a Playstation. They can all just buy the chip and plug it in. But, for example, if you’re buying components for a phased array system, you have to model the entire system, because it’s type 2, not type 1. A design for a Navy ship won’t work in a Starlink terminal. The EM fields interact with everything around them—the metal casing, the mounting structure, nearby components. Change the environment, and you need a completely new design. Everything becomes a custom services problem that is rate limited by these rare experts and simulation.

In short, the solvers are slow, even with supercomputers or programs like Ansys, because the equations are tough to solve and require expertise to wield. The reason the equations are very hard is boundary conditions (the edges, where smooth calculus breaks down), e.g., a sharp metal edge can cause problems by creating strong electromagnetic reflections that cause fields to concentrate in unexpected ways.

Running a full simulation of a proposed design can take hours or days. Here’s an example design loop: make your best guess at a shape, wait hours for simulation, discover it doesn’t quite work, adjust the shape, wait hours again. As we collaborate with field experts, we witness each simulation iteration with legacy tools taking a week. That’s not enough ‘shots on goal’ to develop Electromagnetic Superintelligence.

RF design can’t be done by brute-force computation. The search space is infinite and each evaluation takes too long.

Take a simple two-layer circuit where each pixel in a 64×64 grid can be either metal or dielectric. That’s already roughly 2^64x64, or 10^1,233, possible configurations for a single, small component. The entire history of human RF design has explored a vanishingly small fraction of that space, obviously. Here, see how many of those configurations you can come up with.

Navigating this search space requires intuition. You need someone who can look at the desired field pattern and just... sense what shape might produce it.

The people who can do this have spent decades building up a feel for how electrons move through structures, how fields bend around corners, how waves interfere. They can sketch a spiral on a whiteboard and tell you roughly what frequencies it will emit strongly and which it will absorb. Aside from my classmate who could see like an electron, though, this intuition isn’t natural even to those special few. They acquire it painstakingly over long careers. Because, unlike gravity, there was never evolutionary pressure to understand electromagnetic fields outside of the visible spectrum. We don’t feel them. They’re invisible to us.

I watch my baby daughter learning about the world. She already has an intuition for mechanics. She knows that if you roll a glass off a table, it will break. She has no intuition for electromagnetism, which is probably genetic. 99.99% of people don’t.

It is a miracle that we’ve been able to manipulate EM waves to our purposes to the extent that we have. But the world is only getting more electromagnetic, and we will need a lot more shapes.

That means we need to build something that does have intuition for electromagnetism.

AlphaGo for Electromagnetism

In 2016, DeepMind’s AlphaGo defeated Lee Sedol, one of the greatest Go players in history.

The moment that stuck with everyone was in Game 2, Move 37.

The expert commentary went something like this: “That’s a mistake.” Then: “That’s stupid.” Then: “That’s a very strange move.” And finally: “That’s beautiful. That’s elegant.”

AlphaGo had done something no human would have tried, a move so unconventional that the world’s best players initially dismissed it as an error. But it worked. The machine had discovered a strategy that humans, despite thousands of years of playing Go, had never found.

What made AlphaGo possible? Two things. First, Go has clear rules and a perfect simulator. You always know exactly what state the board is in and which moves are legal. Second, because of these limitations, a computer can play millions of games against itself very quickly. AlphaGo learned by playing more games of Go than all humans in history combined.

We want to do what AlphaGo did for Go — but for physics. What if we could build a system that played millions of “games” of electromagnetic design and developed an intuition that humans simply can’t acquire?

There was an obvious obstacle between us and that dream. AlphaGo worked because Go is a perfect simulation. You know exactly what happens when you place a stone on the board. But physics is more complicated, and the simulators are slow. Maxwell’s equations take hours to solve. You can’t “play a million games” overnight.

So we needed to build the simulator first.

The EM foundation model that we’ve built, and continue to scale, is the simulator for EM physics.

The starting point for what we’ve built is known as a neural surrogate. The idea is simple: instead of solving Maxwell’s equations from scratch every time (which is slow), you train a neural network to approximate the solutions (which is fast). It’s like the difference between calculating the sine of an angle by hand versus looking it up in a table, except the “table” is a neural network that can interpolate to angles you’ve never seen before.

Traditional physics simulators work by brute force. They break space into tiny chunks, apply the equations at each point, and iterate until the solution converges. It’s accurate, but a single simulation can take hours.

But what we’re building goes beyond a surrogate. Most surrogates in physics are narrow: trained to approximate one specific simulator for one specific class of problems. Arena Physica’s model learns the relationship between shapes and fields directly, allowing it to generalize. It’s not a faster calculator (it’s not a calculator at all). The neural surrogate is learning the syntax of physics. Just as GPT learned the “logic” of language, our model is learning the “logic” of fields. Show it enough examples of “this shape produces this field pattern,” and it learns to predict new patterns for new shapes almost instantly. We’re talking about 18,000x speedups, hours to milliseconds.

If you’re reading closely, you might be thinking to yourself: of course you can go faster if you’re just trying to get an approximate answer and not a perfect solution.

Good catch. This is where the magic happens.

When you’re searching for good designs, speed and direction matter more than precision.

Think about how an experienced RF engineer actually works. They use their intuition to filter out ideas that probably won’t work and to get the rough shape of ideas that might. Then, they simulate those. They make fast, approximate judgments to decide where to invest their slow, precise simulation time.

Arena Physica’s model does the same filtering, just much faster. It doesn’t need to tell you exactly how well a shape will perform. It needs to tell you how each shape will perform relative to the others. Good enough for search is a much lower bar than good enough for publication.

Speed lets us flip the problem around. Instead of asking “what field does this shape produce?” we can ask “what shape produces this field?” That’s generative design. We specify what we want, say, an antenna that transmits strongly at 28 GHz but rejects interference at neighboring frequencies. The system uses our desired state to generate shapes that might achieve our goals.

Then, we pair two models in a loop: one that generates designs, and one that evaluates them5.

The generator proposes a batch of shapes, many of them wild, strange, things that no human would come up with. Move 37 shapes. The evaluator characterizes them all in seconds, directionally: this one’s terrible, that one’s promising, this one’s interesting. The best candidates get refined via small variations and perturbations. The evaluator evaluates the refinements. Repeat.

At each step, we ask, “Is this one better for my goal than the shape we had before?” Because we know what the rules are, like AlphaGo, and what our goals are, like AlphaGo, we can reward the model for getting closer. And like AlphaGo, by making simulation cheap, we can explore much more of the design space than we could have if everything we wanted to try required a precise multi-hour simulation.

In Asymmetry of verification and verifier’s rule, OpenAI’s Jason Wei describes the “verifier’s law.” Essentially, it says that any task that’s easy and fast to verify will be automated by AI. The hard part about our world is verification relies on special humans and simulators that are slow and expensive. By attacking this first via the foundation model for fields, acting as a fast simulator, we’ve made this problem accessible by generative AI for the first time. Our generator learns the weights through our feedback loop.

This is the same loop that made AlphaGo work: generate, evaluate, learn, repeat.

Run the loop for yourself here. You get a visceral sense for the importance of speed.

Of course, this loop only works if you have enough training data to feed it, and unlike LLMs, which can scrape the internet for training data, EM field simulations don’t exist in the wild. Nearly every single data point has to be created. So we’re building our own Data Factory.

To do that, we’re hiring the best RF lead designers we can find. Put them in one place, where, theoretically, we can amortize the capabilities of this scarce group of people over all sorts of existing and newly possible customers. Have them create designs, give feedback, test their designs, run through the loop.

We generate random designs synthetically, our experts create seed designs that our system can then procedurally amplify, and we fabricate the top candidates and pipe real-world measurements back into training.

The Data Factory has three layers: high-volume synthetic data, high-information expert-seeded data, and ground-truth fabrication data.

Because of how manual the Data Factory is today, we need to pick our use cases strategically. We are starting with analog silicon (chip packaging, phased array components, RF front-ends) and the full phased array system we mentioned above. We’ll be expanding into new domains, like superconducting quantum computing, in conversation with our partners.

The factory is the moat. You can’t build a foundation model for EM without it, nobody else that we know of is building one, and to try, you’d have to hire from the very small pool of experts that we’re bringing together at Arena Physica.

Then, when we’ve converged on something promising, by feeding the Data Factory’s output into our loop, we validate it by running the slow, precise traditional solver on our best candidate. Or better yet, by fabricating the design and validating it in the real world.

Remember why there’s no ARM for analog? At RF frequencies, the wavelengths are long enough that EM waves interact with everything around them. When designing a phased array for a Starlink terminal, you can’t just model the chip; you have to model the chip, the circuit board, the metal casing, the mounting structure, everything. It all affects how the EM waves behave.

Because of that, we’re even using our own components to build an entire phased array system for imaging and detection from scratch.

We’ll tape out silicon before the end of the year. And, for any part of the problem that’s not analog, we’re actually using our agentic stack — our hardware-aware agents, operating on our metagraph – a dynamic graph representation of the hardware – talking to tools via MCP — to speed up every single aspect of the process, so that we can go end-to-end faster. In this way, we benefit from all the amazing leaps coming from the foundation models, but also own something they can’t replicate: an EM foundation model, fed by our own Data Factory.

The system compounds: fast approximate evaluation enables broad search, broad search finds promising candidates, fabrication validates and generates training data, training data improves the generator, better generator enables even broader search.

If you can do all of the RF design work in our loop, you can build an analog IP Factory.

The IP Factory for RF and the Compiler for Atoms

Our EM Foundation model’s key advantage over existing surrogates is that it can generalize.

Take a philosophical leap with me.

LLMs don’t mechanically learn to classify a sentence. They train on the structure of words and sentences in relation to each other. The rest of their behavior is emergent.

Before LLMs, you had spam detection as its own major problem. You had summarization as its own problem. Translation was its own separate problem. There were actually good companies with good ML teams doing each. The thing that they all got wrong was focusing on narrow problems. What we’ve learned is that if you can understand language at the root level, and you see scaling laws, you can get all the downstream applications for free.

So if it is true that there’s a fundamental relationship between geometry and EM fields, just like there is in language, and if scaling laws are true, then this model should generalize.

EM simulation looks a lot like language pre-LLMs. Where Ansys simulators fail – e.g., this tool is for antenna simulation, this tool is for EMI simulation for an automotive engine – we could tackle all of that with one model, like an LLM tackles translation, sentiment analysis, spam detection, and so much more all in one.

We believe that LLMs were the first foundation models, not the last. Language is one primitive of intelligence – the one humans use to communicate and think. But the universe has other primitives. Newton created calculus because the language of the universe is not English. LLMs will let us interface with new foundation models that can push humanity’s understanding of the universe farther, giving us intuition for things our biology wasn’t evolved for. Such models are a critical part of a positive-sum future of AI.

This is the huge bet that we are making at Arena Physica. It’s our thesis. We are already building a strong business that helps companies better understand and iterate on their electromagnetic systems, but we think generalization will also allow us to build the IP Factory for RF.

With our loop, we can automate IP creation and design. This is the key distinction between our model and ARM’s. If generalization works, then instead of selling one design to many customers like ARM does, we can generate a unique design for each customer and each use case almost instantly, with roughly the same effort it takes ARM to create one-size-fits-all IP.

And we think, and are seeing early signs, that our model can scale this automation across the whole electromagnetic spectrum, in which case our TAM is anything with a wave.

One question we get often is: can’t LLMs just do this? We’ve been running our own internal tests against frontier LLMs – both regular and extended thinking - and their performance gap to our base model is substantial. Our model achieves a magnitude weighted-MAE (Mean Absolute Error) well under 1 dB (for context, the range that RF engineers typically care about spans roughly 20-30 dB, so <1dB is a very strong result)6.

What we’re building is different than LLMs, and better at what it was built to do. A useful way to think about what this unlocks is as a “Compiler for Atoms.”

In software, compilers translate high-level programming languages into binary instruction sets that a CPU can execute. We went from assembly to C++ to Python, and now Claude Code is arguably a compiler with English as the programming language. The LLM compiles down to the programming language it decides is best, which then compiles down further to machine code. At each step in this progression, the abstraction gets higher and the number of people who can “program” gets larger.

Physics doesn’t have a compiler yet. The universe has an instruction set: materials and geometries arranged in specific configurations. Place them one way, you get a motor. Place them another, you get an invisibility cloak. We know all of this is possible because the equations tell us so. Human physicists have spent centuries learning this instruction set. But to access any of it, you still need to hire the equivalent of an assembly language programmer: a physicist who has spent decades learning to translate between human intent and the physical world’s instruction set.

What we’re building at Arena Physica is, in some sense, a Compiler for Atoms: a way to express what you want in high-level terms and have it compiled down into the geometries and materials that produce it, starting with Maxwell’s equations and eventually, we hope, adding Schrödinger’s.

When we launch our EM foundation model next week, we’ll make it available to interact with via an agentic UI. You’ll be able to type a request in plain English like, “I need an eight gigahertz band pass filter for a satellite uplink.” The LLM translates that target scattering parameters, the technical parameters we’re optimizing for. Our LFM generates candidate geometries using physics as its reasoning substrate rather than language. Then, the LLM engages again to explain what the model did and why, drawing on the foundational RF engineering knowledge we’ve built into the system.

Andrej Karpathy says LLMs are “people spirits.” In our system, the spirit of David Pozar, author of the bible of RF engineering, provides his best explanation of the geometries generated by our EM foundation model. It’s like working with an intern who happens to speak electron, except the intern is channeling decades of accumulated RF wisdom.

I think this will be a powerful paradigm for the future. Everyone is thinking about human-to-model interactions. But the bulk of the work in systems like ours is model-to-model: the LLM talking to the EM foundation model, the EM foundation model responding, the two iterating through design space at machine speed. The human-to-model layer becomes a minority of the interactions, the intent-setting and interpretation. The real work happens in a language we can’t speak, translated for us by models that can.

In the future, language models might serve as the universal interface between humans and an entire ecosystem of specialized foundation models: for EM fields, for biology, for materials science. In this future, the LLM becomes the diplomat between species of intelligence.

Eventually, we want to get to a world in which anyone can say, “I want an invisibility cloak” or “I want a cheaper Starlink,” and the machine will design it. But today, even with our compiler, we’re still in the C++ era. Expert “physics coders” still need to tell the machine, “I need an 8 gigahertz band pass filter for a satellite uplink.”

In the meantime, we want to deliver that “eventually” future today.

Taking Problems Off Customers’ Plates Entirely

Since it became clear that our model had the potential to scale and generalize, I’ve been thinking a lot about the right way to deliver the capabilities it provides to customers.

I’ve thought about whether we should build the ARM for Analog. I’ve thought about selling access to the model, or the IP Factory, directly. I don’t think either are quite right. I think it’s worth talking through where we landed and how, because what constitutes the right business model is changing a lot with AI, and the model we landed on is probably not the one we would have pursued a few years ago.

If you think through what makes Arena Physica unique, it’s really four things: we have a talent-dense collection of some of the world’s best RF engineers, a “Compiler for Physics” that can generate valuable IP in experts’ hands, a Data Factory that improves as those RF engineers generate and validate more IP, and a software platform (complete with FDEs) that makes it easy to apply LLM-based agents to reason about hardware.

The business model that falls out of that is services. We’ve hired some of the world’s best RF designers and a team of excellent electrical engineers. Each one, working with our EM foundational model, with our agents, and with LLMs, can cover enormous ground. So instead of selling people the tools and asking them to figure it out on their own, we’re starting to just take problems off people’s plates.

You want to launch a space company and you have an unsolved problem with how your racks communicate in orbit? We’ll solve that for you. You need a silicon layout for your next chip? We’ll do the layout. We send in RF and electrical engineers, armed with our tools, and deliver the product the customer needs. It’s full-stack electromagnetic engineering as a service, at a speed and cost that wasn’t possible before, because each of these rare experts can now do what used to require an entire team, faster, better, and cheaper.

And because we’re so early in our journey to build the EM foundation model, working directly with our customers lets us learn faster and improve our models based on real-world needs.

To that end, we’re also pursuing research partnerships. This falls out of our model, too. As we scale the LFM from version one to version two and beyond, we need to decide which training data to generate next, and that depends on the problems we’re solving. Partners working on chip packaging need us to model different structures than partners working on superconducting quantum computing. They get early access to our model and our team for their specific problems, and we get the data we need to drive generalization. The research partnerships feed the Data Factory, the Data Factory feeds the model, and we eat more of the EM spectrum.

One counterintuitive move here is that we don’t plan to sell the model. We plan to publish it, and sell everything around it: the platform, the experts, and solutions to our customers’ problems. The model is what makes all of that possible, but it’s not the product. As Packy wrote in Power in the Age of Intelligence, “If your technology is so good, why aren’t you using it to compete?” Our product is: bring us your electromagnetic problem, and we’ll solve it.

Companies shouldn’t be bottlenecked by the fact that, as it stands, you need hundreds of millions of dollars to build the types of systems we can build. We should be able to almost AWSify expertise for them.

There’s a lot more I want to do here.

As our model improves, and we move from the C++ to Python and even Claude era of EM foundation models, I suspect our business model changes, too. We can sell the “designer” to companies and they can use it to generate their own IP. A sort of Golden Analog Silicon Goose. As that happens, the cost to manipulate the EM spectrum goes down, and humanity’s capabilities increase.

Over the longer term, my dream is to run Arena Physica as a modern Bell Labs, to use the commercial side of the business to fund a new kind of research network.

What if, once we’ve proven the foundation model works, we opened it up? We could give academics free access to our model and our compute. In exchange, when they use it to design novel structures or discover new phenomena, the IP flows back through Arena. We take a cut (maybe 20%, like an app store) and they keep the rest.

Right now, a professor working on some exotic antenna geometry has to write grants, wait for funding, hire grad students, and slowly iterate through simulations on whatever compute they can scrounge. What if instead they could just... use the model? Explore design spaces that would take years to search manually? And get paid when their discoveries become commercially valuable?

If our model can auto-generate IP for any electromagnetic application, we become the platform. The rare humans who can push the boundaries become contributors, and get rewarded for it as their rare knowledge turns scalable. Because the universe of people with this expertise is so small, we can actually get our arms around sharing the upside of their work with them.

Of course, as we expand across the EM spectrum, and our models become multiphysics models, we’re not going to hire all of the world’s great physicists. But by opening up the platform and becoming a new kind of Bell Labs, we can work with them.

I want Arena to be a place where academics can sabbatical in and access our models. Where we’re not just selling software, but funding experiments. Where we create incentive structures that let brilliant people do fundamental research without the soul-crushing grant cycle. Where we compress the cycle between physics research and application, and wield AI not to do what humans can do cheaper, but to do things humans can’t do at all today. If we win, humanity benefits.

But I’m getting ahead of myself. First, we need to take problems off of our customers’ plates.

What Can Our Customers Build?

So what could our customers build if we provide them best-in-class RF and EE?

To start, by working with Arena Physica, any company that wants to do anything with phased arrays can get custom ones, and much more quickly than they could have before.

What they’ll be able to build isn’t limited to what you’d traditionally think of as radar.

I heard a phrase once that stuck with me. I’ve always thought of radar as a detection system, but this guy told me that radar is actually an imaging system. As the frequency gets shorter, you get higher resolution. You can image things. So radar doesn’t just tell you that “something is there.” It also tells you what that something looks like. Radar can create images, like a camera.

Take drones. Everyone is talking about drones as the future of warfare, but currently, we can’t see them with radars or sensors. It might surprise you to learn that the United States Navy doesn’t have counter-drone phased array radar at scale. That surprised me too, so when we dug a little bit, what we heard was that in the current system, it’s just too expensive. To get Raytheon to build a new ship-based phased array radar for them would end up costing as much as half the ship.

It should not cost close to a billion dollars to make phased array radars. But remember that slow process of imagining and simulating designs that we discussed earlier? Now, imagine that happening inside of a slow-moving legacy prime that gets paid a margin on top of every dollar it spends. The result is that our Navy doesn’t have phased array radars that are good at detecting drones.

We can help the Navy see drones much more cheaply. And we should. For drone detection, you might have 1,000 incoming targets. The benefit of phased arrays is that you can form multiple beams from one antenna, like Starlink does, instead of something spinning like an old-timey radar or LiDAR. (As a side-note, those spinning LiDARs on top of Waymos will almost certainly become phased arrays, too, and when they do, the whole unit goes solid-state, which is cheaper and more reliable with no moving parts).

We also have people reaching out to us who want to design phased arrays for drone capture. They’re building systems that catch drones in mid-air with a robot arm. The drone capture systems are smart; they have onboard computing and sensors to actively track and intercept the incoming drone. They need to be able to see, which means they need custom phased arrays that can track a fast-moving drone with high precision, work in any visual conditions, and are small and cheap enough to put on the capture mechanism itself. They certainly couldn’t make the math work at a billion dollars, but we can bring those costs down at least an order of magnitude in the near-term with automated design and fast iteration.

The applications only multiply once you realize that you’re dealing with an all-condition imaging system, and that drones are just one application of phased arrays. The same physics that steers radar beams also steers communication beams.

If space stuff continues to grow the way everyone thinks it will, you’re talking about tens of thousands of satellites and millions of ground terminals. Every single one needs these precisely shaped silicon tiles. Every ground station needs phased arrays. Northwood just raised a ton of money to build phased array ground stations. Every Starlink antenna on a house and satellite is powered by phased arrays.

Now, let’s say an adversary is trying to jam your radar signal. Notice how these little things dive down to zero:

Remember I said you can dynamically move this beam? This is one of the benefits of the phased array. It lets you move those points. Those are called null points, and it’s where the interference pattern zeroes out, like your noise canceling headphones. So imagine I’m being jammed. What’s crazy cool about a phase array is that I can transmit my signal, and then I can cyclically move my null point to absorb the jam signal. Just think about it. It’s remarkable, actually. You’re still transmitting and receiving, but you’ve carved out a little pocket of silence right where the enemy is screaming at you.

But if you can exploit the physics really well—like, in this case, we’re still using our good old digital transistors to say what to do, while the analog shapes determine the quality of the EM fields produced—this is where Gordon Moore’s dream might come true: his silicon boolean transistors are talking to analog silicon. And that’s really cool. With digital silicon, we can compute. But with analog silicon, we can transmit power, we can transmit directed energy, we can absorb energy. Analog silicon makes it much more physical.

The same physics that lets you communicate also lets you deny communication to others. And the same physics that lets you transmit also lets you absorb. Stealth is just the inverse problem of radar: instead of bouncing signals back, you’re making them disappear.

It’s all shapes.

Imagine we could change the economic structure of all of this.

That is exactly what we are trying to do at Arena Physica. Our mission is to create “Electromagnetic Superintelligence.” It sounds audacious, but remember, it is much easier to achieve superintelligence—as in, relative to humans—in electromagnetism than it is in language or even math. And it describes precisely what we’re building: a system that develops superhuman intuition for how geometry shapes electromagnetic fields, a mind that can see what we can’t.

As a software engineer, I don’t know how to be an infrastructure engineer. That’s because I don’t need to. Amazon takes care of it for me. If you could say companies no longer need RF expertise, you could lower the cost of everything we’ve discussed dramatically by a factor of 10 or more. We could give RF capabilities to everyone, from small companies to those that serve the Navy.

One obvious ramification is we’ll probably see more satellite companies and space companies, because they can now design their own phased arrays. More competition in the radar space. More competition in the jamming space.

Another less obvious one might be backpack radars. Think about troops going into a situation like Ukraine. One of the big risks they face is drones sneaking up on them. They should have backpack-mounted counter-drone radar: small, cheap phased arrays that let every warfighter see what’s coming.

Backpack radars are a very specific thing, but the point is that they’re something that made no practical or economic sense before that becomes practically and economically feasible.

We can even help AI models get better in a sort of indirect way.

Data centers need to move enormous amounts of data between chips very fast. The problem is that, at those speeds, the wires connecting chips start acting like antennas. They can accidentally broadcast and pick up signals. Which means they hit bandwidth limits on chip-to-chip communication. So their goal is the opposite as ours: they’re trying to make really bad antennas. They don’t want the electron traveling between their GPU and CPU to pick up a signal or transmit one, because then the data gets garbled. This is called signal integrity. The solution is the same shaped-silicon approach: carefully designed structures that guide high-frequency signals without interference.

It’s not just chip-to-chip. Think bigger. A data center company recently asked us whether we could beam data rack-to-rack wirelessly, because the cabling itself is becoming a bottleneck to how fast they can deploy. It’s not just terrestrial, either. For orbital data centers, there won’t be any option. You’re not going to be running optical cables between racks in space.

The most interesting thing is how the market has responded by asking us for use cases we would never have thought of.

For example, high-frequency trading firms have reached out to us about helping them trade faster. I’d assumed fiber optics already transmitted at the speed of light, but due to total internal reflection inside the glass, signals travel at only 60-70% of light speed. A phased array transmitting through free space goes at actual light speed. Over the distance between New York and Chicago, that difference could be enough to make a lot of money.

If you look at what we’ve just described, there are really two different things happening. I think we‘re going to see a K-shaped future for hardware.

The lower leg of the K is making commodity devices dramatically cheaper and more accessible. For example: the Navy acquiring drone-detection radar without writing a billion-dollar check to Raytheon, satellite startups designing their own phased arrays instead of outsourcing to a prime, backpack-mounted counter-drone radar for every warfighter, and data centers deploying faster with wireless rack-to-rack links are capabilities that exist today in expensive, bottlenecked forms. We want to remove these bottlenecks and make the capabilities cheaper and more abundant.

The upper leg of the K is making exquisite devices at the frontier of what’s physically possible newly achievable. This is what Bell Labs enabled, entirely new and more powerful capabilities than were possible before their research. There is a lot of excitement about cheap, attritable systems in the future of warfare, and rightly so. But in the conversations I’ve had with military leadership, they believe that to win (or deter) a conflict in the Indo Pacific, we’re going to need some of the world’s most exquisite machinery. The F-117 was crucial to winning the Gulf War; it made up just 2.5% of coalition air power but destroyed 40% of all strategic targets. We want to make new types of exquisite hardware possible, for defense and beyond.

By making it easier, faster, and cheaper to design more capable RF components, we think we’ll help expand the market beyond current analyst estimates, which don’t anticipate the unimagined. Those estimates project a 12% CAGR for RF components, from $44.8 billion to $140.5 billion over the next decade. I think that’s wrong, almost certainly too low. One of the reasons they’re projecting relatively slow growth is that RF is just too hard today. But if you democratize the expertise and transform the cost structure, would everyone just add better all-weather sensing equipment onto their robot? Would everyone just add backpack-mounted radars to every soldier?

What I’m most excited about is this open possibility space. I don’t even know what else people might think up now that they have the ability to manipulate this stuff. If we’re right, their ambitions won’t be limited by speed, economics, or even by the shapes required to deliver the capabilities that they need.

Alien Designs

A lot of the shapes coming out of our system already look nothing like what a human would design. Basically, we’re working with near-alien humans to create a system that will ultimately produce alien designs.

Human RF engineers have been trained on certain canonical structures: dipoles, patches, spirals, horns. They know these shapes work because they’ve been refined over decades. When they design something new, they start from these familiar forms and tweak them.

Our system doesn’t care about any of that. It starts from noise and evolves toward function. The results often look like QR codes, or random stippling, or structures that seem to follow no logic at all.

When we show these designs to expert RF engineers, their first reaction is usually skepticism. “That doesn’t look like an antenna.” “I would never have come up with that.” “Are you sure this works?”

Then, we fabricate it.

Today, we’re fabricating at the PCB board level. The goal I have for the team is that we’ll do our first silicon tapeout this year. As in, we’ll manufacture actual silicon chips. Analog silicon has the advantage that we don’t need TSMC’s bleeding-edge fabs; older, cheaper factories like Samsung, Global Foundries, and some defense fabs can do it, because the node size is typically larger.

And it works. This is the AlphaGo moment. Remember Move 37 expert commentary?

We’re seeing the same pattern. Engineers look at our designs and say, “That’s super unconventional. That’s unusual. That’s not in the textbook.” And then it works. And then they say it was “creative.”

Our goal is to do more than just match human performance autonomously. We want to exceed it and find all of the EM equivalents of Move 37: designs so counterintuitive that no human would have tried them, but so effective that they outperform anything we would have tried.

My hunch is that this will happen fairly quickly in EM because of what we discussed at the top: most humans didn’t evolve to intuit which shapes produce which EM waves. Therefore, we are terrible at intuiting the answer. On the other hand, computers can become superhuman — we can evolve them to get there in simulation.

There’s this great story in Skunk Works, Ben Rich’s personal memoir from his time atop the famed Lockheed division, about how the F-117 stealth bomber came to be.

A thirty-six-year-old mathematician and radar specialist named Denys Overholser happened to read a translation of a dense technical paper by Pyotr Ufimtsev, the chief scientist at the Moscow Institute of Radio Engineering, titled Method of Edge Waves in the Physical Theory of Diffraction. Ufimtsev had “revisited a century-old set of formulas derived by Scottish physicist James Clerk Maxwell… these calculations predicted the manner in which a given geometric configuration would reflect electromagnetic radiation” and took them a step further. “Ben,” Overholser told Ben Rich, “this guy has shown us how to accurately calculate radar cross sections across the surface of the wing and at the edge of the wing, and put together these calculations for an accurate total.”

With Ufimtsev’s work, the Skunk Works team could create computer software to calculate the radar cross section (how visible an object is to radar) as long as the shapes were in two dimensions. If they designed the bomber as thousands of flat triangles, they could add them all up and get the radar cross section.

That’s exactly what Overholser did, and the design that emerged “was a diamond beveled in four directions, creating in essence four triangles,” which, viewed from above, “closely resembled an Indian Arrowhead.”

He called it the Hopeless Diamond, and calculated that it would be “one thousand times less visible than the least visible shape previously produced at the Skunk Works.” On a radar screen, it would appear to be the size of an eagle’s eyeball.

Kelly Johnson is the Skunk Works founder and boss who was so magnificent at airplane design that his boss said of him, “That damn Swede can actually see air.”

He was so unimpressed by the design that, when he saw it, he physically kicked Rich in the butt, crumpled the proposal, threw it at Rich’s feet, and stormed, “Ben Rich, you dumb shit. Have you lost your goddamned mind? This crap will never get off the ground.”

As it turned out, the model was right and even the great Kelly Johnson was wrong. The Hopeless Diamond became the F-117 Nighthawk, the stealthiest plane built to date by more than three orders of magnitude, and flew over 1,300 sorties during the 1991 Gulf War without a single combat loss.

The F-117’s was an alien geometry, and it gave the US otherworldly capabilities.

We’re attempting to do something similar, but at silicon scale (to start). And with AI doing the search instead of human engineers sketching on whiteboards, we’re planning to do it over much wider problem spaces and search spaces than currently exist.

The Shape of Things to Come

We have a lot of work ahead of us to solve the core problem: building a Large Field Model that can do for EM what LLMs did for language.

The simulation has to get faster. The generative model has to get smarter. The fabrication loop has to tighten, and we need to actually fab analog silicon, and then we need to do it for a lot of customers. We need to hire more of those rare humans who think like electrons and work with them to create training data.

It’s hard to see what our models can do today and not imagine what they might be able to do in the future, though.

If it scales across the entire EM spectrum like we think it will, things will get very interesting.

How interesting? Packy asked me whether our models might one day help finally produce a Grand Unified Theory, if one is possible.

I don’t know. But here’s how I think about discovering new physics more broadly.

All of our tools today — since the dawn of computers — have been about deduction. Solve, compute, calculate, predict. Here’s an input; tell me what happens. But a lot of really creative human reasoning is inductive. You postulate a question or an idea, and then you research it. That’s deeply creative and deeply human.

This connects to something personal for me. My undergraduate research advisor at Stanford, Hari Manoharan, did an experiment that made the cover of Nature in 2000.

He arranged 80 cobalt atoms in an ellipse on a copper surface. You know those whispering galleries where you stand in a corner and whisper and someone on the other side can hear you? That’s constructive interference of sound waves. Hari knew that electrons also behave like waves, and he suspected they should interfere in the same way.

And that’s exactly what he saw. A “quantum mirage.” A ghost atom appearing at one focus when a real atom sat at the other. What blew everyone’s mind was that there was no time delay. Physics predicted some tiny delay to account for information traveling at the speed of light. Instead, it was instantaneous. This kicked off a whole field of quantum communication research.

Hari knew the Schrödinger equation. He knew Maxwell’s equations. Everybody knew those equations. But he combined them in a way no one had thought to try, postulated what might happen, and built an experiment to test it. That’s how new physics happens.

How amazing would it be if we had machines that could actually understand these equations and help us? Currently, we have to wait decades for a genius to come along and push a field forward. What could we benefit from if those geniuses had a little help? How much closer could we pull the future? How much more of the universe could we understand in our lifetimes?

What if our foundation model could start making those leaps? Right now, it’s not yet breakthrough inductive, meaning it can’t create its own experiments. It’s like an applied physicist: we can tell it our engineering goal and it will come up with a design. But as it learns more, develops something like intuition, could the model start postulating? Could it notice patterns humans have missed and suggest experiments?

And maybe (this is where I let myself dream), maybe if we build foundation models for each of the fundamental forces, and they start talking to each other, we get closer to something bigger. There are four fundamental forces: strong, weak, gravity, electromagnetism. We’re just trying to make a dent in one of them. But physics is deeply interconnected. Hari’s quantum mirage happened because electromagnetism and quantum mechanics intersected in a way that nobody expected.

What happens when foundation models that understand multiple forces chain together, exploring the spaces between them? I can’t stop thinking about this.

Ultimately, that’s what this is about. It’s why I pursued my PhD in quantum electromagneticism for four years (then dropped out in true Silicon Valley fashion) and why I started Arena Physica.

I want to understand the nature of reality in order to manipulate it for the betterment of humanity. To do that, we need models that learn directly from physics, that develop intuitions we’ve never evolved to have.

One of the questions that came up in grad school, and I remember thinking, how do people even ask these questions, was: Why does physics work? Why does math work? Isn’t it strange that reality is so... describable? Why isn’t it much more random?

Maybe these models will help us find out.

Electromagnetism secretly runs the world. We’ve been manipulating it for a century with our hands tied behind our backs, limited by the rarity of humans who can see what we cannot see.

We’re creating something to understand the interplay between geometries and electromagnetic waves, and evolving them to develop a new intuition.

The universe is made of fields. Fields are shaped by geometry. Geometry, it turns out, is something computers can learn much better than we can. We should lean on them so that we can get to new problems.

If we learn to shape the waves, we might be able to shape the future.

Big thanks to Pratap and the whole Arena Physica team for sharing their knowledge, and to Badal for the cover art.

That’s all for today. We’ll be back in your inbox with a Weekly Dose on Friday.

Thanks for reading,

Packy

Hi friends 👋 ,

Happy Friday and welcome back to our 185th Weekly Dose of Optimism. Fresh off the heels of my World Models primer with General Intuition’s Pim de Witte…

… I’m excited to bring you my favorite Dose in a long time. Kalanick, Bezos, copper, cancer drugs, meditation stimulation, and the return of Ulkar. I can’t wait any longer…

Let’s get to it.

Today’s Weekly Dose is brought to you by… Framer

Framer gives designers superpowers.

Framer is the design-first, no-code website builder that lets anyone ship a production-ready site in minutes. Whether you’re starting with a template or a blank canvas, Framer gives you total creative control with no coding required. Add animations, localize with one click, and collaborate in real-time with your whole team. You can even A/B test and track clicks with built-in analytics.

The weekend starts today, and the weekend is the perfect time for side-projects. Get to it.

Launch for free at Framer dot com. Use code NOTBORING for a free month on Framer Pro.

Just Publish it With Framer

(1) Travis Kalanick is Back with Atoms

Travis Kalanick &

There’s a version of the story where Travis Kalanick disappears into the wilderness after his investors ousted him from Uber. He’s rich. He doesn’t need to deal with this shit.

If you watched the Uber saga, though, you knew that wasn’t how the story was going to play out, and luckily, you were right. Travis. Kalanick. Is. Back.

Kalanick spent nearly eight years in what might be the most extreme version of stealth mode any modern founder has pulled off: he’s hired thousands of employees, in 30 countries, bought and developed hard real estate assets, built a full-stack food infrastructure business, and did it all like lasagna. Employees weren’t even allowed to list the company’s name anywhere.

The company was called City Storage Systems. Its most visible subsidiary, CloudKitchens, operated ghost kitchens. It was valued at $15 billion in 2022. And it was just, to use a food term, a little amuse-bouche before the main course.

Last week, Kalanick unveiled Atoms: a robotics company spanning food, mining, and transport, built on everything CloudKitchens learned about physical automation. CloudKitchens is now Atoms Food, which includes Lab37's Bowl Builder robot (200 meals per hour, no humans), the Otter restaurant OS, and Picnic delivery.

With Atoms, TK is expanding into mining (via the acquisition of Pronto AI, an autonomous haulage startup for mines and quarries founded by Waymo founder Anthony Levandowski) and transport, where Atoms is building what it calls “a wheelbase for robots.”

Kalanick’s robotics bet is explicitly anti-humanoid. It’s the same bet that we wrote about in Many Small Steps for Robots, One Giant Leap for Mankind with Standard Bots’ Evan Beard and in yesterday’s World Models: Computing the Uncomputable with General Intuition’s Pim De Witte. “The recent humanoid Olympics in Beijing highlighted many advances in humanoid development,” Kalanick writes in the company’s Vision doc, “I watched the half-marathon and couldn’t help but think how much better it would be if they just had wheels.”

Uber was Kalanick’s first attempt to digitize the physical world. Atoms seems to be the rest of it.

It’s a bad idea to bet against Travis Kalanick. Watch the TBPN interview to get a taste of why. Which means it’s a good idea to bet on our autonomous everything future.

(2) Jeff Bezos Raising $100B AI Manufacturing Fund

The Wall Street Journal

THE BOYS ARE BACK IN TOWN. First Kalanick, now Bezos. Elon’s been living here.

Imagine going out to raise one of the largest funds in history, a SoftBank Vision Fund-sized whopper of an investment vehicle, one that exceeds the total amount of US venture capital funds raised in 2025… and that is less than 50% of your personal net worth. Jeff Bezos doesn’t have to imagine.

The WSJ reported that Bezos is looking to raise a $100B fund “to buy companies in major industrial sectors such as chipmaking, defense and aerospace.” The plan is to buy existing manufacturing companies, and then use Bezos’ new startup, Project Prometheus, which sounds like a World Models company (“building AI models that can understand and simulate the physical world”), to boost efficiency and profitability.

Basically, the fund is to AI rollup funds targeting accounting firms as new Jeff Bezos is to scrawny old Jeff Bezos. It’s the gigachad version of the AI rollup playbook.

Details are sparse, but if you told me a couple of years ago, in the depths of the bear market, when everyone said that America couldn’t build things anymore, that Jeff Bezos, Travis Kalanick, and Elon Musk were going to be competing to industrialize America harder, all while Chris Power launches a new Hadrian factory in Alabama, I would have told you… welcome to The American Millennium.

(3) Copper One: The World’s Only Autonomy-First Mine and Refinery

Mariana Minerals

Wait… and we’re mining and refining autonomously here, too?

Long time not boring readers will know that I am a big fan of mining startups, given how critical (get it) the industry is to the modern world and how critically underinvested-in it’s been. I think sell side analysts will write poems about Earth AI’s business model one day.

It’s also a unique industry in that it’s structurally not winner-take-all. Earth AI can build a very, very big business while Kobold, Durin, and Mariana Minerals do, too. If you have enough high-grade metal in your deposit and can get to it economically, you’ll make money.

So I was pumped to see Mariana announce Copper One, its first autonomy-first copper mine and refinery in Utah. The blog post in which they announced the project is one of the best company announcement blog post’s I’ve read in a while, with clear explanations and great graphics, so I encourage you to go read it. Here, I’ll just share the plan they laid out to turbocharge the existing copper mine they acquired last year

1. Deploy PlantOS at scale to maximize copper recovery and reduce copper refining costs throughout heap leaching, solvent extraction, and electrowinning.

2. Restart mining operations with autonomous equipment and orchestration via MineOS.

3. Integrate copper scrap processing into the refining circuit, leveraging PlantOS to manage feedstock variability and to put a meaningful dent in US copper scrap exports.

4. Scale combined output at the site to 50,000 metric tonnes per year from both geologic and scrap feedstocks (leveraging CapitalProjectOS to accelerate capital project delivery).

We’re going to need to find a lot of metals and minerals, we’re going to need to get them out of the ground and refine them more efficiently, quickly, and cleanly. I’m excited to see what improvements Mariana can dig up in Utah.

(4) Scientists Create Cancer-Fighting Immune Cells Right in the Body

University of California San Francisco

Long time not boring readers will also know that we have long taken a bold but important staunchly anti-cancer stance here in the Dose. We just don’t like cancer, and we want to see it gone.

’s piece on Sid Sijbrandij’s extraordinary care journey, Going Founder Mode on Cancer, remains our favorite in the anti-cancer canon. It tells the story of one person’s against-the-odds battle to bend the medical system and cure his own cancer. But it shouldn’t have to be that hard. Enter this week’s entry.

CAR-T therapy is one of the most powerful weapons against cancer. It works by pulling a patient's T cells out of their body, genetically reprogramming them to hunt cancer, growing them up in a lab, and infusing them back in. Seven CAR-T therapies are now FDA-approved for blood cancers. The problem is, the process takes weeks, costs over $400,000, requires specialized manufacturing facilities, and demands lymphodepleting chemotherapy just to make room for the new cells. Most cancer patients in the world will never have access to it.

This week, a team at UCSF led by Justin Eyquem published a paper in Nature showing they can skip almost all of that. Instead of the extract-engineer-expand-reinfuse pipeline, they designed a two-particle injection system that reprograms T cells inside the body. One particle delivers CRISPR-Cas9 to make a precise cut in the T cell’s genome. The second delivers the DNA template for a chimeric antigen receptor (the cancer-targeting weapon). The whole thing is designed so only T cells get edited, and only at a specific, safe genomic location, avoiding the random integration that can, in rare cases, cause secondary cancers.

In mice with humanized immune systems, a single injection cleared detectable leukemia in nearly all animals within two weeks. The engineered cells made up as much as 40% of T cells in organs like bone marrow and spleen. The approach also worked against multiple myeloma and sarcoma, a solid tumor that is historically much harder for CAR-T to crack. Best of all, the in vivo cells actually outperformed lab-manufactured ones, because cells that never leave the body retain their “stemness” and ability to keep dividing.

As always, it’s important to remember that mice are not people, and we need to see this thing work in humans before we start popping the champagne. To that end, Eyquem and his collaborators founded a company, Azalea Therapeutics, to push toward human trials. If it translates, this could turn CAR-T from a last-resort therapy available at a handful of elite cancer centers into something closer to a vaccine, a single injection that any hospital could administer.

Just a quick shot, a follow-up, and back at it. Say it with us… get fucked, cancer.

(5) Facilitating Mindfulness Training with Ultrasonic Neuromodulation

Brian Lord, Erica N. Lord, Jessica Schachtner, Laura Beaman, Shinzen Young, John J. B. Allen, Joseph L. Sanguinetti

A little while back, I read Shinzen Young’s The Science of Enlightenment. It was one of the best books I’ve read on meditation. Shinzen is a Jewish-American who trained as a Shingon monk in Japan, did deep practice in all three major Buddhist traditions (Vajrayana, Zen, Vipassana), and then came back to the West to fuse contemplative practice with scientific rigor.

Most relevantly, near the end of the book, he makes a wild prediction about Maitreya, the prophesied “future Buddha” of Buddhist tradition. Maitreya won’t be a person, Shinzen argues. It will be a collective of scientists, technologists, and contemplatives working together to make liberation accessible at scale through tools and systematic methods, rather than requiring every individual to spend decades in a monastery.

This week, Shinzen and a team of researchers dropped a preprint on bioRxiv that might be a step in that direction.

After dozens or hundreds of hours of sitting, experienced meditators have brains that look different from everyone else's. Their default mode network (DMN, the brain's self-referential chatter machine, the thing running rumination loops, and the thing that psychedelics seem to quiet) decouples from the central executive network, the system that handles focused attention. That decoupling is the neural signature of equanimity: letting experiences arise and pass without getting caught up in them. It's associated with reduced stress, lower depression, and generally being a calmer human. But it takes hundreds of hours to get there, and most people quit long before anything rewires, because the early stages are so uncertain and even boring.

What if you could use technology to do the same thing, much faster?

Shinzen and the team at the University of Arizona ran a randomized controlled trial with 24 meditation-naïve participants who did a two-week mindfulness program. Half received transcranial focused ultrasound targeting the posterior cingulate cortex (the hub of the DMN) during four in-person sessions. Half got sham stimulation.

After two weeks, the active group's brains showed the decoupled network pattern that normally takes experienced meditators hundreds of hours to develop (p < 0.001). The sham group's connectivity actually increased; they got more tangled up, not less, which is exactly what frustrated novices do. Within the active group, greater decoupling also predicted bigger increases in self-reported acceptance and longer voluntary meditation sessions. The ultrasound made the practice better and gave novices experienced meditator brains.

It's a small study and a preprint, so it needs replication, but as someone who has spent a lot of hours trying to go deep, it also seems remarkable.

I love this because I’ve written that the worst outcome would be for us to get all of the technological wonders we could have asked for and still be unhappy, and I think meditation, and the associated ability to pay attention, can help there.

It’s also just very cool to see someone who spent his entire career arguing that the cross-fertilization of Eastern contemplative technology and Western science would eventually produce something neither could produce alone being proven right.

“The next Buddha is a sangha,” indeed.

EXTRA DOSE:

1. Scientific Breakthroughs with Ulkar Aghayeva

2. My Favorite Essay of the Week

Read more

Subscribe now

Hi friends 👋 ,

Happy Wednesday!

A few months ago, Pim De Witte and Kent Rollins invited me to their office right here in New York City to show me what they’ve been cooking up at General Intuition. I’d heard about the company, from the announcement of their leet $133.7 million Seed round, and I’d heard about the class of product they were building, World Models, but I didn’t know much beyond that.

What they showed me that day, models that learn to predict the near future from action-labeled gaming clips, and what I’ve learned from many conversations and dozens of hours of research since, has changed my perception of what models can do. I am on the record as being skeptical that LLMs will take us to superintelligence, but I think there is a real shot that World Models will drive superhuman, complementary machines that do things that we can’t, or don’t want to, do.

Since that first meeting, the World Models space has heated up. Fei-Fei Li’s World Labs raised $1 billion. Yann LeCun’s AMI raised $1.03 billion. World Models were one of the stars of this week’s NVIDIA GTC. But the field is so nascent and there is so much going on, so many geniuses pursuing competing and collaborative approaches, that it’s hard to make sense of it all.

So I asked Pim to team up with me on a co-written essay about the history, theory, progress, and potential of World Models. He agreed, and both he and the General Intuition team have been incredibly generous with their time and human intelligence in helping me get up to speed, so that I can help you get up to speed.

I have the coolest job in the world. Over the past couple of months, I’ve gotten a front row seat to the future of embodied AI, of Models and Agents, trained in dreams, that direct machines to do things for us in the physical world.

I’m thrilled to share the fruit of that exploration, what I think is the most comprehensive guide to World Models that exists. Obviously, Pim and the GI team have a perspective on the best way to build World Models, but I was impressed with how careful they were to present the pros and cons to every approach, including theirs, and with their admission that the future is not yet determined.

The space continues to change and progress incredibly fast. I hope this will help you navigate and make sense of all of the exciting news that continues to drop.

Let’s get to it.

Today’s Not Boring is brought to you by… Framer

Framer gives designers superpowers.

Framer is the design-first, no-code website builder that lets anyone ship a production-ready site in minutes. Whether you’re starting with a template or a blank canvas, Framer gives you total creative control with no coding required. Add animations, localize with one click, and collaborate in real-time with your whole team. You can even A/B test and track clicks with built-in analytics.

Launch for free at Framer dot com. Use code NOTBORING for a free month on Framer Pro.

Just Publish it With Framer

World Models: Computing the Uncomputable

A Co-Written Essay with Pim De Witte

“I wanted to fall asleep last night. Instead, I started imagining all of the scenarios I might run into the next day, and how I might react to them.”

This is a common experience. As humans, we imagine easily, whether it’s complex sports stadiums, potential romance, or heated discussions. We don’t have to work harder to imagine ourselves at the next Manchester United game than we do to imagine talking to a friend we’ve known for years, even though imagining a Manchester game includes simulating and modeling the behavior of thousands of people, something that would take years for traditional computers and game engines today1.

Think about writing the code to describe the Man U match: at any moment, a fan might bring a random, home-crafted flag. The entire stadium starts singing a song related to it. Only some will sing, though; others will jump with their kids, while an old couple sits still, wondering if this is their last game together, soaking in every second in silence.

The world is a place where unexpected futures unfold, but in somewhat predictable ways. As humans, we can envision almost all of them with roughly the same amount of effort with a very similar amount of time given to each thought. Computers can’t.

It’s no wonder traditional computing struggles with this complexity. Imagine anticipating and coding each and every action, as well as the interactions between all of those actions. Mathematically, in a traditional engine, simulating N fans is at least an O(N) or O(N²) problem. Each person, flag, chair, and ball must be explicitly calculated — and really, the interactions between them need to be calculated, too.

In robotics, machines must respond to situations in the real world in the same amount of time, regardless of their complexity, even though, in traditional computing, different situations can take wildly different amounts of time to simulate. This has been a major bottleneck for robotics and embodied AI progress.

World Models are a solution to that problem.

World Models learn to predict those dynamics from video and, often, the actions taken in them. They reduce situations that are dynamic and computationally difficult to simulate at scale — including stochastic, action-dependent group behavior like soccer games — into a single fixed cost operation in a neural network.

In a World Model, the entire stadium is simulated as a fixed cost forward pass through the neural network. The complexity of the scene doesn’t exponentially slow down the ‘engine’ during inference because the weights have already absorbed the patterns of the world in training.

How? Actions.

Actions act as a form of compression to predict unfolding dynamics: they hold the information to unroll future states in an environment, until more actions take place and add new inputs into the environment. Each action carries enough information to predict what happens next, until the next action updates the picture.

This action-conditioned approach allows models to learn and plan interactively. Today, this is intractable in even the best simulation engines, and definitely not at predictable compute costs. Actions help models interact with the world like we do.

Over and over again, every single day, you observe, you compute, you decide what to do, you act. This is life. At any point, all gathered information about space and time collapses into the action you take.

For computers, actions are a cheat code around the costs of simulation. If human brains are much more efficient than best-in-class LLMs, then we can get all of that computation practically for free by observing how humans respond to the countless variables in their environments. This gives us a way to do non-deterministic computing efficiently and create simulations that shouldn’t be possible under traditional compute constraints.

This ability to compute the uncomputable is why we believe World Models will unlock progress in embodied AI in a way that current model architectures can’t.

Think about models like dreams.

Have you ever had a dream where you simply stood and watched what was happening without the ability to intervene? That’s a video model.

The real world is different. It responds to what you do or instruct to do, and predicts the full range of things that could happen as a result, not just the single most likely or most entertaining next frame.

Have you ever had a lucid dream in which you were able to shape the story inside the mind-generated dreamscape? That’s a World Model.

I coded up a comparison that you can play with here.

More formally, while a standard video model predicts the next frame based on probability, P(x_t+1 | x_t), a World Model predicts the next state based on intervention, P(s_t+1 | s_t, a_t).

That a_t, the action at time t, is the magic.

At General Intuition, we believe (and are seeing early signs) that World Models are a new and potentially more powerful class of foundation model than LLMs for environments that require deep spatial and temporal reasoning. Environments like our real world.

World models — these systems that learn from watching the world and the actions taken in it — are a fundamentally new kind of foundation model. They can compute what was previously uncomputable.

They will matter far more than anyone currently realizes, because they offer a path to general intelligence that language and code alone cannot. Being human, after all, is spending a lifetime taking actions based on what we experience, observe, and learn.

Pause. You might be confused by that claim, that World Models offer a path to general intelligence that LLMs cannot. Understandably so.

World Models are getting a lot of attention as of late. Yann LeCun, who has been skeptical that LLMs are the path to general intelligence, just announced that he raised $1.03 billion for AMI. Fei-Fei Li’s World Labs has also raised more than $1 billion to pursue World Models. Google DeepMind, which has the closest thing to an infinite money printer in tech, is betting money on World Models too. But what we’ve seen so far from that investment are cool videos and 3D worlds.

LLMs can quote Shakespeare and solve Erdős Problems. World Models, on the other hand, still seem more like a path to the Metaverse than a path to general intelligence.

But part of the reason World Models don’t yet have the hype of LLMs is that their definitions are still shaky.

What are World Models? We’ve already said that video models don’t fit the definition. 3D space models don’t, either. That said, both may be paths to World Models. Are the models that animate robots today World Models? Not really, although some are, and even the ones that aren’t share features with World Model architectures.

As always, hype adds to confusion. “My prediction is that ‘World Models’ will be the next buzzword,” Alexandre LeBrun, the CEO of AMI Labs (which is definitely a World Model company) told TechCrunch. “In six months, every company will call itself a World Model to raise funding.”

Hype is a small part of it. What we — and everyone else building in this space — believe is that World Models are the path to controlling machines in the physical world. There are differences in what we believe this path will look like. But all of us believe that the future runs through World Models.

“...very few understand how far-reaching this shift is…,” NVIDIA Director of Robotics and Distinguished Scientist Jim Fan said recently. “Unfortunately, the most hyped use case of World Models right now is AI video slop (and coming up, game slop). I bet with full confidence that 2026 will mark the first year that Large World Models lay real foundations for robotics, and for multimodal AI more broadly.”

Today, we’d like to welcome you into the group of the “very few” who “understand how far-reaching this shift is.” We are going to share the history of World Models, the state of the field as it stands today, broad explanations of the approaches each major lab is taking, and the convictions that drive General Intuition’s directions.

Whether you come with us is up to you. You take the blue pill, the story ends. You wake up in your bed and believe whatever you want to believe. You take the red pill... you stay in Wonderland, and we show you how deep the rabbit hole goes.

For example…. how can you be sure that you’re not an Agent operating inside of a World Model yourself?

Can Agents Learn Inside of Their Own Dreams?

Wake up, Neo.

World models aren’t a new idea. They are one of our oldest. Since humans gained the ability to think about our place in the universe, to ask why we are here, we have pondered whether our reality is just a simulation.

In 380 BC, Plato, via Socrates, offered The Allegory of the Cave. Imagine human beings who live underground in a cave, necks chained, forced to look ahead at the shadows on the wall. Those humans would believe those shadows to be reality, when in fact they are mere shadows of reality. This was Plato’s metaphor. He suggests that we are all stuck in the cave, necks chained, mistaking our perception for true reality.

Eighty years later, Chinese Daoist philosopher Zhuangzi contemplated similar questions in a passage of his Butterfly Dream:

Once Zhuang Zhou dreamt he was a butterfly, a butterfly flitting and fluttering around, happy with himself and doing as he pleased. He didn’t know he was Zhuang Zhou. Suddenly, he woke up and there he was, solid and unmistakable Zhuang Zhou. But he didn’t know if he was Zhuang Zhou who had dreamt he was a butterfly, or a butterfly dreaming he was Zhuang Zhou. Between Zhuang Zhou and a butterfly there must be some distinction! This is called the Transformation of Things.

As the centuries passed and our technological capabilities evolved, sci-fi writers joined the long lineage of thinkers inquiring about the true nature of reality. Frederik Pohl’s 1955 The Tunnel Under the World. Daniel F. Galouye’s Simulacron-3. Stanislaw Lem’s Non Serviam. Vernor Vinge’s True Names. William Gibson’s Neuromancer. Neal Stephenson’s Snow Crash. All painted textual pictures of simulated worlds.

During a 1977 speech in Metz, France, sci-fi legend Philip K. Dick confidently told the audience: “We are living in a computer-programmed reality, and the only clue we have to it is when some variable is changed2, and some alteration in our reality occurs.”

Your first interaction with the simulation was probably The Matrix. Ours was. In the original script for The Matrix, the Wachowskis conceived of the Matrix as a simulation collectively produced by human brains chained into a neural network.

The studio thought humans-as-computers was too confusing a concept for mass-market audiences, so they made the thermodynamically questionable decision to turn humans into batteries that powered the simulation. That was probably the right commercial call. The Matrix franchise has done nearly $2 billion in worldwide gross. More impactfully, it introduced the masses to the idea of a simulated world generated indistinguishable from the “real” one.

It’s no wonder that this idea has taken hold of our collective imagination. It’s certainly the right kind of weird but it’s also surprisingly hard to disprove. If the observations are the same, and the actions are the same, then the computation is the same. If what you see is the same and what you do is the same, it doesn’t matter whether you’re in a simulation or reality. It doesn’t matter whether you’re walking down a real street or a simulated one. Your brain processes both identically. Neo had no idea he was in the Matrix until Morpheus woke him up.

Christopher Nolan, throwing audience confusion to the wind — savoring it, even — released Inception3 in 2010. Dreams within dreams within dreams.

Nolan’s central premise is that the dream is a controllable space from which information can be extracted or, more importantly, into which information can be implanted.

But it’s all just sci-fi, right?

In 1990, Jürgen Schmidhuber, a young researcher at the Technical University of Munich, published Making the World Differentiable.

The paper proposed building a recurrent neural network (RNN), a neural network with two jobs: first, learn to predict what happens next in a simulated world and second, use that simulated world to train an Agent to act in it.

The Agent wouldn’t need to interact with a “real” environment at all. It could learn inside the model. Inside a dream.

The following year, Richard Sutton, of Bitter Lesson fame, dreamt up a similar idea. In Dyna, an Integrated Architecture for Learning, Planning, and Reacting, he argued that learning, planning, and reacting shouldn’t be separate systems. They should be unified in a single architecture. Which would mean that it’s technically possible to build a model of the world, practice inside it, and transfer what you learn back to reality.

Both papers were visionary. They would have a lasting impact as progress in the field enabled the researchers’ visions to become reality. But coming when they did, both papers may as well have been sci-fi.

In 1990, the world had something like 100 trillion to 1 quadrillion times less compute than we have today. Back then, the entire world had maybe 10-100 gigaFLOPS of total capacity. Tens of zettaflops (10^22 FLOPS) of computing power were sold in 2024 alone. In 1990, the global digital datasphere was approximately 10 petabytes, a volume so small it could barely hold 0.005% of the video data we now use for a single training run. By 2026, that volume has exploded by a factor of 22 million to 221 zettabytes.

But technology improves, and the most powerful dreams do not die.

Nearly three decades later, in March 2018, David Ha (then at Google Brain) and Schmidhuber published a paper titled World Models.4

The paper asked: Can agents learn inside of their own dreams?

To answer their own question, Ha and Schmidhuber built a fictional system with three components: a vision model (V) that compressed raw pixel observations into a compact representation, a memory model (M), a recurrent neural network that learned to predict what happens next, and a tiny controller (C) that decided what to do based only on V and M’s outputs.

The World Model was V + M: it could take in observations and imagine plausible futures. The controller was the Agent or policy: it chose which actions to take.

The paper joined in conversation with those centuries of thought experiments, novels, and movies. A dream might be reality, reality might be dreams. But what if we could actually act in our dreams? What would that do to reality?

Ha and Schmidhuber trained their World Model on observations from a car racing game and a first-person shooter game. The World Model generated new digital worlds. Then, they let the Agent practice entirely inside the World Model’s hallucinated dreams. Afterwards, they transferred the learned policy back to the actual environment.

And... it worked. The Agent could solve tasks it had never encountered in reality. The dream was real enough.

It was shocking, from a computer science perspective. But was it really so surprising? Isn’t this how humans navigate the world?

Ha and Schmidhuber noted that humans constantly run World Models in their heads. A baseball player facing a 100 mph fastball has to decide how to swing before the visual signal of the ball’s position even reaches their brain. The reason that every at-bat doesn’t result in a strikeout is that batters don’t react to reality, but to their brain’s “internal World Model’s” prediction of where the ball will be.

Donald Hoffman, Professor of Cognitive Sciences at University of California, Irvine, takes that idea a million steps further. He believes that we all walk around wearing “reality headsets” that simplify the staggering complexity of the quantum world into a user-friendly interface. Reality is too rich, so we navigate it via a sort of persistent waking dream.

This rabbit hole goes as deep as you want it to. But it’s World Models all the way down.

Ha and Schmidhuber showed that computers might be able to approach the world like we do: creating simulations to predict future states based on actions, acting based on those predictions, updating, and looping.

Actions, not words.

Language is Not Enough (Neither is Code)

Let’s play a game.

Clap your hands five times.

Now, instead of physically clapping your hands, I want you to describe clapping your hands using just words.

Where they are positioned in space, where they are relative to each other, by the picosecond. The points of contact. The sounds. What your hands look like as they move closer to each other, make contact, and pull apart. How they squish each other. What happens to the air between your two palms. What you see while your hands clap. Don’t forget your arms. How do they bend to facilitate the claps? Remember to do this by the picosecond, too. How does the fabric on your sleeve respond? What is happening in the background? Did the person next to you notice you clapping? How did they respond? Did you get fired for clapping in the middle of the meeting, following the instructions of an essay you shouldn’t have been reading while you should have been paying attention to work? Describe to me the vein on your boss’ forehead. Is it popping?

You can’t, can you? OK, stop. The point is made.

Language is an incredibly lossy compression of reality.

Language is important, of course. It is how we communicate and coordinate. The game Charades illustrates that to communicate ideas, language can be much more efficient than actions. LLMs are important in that capacity. But language alone is not enough.

What about code? Code is a form of very precise language that makes machines do things.

I asked Claude to “code me a simulation of hands clapping five times in a realistic environment.” It built me this. Which looks very painful.

There is a belief that, with scale, language and code will be able to solve all spatial-temporal intelligence challenges and produce Artificial General Intelligence (AGI) or Artificial Superintelligence (ASI).

Some argue that code is the key to solving many real-world intelligence challenges because it can perfectly instruct all physical form-factors with precision.

We do not share that belief. A code-based simulation is a poor version of a dream. It is rule-bound and unable to handle the stochastic messiness of reality.

To know the world, you must interact with it.

In The Glass Bead Game (Das Glasperlenspiel), a novel by Herman Hesse that won him the Nobel Prize for Literature in 1946, readers are introduced to Castalia, a future intellectual utopia devoted to pure thought. At Castalia’s center is an elaborate game, the titular Glass Bead Game, that synthesizes all human knowledge into a single formal language. Players compose “games” the way one might compose a fugue. A move might link a Bach cantata to a mathematical proof to a passage from Confucius. The game is the ultimate abstraction: all of human culture compressed into symbolic manipulation.

The protagonist, Joseph Knecht, rises to become Magister Ludi, Master of the Game, the highest position in Castalia. But he grows disillusioned. The game, for all its beauty, is sterile. Castalia’s intellectuals have retreated so far into abstraction that they’ve lost touch with the world. They can represent reality with extraordinary elegance, but they cannot act in it.

Knecht ultimately decides he must leave Castalia, and becomes a simple tutor. He chooses the messy, embodied, unpredictable world over the perfect symbolic one. He dedicated his life to the Game, the mastery of which involves operating on a level of abstraction beyond words, something closer to world modeling. But it wasn’t enough. Symbols alone, without contact with reality, eventually run dry.

Large Language Models are our Castalians. They are exquisite manipulators of symbols, capable of drawing connections across the entirety of human textual knowledge. They can discuss physics, compose poetry, write code, and explain the rules of baseball. They are, genuinely, one of the great intellectual achievements in human history.

But they operate entirely in the realm of representation. They can describe clapping, but they cannot clap. They can talk about gravity, but they do not know gravity the way a toddler knows gravity. They do not learn, the way a body learns, through thousands of falls and stumbles, what “down” means.

Language models predict the next token extraordinarily well. The only problem is that tokens are like shadows on Plato’s cave wall. And you cannot code your way to a realistic stadium crowd any more than you can describe your way there.

The real world is — or was — uncomputable.

If language and code, two of mankind’s most powerful inventions, are inadequate to represent our world, what do we have left?

The Answer is World Models

World Models offer another approach on the path to AGI. They offer a path to compute the things that are, today, uncomputable. They learn from the messy contact with reality that Knecht sought.

World Models offer a way to do non-deterministic compute efficiently, and to run simulations that shouldn’t be possible under traditional compute constraints.

World models are not a replacement for LLMs. Language remains essential; text can be used to condition World Models, to tell them what scenario to imagine, what goal to pursue, to give them a long-term goal. The thinking and the doing work together. But the doing has to come from somewhere other than text.

Joseph Knecht must come down from Castalia.

Real intelligence must come from observation of the world; from understanding actions and their consequences; from the things that language can only point at.

The Dao that can be told is not the eternal Dao.

In the beginning was the Word. Then came humans, to act imperfectly and unpredictably.

Maybe this is the way of things. In the beginning were LLMs. Then came World Models.

What Are World Models?

A World Model simulates environments and responds when you act inside them.

More formally, a World Model is an interactive predictive model that simulates spatial-temporal environments in response to actions.

While LLMs predict the next word in a sentence, World Models predict the next state (as in, the immediate future), conditioned on the current state and control input.

More succinctly: LLMs learn the structure of language. World Models learn the structure of causality.

This is a simple definition of World Models. It is accurate, but it’s not enough to understand how World Models work. For that, you’ll need to know four things:

1. What World Models do,

2. How they’re built,

3. Why “action” is so important, and

4. The relationship between World Models and policies.

What World Models Do

Think about what happens when you catch a ball. Your eyes take in a scene: the thrower’s arm, the ball in flight, the wind, the sun in your eyes, all of it. From that flood of sensory data, your brain builds a compressed model of what’s happening and, crucially, what’s about to happen. It predicts the ball’s trajectory a few hundred milliseconds into the future. Then it sends a motor command to your hand. You catch the ball. The whole loop — observe, predict, act — takes a fraction of a second and involves no language or “thinking” whatsoever.

A World Model does the same thing, computationally. It takes in observations (often video frames, though it can use any sensory data), builds a compressed internal representation of the environment’s state, and predicts how that state will change in response to actions.

It is, in essence, a learned physics engine, but one that doesn’t rely on hand-written equations. Instead of calculating gravity, collision, and friction from first principles, it has watched gravity, collision, and friction billions of times and learned the patterns.

This makes World Models a powerful tool for building Agents, AI systems that act in environments. World Models help Agents in three ways:

1. They serve as surrogate training grounds. An Agent can practice inside the World Model (basically, inside a dream) and transfer what it learns back to reality. This is important for safety (some things should not be tested or trained in the real world) and cost or sample/data efficiency (real world data is expensive, costly to gather, not available, you need a lot of it, etc.).

2. They enable planning over longer time horizons. An Agent can “imagine” the consequences of different actions before committing to one, the way a chess player thinks several moves ahead, except here, the board can be any environment or the real world.

3. They provide rich representations of the world for Agents to learn behaviors from. An Agent trained on a World Model’s internal representations learns to “see” the world in terms of the features that matter for acting in it, rather than raw pixels.

For these three reasons, the promise of World Models is that they are a path towards generalization. If you can create worlds that respond to actions the way the real world does, you can use them to safely, economically, and efficiently train embodied agents that can act in any virtual world, or the real one.

To be clear, this is the massive question in World Models: whether the simulated environments are faithful enough to reality that you can train on them and have that training transfer to the real world or more generally, whether you can “pre-train in sim.” Increasingly, the answer seems to be yes.

Ai2, the Allen Institute for AI, is a non-profit founded and funded by the late Microsoft co-founder, Paul Allen. It does great open source research and tooling, including its recent release of MolmoBot, an “open model suite for robotics, trained entirely in simulation.”

“Our results show that sim-to-real zero shot transfer for manipulation is possible,” they tweeted.

Dhruv Shah, a Princeton professor and Google DeepMind researcher who worked on the project, shared: “Within the scope of easily simulate-able tasks, a purely sim-trained policy outperforms SOTA VLAs trained on thousands of hours of real data!”

It is a pretty astonishing finding. A big focus of ours, and of the broader World Models field, is to expand the scope of tasks that are easy to simulate.

This is how it works. First, World Models imagine realistic environments and future states, ideally that respond to actions or instructions in the way the real and virtual worlds they’ve been trained on do. Next, the Agents are let loose inside of the generated worlds to train. Then, the Agents are brought back into real environments and are tested on what they’ve learned.

This is what Ha and Schmidhuber demonstrated in 2018. It remains the central promise of the field.

How World Models Are Built

World Models are fairly young. No single approach or combination thereof has proved superior, which means that the final architecture for general World Models is still an open question. There are, however, repeatable ingredients for training.

Start with data; massive quantities of observation data. Often, observations are paired with the actions taken to produce them. This pairing can come about in several ways. Observations (typically video) are collected in advance and actions are either recorded alongside them, or inferred via another model after the fact. Alternatively, the model learns by taking actions itself, generating its own observations and action data through direct interaction with an environment.

When the training data is observations or videos, the raw frames serve as observations of an environment unfolding over time. These videos are ideally labeled with the actions that produced them (either because they were recorded together or inferred with a separate AI model). The actions provide the causal link: what someone did that made the environment change. A gameplay clip where a player turns left and the camera pans to reveal a hallway. A driving recording where the wheel turns and the car follows a curve. A teleoperation session where a robotic arm reaches and a cup moves. In each case, the model sees a before, an action, and an after.

When the model learns through interaction, the same structure applies — before, action, after — but the data is generated on the fly rather than collected in advance, and the actions come from the model’s own developing policy rather than from an external source.

The World Model’s core objective remains the same: given the current state and an action or instruction, predict the next state. It sees frame t and action a, and tries to produce state frame t+1.

But predicting raw pixel worlds for everything can be expensive and often wasteful. Most of what’s in a video frame doesn’t change from one moment to the next; the walls stay where they are, the sky remains the sky. And most of the details within a frame are redundant; the color of the sky, the texture of a wall. They could be described in a more compact form.

So modern World Models involve a latent space: a compressed, learned representation where only the most essential information is retained.

The visual encoder compresses each frame down to a compact vector (a mathematical fingerprint of the scene) and the model learns to predict the next fingerprint — not every pixel in the 4K frame — in response to actions. This is where the computational efficiency comes from.

To accurately model the evolution of the world, World Models must also learn to represent the full set of possible outcomes. This uncertainty in outcomes is usually referred to as the stochasticity of the environment.

World Models have to learn to navigate what they don’t know yet (epistemic uncertainty: for example, a model that has never seen a traffic light will not know that red follows after yellow) and the inherently unknowable (aleatoric uncertainty: the randomness, like rolling dice5).

Even when the model has learned all that’s possible to know about the behavior of the environment (it has reduced its “epistemic” uncertainty to a minimum), there will almost always be some inherent uncertainty (“aleatoric” uncertainty) in what happens next. This is in contrast to pure entertainment video models, which only need to be able to predict a common evolution of the world state to perform well.

If you use a straightforward prediction approach (for example, a model naively trained with Mean Squared Error, or MSE) to predict a car turning a corner, the model can become ‘blurry’ because it averages every possible outcome. The car could turn and stay in the left lane, or it could merge into the right lane. The trajectory that actually minimizes the error is the implausible one where the car stays in the middle of the two lanes. That’s the blurriness, and different models handle it differently.

Diffusion models avoid this problem by gradually diffusing towards the outcome, enabling the model to commit to a specific mode of the outcome distribution, sampling a sharp, plausible future rather than averaging all possibilities.

Autoregressive models with multiple tokens per outcome also handle multimodality; by sampling one token after the other, they ensure that future token predictions are consistent with previous ones.

JEPA-style architectures, by contrast, address blurriness by simply sidestepping it. JEPA largely avoids having to model that distribution explicitly by never decoding back to pixel space at all. It operates in a space where averaging is less catastrophic, because we don’t expect these models to predict frames, but rather to develop representations that are useful for downstream tasks.

What comes out of this process depends on what you need. If you’re building a visual world simulator — something you can watch or explore — you decode the latent predictions back into pixels through a visual decoder, producing imagined video of plausible futures. This is what makes the demos from Google DeepMind and World Labs look realistic and impressive.

There are a number of approaches used to train World Models. We will cover them and how they evolved and built on each other through the lens of the brief eight-year modern history of the field shortly.

For now, keep this in mind: observation data in, paired with the actions that caused what’s happening in those observations, train World Models to predict the next state, Agents train to predict the next action in those Worlds.

Why Actions are the Ultimate Form of Compression

Here is a key insight behind World Models: actions are the ultimate form of compression.

Consider what happens when you decide to step left to avoid a puddle. Your brain processes the visual scene (the sidewalk, the puddle, the people around you, the curb, the approaching bus), predicts the immediate future (the puddle won’t move, the bus will pass, the person behind you will keep walking), evaluates options (step left, step right, jump, accept wet shoes), and selects one.

An outside observer can’t see inside your head, can’t know exactly what you were thinking, can’t know what you’re processing subconsciously. They don’t know if you’re tired or if you’re in a rush. They don’t know your moral code, how you, specifically, would answer the Trolley Problem. They don’t need to. They see the output of all of that near-instantaneous calculation: step left.

That, to me, is magic.

Of course, not everyone makes the right decisions. Play the video forward and you are able to learn the consequences, too. Step left, into an even bigger puddle. Step left, and get clipped by a car. Step left, and knock a baby out of its stroller. Over billions and billions of observations and instructions and actions, we learn not just how humans decide to respond based on inputs, but the consequences of those decisions. The collective World Model learns to act smarter than any individual.

Zoom back into the individual. If you could perfectly reconstruct someone’s stream of observations and actions, you would have a nearly complete record of their interaction with reality. You would know what they saw and what they did about it. The World Model learns exactly this mapping. It compresses space and time into a compact representation, and then uses actions to unroll what happens next. That’s what makes World Models so computationally efficient.

It’s also the same reason why World Models can handle stochasticity that traditional simulation cannot. To understand why, let’s revisit our Man U match with our new understanding of how World Models work.

In a traditional simulation engine, every possible behavior must be coded. If you want a thousand soccer fans to react realistically to a goal, you need to write rules for each type of reaction. The computational cost scales with the number of Agents and the complexity of their interactions.

In a World Model, the cost is fixed to one neural network pass. The stochastic, messy, human reality is already baked into the learned weights and absorbed from the millions of hours of video the model was trained on. The model doesn’t calculate what a crowd should do. It has seen what crowds actually do and it uses this information to make probable predictions.

This is what I mean when I call World Models compute for the uncomputable. Traditional computing is deterministic: known inputs, known rules, known outputs. The real world is not deterministic, so World Models don’t even try to code these things in. They watch, learn, and do, at a fixed computational cost, regardless of how complex the scenario gets.

World Models and Policies

There is one more distinction to make before we go further, one that gets muddled in typical conversations about World Models.

A World Model is a simulation of the environment; it takes in actions and produces predicted observations; it shows you what will happen if you do something.

A Policy is the brains of the Agent that acts within that environment. It takes in observations (and often instructions) and produces actions; it decides what to do.

The World Model is the dream. The Policy is the dreamer. The dreamer acts, and the dream responds. The dream responds, and the dreamer acts.

In practice, the relationship between the two turns out to be even more intimate and intertwined than that distinction suggests. Recent research has investigated training policies on top of World Model foundations or building them together from the get go. Start with the weights of a World Model — a system that has learned how to predict what happens next — and then, instead of training the model to predict future frames, or states, you train it to predict future actions.

A system that learns to predict the world can also learn much faster how to act in it. Understanding and doing aren’t two separate skills bolted together. They are the same skill, seen from different angles. At least this is what our research, and that of other labs, is starting to suggest.

That means that if you build a good enough World Model, you can also more effectively train a policy to act in the worlds it generates.

This is one of many important things the field has learned in a very short amount of time. Turns out intuition and imagination are two sides of the same coin.

A (Very Brief) History of World Models

On one hand, it should be very easy to summarize the modern history of World Models. It has only been eight years since Ha and Schmidhuber published World Models.

On the other hand, an awful lot has happened in just eight years. In that time, the field has gone through four waves: major periods when the field shifted its focus to prioritizing new questions. We highlight some of the most important papers here, and not boring world subscribers can find a full downloadable list of key papers at the end of the essay.

Wave 0, in 1990-1991, was the pre-deep learning era. Researchers first articulated the idea that Agents could learn internal models of the world and use them for prediction and planning. They asked, and answered, the question: what would a World Model do?

This is Richard Sutton and Dyna. This is Jürgen Schmidhuber and Making the World Differentiable. Before we had the compute, the data, or the architecture, we had the dream, waiting in dreamspace for reality to catch up.

Wave 1, in 2018-2019, asked: “Can this even work?”

Based on Ha and Schmidhuber’s work, the first paradigm involved using Video Auto-Encoders (VAE) to compress frames, model dynamics with Recurrent Neural Networks (RNN), and train policies inside the resulting dreams. So: compress what you see, predict what comes next, and train Agents to act inside that simulation.

At the time, the question was whether learning in imagination — dreams — was feasible. Researchers attempted to answer it using small models and simple environments to generate proof-of-concept results. Quite literally, the next big thing started out looking like a toy. Model Based Reinforcement Learning for Atari introduced the Atari 100k benchmark: whether the SimPLe algorithm could learn Atari games with only 100,000 real environment steps, or about two hours of gameplay.

The answer was yes. SimPLe learned how to play 26 Atari games and beat a competitor model on sample efficiency, or how many steps it took to reach a given score.

But could it play as well as humans?

That was the question that drove Wave 2 (2020-2022): “Can the World Model match human performance?”

DreamerV2, developed by Danijar Hafner at Google DeepMind, reached an answer quickly. They used a Recurrent State-Space Model (RSSM) with discrete latent representations — a system that maintains a compressed, running memory of the world and updates it with each observation. DreamerV2 became the first World Model Agent to achieve human-level performance across the 55-game Atari benchmark6. It was trained entirely in imagination, on a single GPU.

That same year, another DeepMind team published Mastering Atari, Go, chess and shogi by planning with a learned model in Nature. The paper described its MuZero model, which also beat Atari games (and others like Go), but did so by taking almost the exact opposite philosophical approach.

Whereas DreamerV2 generated observable dream environments and trained inside of them, MuZero never generated anything observable at all, planning entirely in abstract latent representations it invented for itself, and it did well.

It did so well, in fact that it leapfrogged the Go-specific models. In 2016, DeepMind’s AlphaGo beat human Go Champion Lee Sedol 4-1. It had been trained on a large database of human expert games plus self-play, with the rules of the game hard-coded in. The next year, AlphaGoZero beat AlphaGo 100-0 after being trained entirely from self-play with no human game data at all, just the rules. That same paper season, AlphaZero generalized AlphaGoZero’s approach to other games, like chess and shogi, both of which it came to dominate within hours. Then in 2019 (pre-print), MuZero learned everything, including the rules, the game dynamics, and the value function, from scratch, purely from observation and outcome. It matched AlphaZero on Go, chess, and shogi (where AlphaZero knew the rules) while also generalizing to 57 Atari games (where “rules” aren’t even a well-defined concept).

With each new model, something that humans had previously hard-coded — the rules, the strategy, the value of a position — was removed. The model learned each from scratch instead. MuZero was the terminus of that progression, entirely learned.

And MuZero did this without imagining future board states at all. It imagined hidden states, or abstract vectors it invented for itself during training that have no guaranteed correspondence to anything human-observable or interpretable. A human looking at MuZero’s internal representation of “three moves from now” would have absolutely no idea what it was thinking. And yet… it outperformed all previous models.

With MuZero’s success, the field now had two opposing schools of thought: generative World Models that produce observable futures, and latent World Models that predict in abstract space, even if they weren’t called “latent” yet.

From then on, progress in World Models has happened in both directions, generative and latent.

On the latent side, in 2022, Yann LeCun published a sweeping position paper from his dual positions at Meta and NYU Courant proposing a fundamentally different philosophy from generative models, one that looked more like MuZero: A Path Towards Autonomous Machine Intelligence. His new World Models company, AMI, is named after this paper.

LeCun’s Joint Embedding Predictive Architecture (JEPA) argued against generating pixels entirely. Similar to MuZero, instead of predicting what the world will look like, JEPA predicts what it will mean. It forecasts abstract representations of future states, deliberately discarding unpredictable visual details.

That same year, on the generative side, IRIS (2022), developed by Vincent Micheli and Eloi Alonso, two of General Intuition’s future co-founders, reframed World Modeling as language modeling over a learned vocabulary of image tokens. Instead of recurrent state-space models, IRIS used a GPT-style autoregressive transformer over discrete visual tokens. Basically, IRIS borrowed the machinery of language models and applied it to World Modeling.

In doing so, IRIS filled a number of previous gaps. The IRIS World Model was, in effect, a language model, but its vocabulary was images and actions instead of words. This brought the scaling properties of LLMs directly into World Modeling: efficient attention, scaling laws, and all the engineering infrastructure that had been built for large language models could now be applied to learning about the physical world.

Where Dreamer was missing the ability to model the joint law of the next latent state (for example, to handle multimodality), IRIS represented the next latent state as a series of discrete tokens to predict autoregressively, which meant that it was now able to predict multiple outcomes. And while Dreamer beat humans by using much more data than they do, IRIS was the first learning-in-imagination approach to beat humans with the same amount of available gameplay data (two hours).

JEPA aside, practically all of the work up to this point in World Models happened within games, and it’s worth floating between Wave 2 and Wave 3 for a second to appreciate the special relationship between AI and games.

Games have always played an important role in the development of AI. Claude Shannon’s 1950 paper Programming a Computer for Playing Chess is one of the founding documents of AI. In 1959, Arthur Samuel’s checkers program introduced the concept of machine learning itself. The first time the world woke up to the idea that intelligent machines could beat humans at anything was when IBM’s Deep Blue beat Garry Kasparov in chess.

Before DeepMind was an AI lab, Demis Hassabis was a game designer. At 17, he designed the commercially successful Theme Park. DeepMind’s founding breakthrough is detailed in the DQN paper, published in Nature in 2015, in which it was demonstrated that Atari games could be played from raw pixels using deep reinforcement learning. Then came AlphaGo in 2016, which beat the world champion at Go, a game that once was believed to require the kind of intuition that was uniquely human, with more possible board positions than atoms in the universe.

The path from AlphaGo to AlphaFold ran through exactly the insight that World Models formalize. As Hassabis put it:

Wouldn’t it be incredible if we could mimic the intuition of these gamers, who are, by the way, only amateur biologists?

General Intuition is named after this quote from Demis, which points towards a future where our models power research far beyond the dynamics of what pixels can describe today, beyond games themselves, and into our bodies.

And then DeepMind taught machines how to fold proteins. AlphaFold won Hassabis and his DeepMind teammate John Jumper the 2024 Nobel Prize in Chemistry.

Games are fun, of course. But the reason games keep showing up is that games are the only domain where you get massive amounts of labeled spatial-temporal data with clear action-outcome pairs, consistent physics, unambiguous reward signals, and a controlled environment where you can run millions of experiments. The real world has none of these properties.

Early World Models, like a human child, spent most of their time watching and playing games. The Atari 100k benchmark became the standard arena for World Model research, DreamerV3 played Minecraft, and many current World Model companies retain a connection to games, with many World Models being “playable.”

Games are the lab bench of embodied AI. But they are only a small fraction of the ambition.

For World Models to be truly useful, they need to interact with the world.

That’s Wave 3 (2023-2024). It asked: “Can World Models be truly interactive?”

We got the first answer from driving. GAIA-1 (2023), developed at Wayve, scaled the sequence-modeling approach pioneered by IRIS to 9 billion parameters and trained on real-world driving video. It could generate driving scenarios in response to actions (steer the car), text prompts (“rainy day, highway”), or both. Anthony Hu, who led this research, now leads World Modeling at General Intuition.

GAIA-1 confirmed that the scaling laws everyone had observed in LLMs also held for visual World Models. More data and more parameters yield predictably better performance for World Models, too. This was not a given. It meant that the path forward was clear even if it was expensive: scale up and the models get better.

The following year, DIAMOND (2024), developed by future General Intuition co-founders Eloi Alonso, Adam Jelley, and Vincent Micheli, opened a new architectural frontier. Rather than compressing observations into discrete tokens and predicting them autoregressively, as researchers had been doing since IRIS, DIAMOND used diffusion models to predict future frames directly.

The visual fidelity was meaningfully richer, and that richness translated directly into better Agent performance. The subtle visual details that discrete tokens discarded, the little clues that tell you a surface is slippery, a door is ajar, a person is about to change direction, turned out to matter for decision-making, which is unsurprising when you think about it.

As a brief aside, it’s worth noting that many of the open source advancements that have been made in World Modeling were built on top of the DIAMOND architecture. Multiverse, the first AI-generated multiplayer game, is DIAMOND-based, as is Alakazam, the “1st ‘World Model game engine’.” DIAMOND is essentially the Deepseek or Llama of Generative World Models.

DIAMOND itself set a new best on Atari 100k and demonstrated something that captured the public imagination: trained on Counter-Strike gameplay, it produced a fully interactive, playable neural game engine from roughly 87 hours of footage on a single GPU.

It showed that it was possible to run an interactive 3D World Model in real time too.

Counter-Strike environment generated by DIAMOND with just 87 hours of footage

DIAMOND got very good at playing Atari. The Agent plays a real game and gathers real data there, with which it trains the World Model. Then it tests itself inside the World Model’s synthetic environment, gets better in there, and goes back out for more real interaction, to test itself in the wild. This loop between ground truth and synthetic, back and forth, is how World Models improve, almost like working problems out in a lucid dream then testing them in reality upon waking. This is the Dyna paradigm mentioned earlier.

Would that loop work in real-world conditions?

It turns out that the answer is yes, too. And that it would work beautifully.

GAIA-2 (March 2025) pushed the diffusion approach to its most ambitious application yet: multi-camera autonomous driving simulation. Using latent diffusion with flow matching and space-time factorized transformers, the model could generate high-resolution surround-view driving video conditioned on ego-vehicle dynamics, other Agents’ trajectories, weather, time of day, road structure. In short, it could reproduce the full complexity of real driving. It could simulate scenarios that were too dangerous or too rare to collect from real roads: sudden cut-ins, emergency braking, pedestrians stepping off curbs.

GAIAs 1 and 2, and DIAMOND, like IRIS, were the products of researchers we now get to work with at General Intuition. Diffusion or flow-matching models like GAIA-2 were the starting point of our team’s current research efforts.

But they are not the only approach.

Google DeepMind is one of the central players in this space. Their World Model, Genie (2024), is an 11-billion-parameter model trained on unlabeled internet video of 2D platformer games. It learned an action space entirely from scratch; no one ever told the model what the controls were. Give it any image and it can generate a playable world from it.

OpenAI’s Sora (2024, with Sora 2 following in 2025) and Google’s Veo 3 (2025) pushed video generation to extraordinary visual quality and framed these systems explicitly as “world simulators.”

The field’s vocabulary can get muddied. Let’s make it clear.

Video generation models produce beautiful visual sequences, but they aren’t quite World Models in the sense that we’ve been describing them. In these videos, you can’t take an action and watch the environment respond live to your intervention. They predict what a scene will look like over time; they don’t model what happens because of what you do.

Think of the difference between watching a movie of someone driving and actually steering a car. The visual output might look similar, but the underlying computation is fundamentally different. Interactivity, the ability to take actions and observe their consequences, is what separates a World Model from a very impressive video.

And interactivity is what it takes to impact the real world.

This is the central question of Wave 4, the wave we’re in right now: “Can models act in the real world?”

As in: Can Agents trained in World Models work outside of research settings, in real vehicles, real robots, real deployments? We are now getting awfully close to sci-fi’s predictions.

This is where the current frontier is being pushed. Right now. As you read this.

Comma.ai took the most direct path in driving from World Model to product: Learning to Drive from a World Model. They trained a driving policy entirely inside a learned World Model — inside the dream — and deployed it in openpilot, their open-source driver assistance system running on production vehicles driven by real people. The World-Model-trained policy outperformed both traditional imitation learning and policies trained in conventional simulators. This is arguably the first consumer product powered by a World-Model-trained Agent.

In robotics, Meta’s V-JEPA 2 animated LeCun’s latent prediction philosophy. The model is the clearest large-scale proof point so far. It’s a 1.2B-parameter model pre-trained on over a million hours of video via self-supervised masked prediction: no labels, no text. In the second stage, it fine-tunes on just 62 hours of robot data from the Droid dataset. It turns out that is enough to produce an action-conditioned World Model that supports zero-shot planning. V-JEPA 2 was deployed zero-shot on real Franka robot arms in new environments to perform pick-and-place tasks. It planned all of this entirely in latent space, without pixel generation, task-specific training, or hand-crafted rewards. And it was fast; where pixel-space approaches took minutes to plan a single action, V-JEPA 2 did it in seconds.

Google DeepMind’s SIMA 2 took an entirely different approach. Rather than build a dedicated World Model, it fine-tuned Gemini, its large foundation model, to act directly as an Agent in 3D game environments. SIMA 2 can reason about high-level goals, follow complex multi-step instructions, converse with users, and generalize to unseen environments.

It represents an alternative paradigm: instead of building a specialized World Model, leverage the implicit world knowledge already embedded in a model trained on the breadth of human knowledge.

This is one of the field’s open questions. Will this path, using a large foundation model or a video model as the basis for an Agent, rather than training an Agent from scratch in a World Model, win out?

In fact, there are many open questions. And nearly as many World Model startups trying to answer them.

The State of the World (Models)

That brings us to the present moment.

What has become clear is that talented researchers and investors alike are excited by World Models’ potential, as evidenced by the massive funding rounds to support companies led by legends in the field.

In February 2026, World Labs, the company founded by legendary researcher Fei-Fei Li, announced that it had raised a fresh $1 billion from investors at a $5.4 billion post-money valuation.

Not to be outdone, Yann LeCun, who launched AMI Labs in late 2025, announced last week that it had raised $1.03 billion at a $3.5 billion valuation.

In October 2025, our company, General Intuition, announced $133.7 million in a very large seed round. Last summer, Decart raised $100 million at a $3.1 billion valuation. In November, Physical Intelligence raised $600 million at a $5.6 billion valuation for its robot foundation models. And just this past February, Wayve, the UK-based self-driving startup whose researchers built GAIA-1 and GAIA-2, raised $1.2 billion at an $8.6 billion valuation.

Google DeepMind, which doesn’t need to fundraise because it’s fueled by history’s greatest business machine, is pouring resources into SIMA, Genie, and Veo, and using it to power initiatives like Waymo. Demis has publicly stated that he believes World Models will become an important part of Gemini’s planning capabilities. GDM is also merging many of these capabilities into a “Video Thinking” team, with the reasoning described best by Shane Gu and Jack Parker Holder from GDM.

What is less clear, but even more interesting, is that we are at the point in this technology’s development where we know that something big is happening, but it’s still unclear exactly which approach, or combination of approaches, will win out. We are seeing breakthroughs almost every day at General Intuition, and we hear rumors of leaps happening in other labs too.

Below is a framework in which to fit any news you see coming out on World Models. We won’t cover everything, and we apologize in advance if we miss your embodied AI of choice. A fun exercise for the reader will be to fit what we’ve missed into what we’ve laid out.

World models have three main types of approaches: Current Foundation Models, World Models, and Embodied Agents.

The thing to keep in mind here is that, despite different World Model approaches, we all share the same end goal. The end goal is to produce Agents that generalize and do things in various environments, including the real world. Some of the Agent approaches get there using LLMs as their stepping stone, others start with video models. Other agent approaches use World Models as their training environments. And some Agents learn directly from experience.

With us? Goed, daar gaan we dan!

Current Foundation Models

Current Foundation Models are the ones that learned to make sense of the world’s data without being able to simulate the stochastic world environment itself. They are models that process inputs — text, images, video — and learn to predict, generate, or reconstruct. But they don’t yet give an Agent a place to act. They are not action-conditioned. They don’t respond or interact. They are potential substrates on which World Models can be built, or even, in some cases, on which Agents are pre-trained.

Three categories of stepping stone models we’ll focus on here are Large Language Models, Video Models, and 3D Reconstruction Models.

Large Language Models

LLMs learned from staggering quantities of text that the world has structure. They know that a glass falls when pushed, that fire is hot, that if you leave the house without an umbrella in a rainstorm you will get wet. They encoded an enormous amount of causal and physical knowledge. But none of this was from experience. Like digital Castelians, they read about the world rather than perceiving it. This makes them extraordinarily useful as a backbone for reasoning and planning, which is why you’ll find LLMs embedded in many agent architectures we’ll discuss later. But a language model alone cannot simulate what happens when a robot arm reaches for a cup.

In our context, LLMs are particularly relevant when we discuss VLAs, or Video Language Action models, which take advantage of the enormous amount of research, capital, tooling, and infrastructure that has gone into developing LLMs in order to bootstrap robots that can do things in the physical world.

Video Models

Sora. Veo 3. Kling. Seedance 2.0. Runway. Pika. Moonvalley. Haiper. Luma AI.

No one confuses an LLM for a World Model, but plenty of people conflate Video Models and World Models.

These models are trained on the enormous amount of video data on the internet, and produce extraordinary videos themselves. Sora can generate a convincing shot of a woman walking through a neon-lit Tokyo street. Veo 3 can render photorealistic scenes with synchronized dialogue.

But you can’t interact with them. You can’t take an action inside of them and watch the environment respond instantly. They predict what a scene will look like over time but they don’t try to model what happens because of what you do.

Of course, the lines get blurry.

Odyssey, founded by self-driving heavyweights Oliver Cameron (ex-Cruise) and Jeff Hawke (ex-Wayve), is building “a world simulator that dreams in video.” Currently, they don’t let you take an action and watch the environment respond, but they do let you prompt the video mid-stream to steer it in real-time. Where do you draw the line?

Wherever the line is, these video models are getting good, and really funny.

Really, really funny.

Video models aren’t quite World Models in the sense we define them; they are a stepping stone. Runway began as a video generation company – its Gen 4.5 is among the best on the market – but has concluded that physics-aware video generation is a path toward something bigger. This thinking led to GWM-1, their explicitly labeled “General World Model, built to simulate reality in real time,” which is interactive, controllable, and general-purpose. The real value, financially and societally, won’t come from video for its own sake, but from models that use video as a training environment on the way to controlling embodied systems.

3D Reconstruction and Generation Models

Take it a step further. What if you could navigate through the scenes depicted in video generation models? That feels like a world, right?

World Labs, led by the legendary Fei-Fei Li, the “Godmother of AI” who created ImageNet, is the most interesting example in this category. While the company is the one most people would associate with “World Models,” World Labs is not currently building what I would define as World Models.

Instead, in its early days, World Labs has focused on immersive virtual worlds, but not action-conditioned ones. Its first product Marble generates and edits persistent 3D environments from text, images, video, or 3D layouts. They call it a “Multimodal World Model.”

Marble is thus far not interactive, other than being able to move through the generated environments. They say this themselves. On the Marble product page, World Labs frames interactivity as a future opportunity:

Future World Models will let humans and agents alike interact with generated worlds in new ways, unlocking even more use cases in simulation, robotics, and beyond.

It is worth noting that World Labs has recently started exploring World Models that do generate frames directly, instead of the underlying splats of the entire world.

World Models

A World Model, as we define it, is an environment that an Agent can act in, and that responds in real time. It is a simulation, a dream, one learned from observation and actions data rather than hand-coded. The Agent takes an action, the world changes, and the Agent observes what happened. Repeat, millions of times, across an enormous variety of situations, and the hope is that you get an Agent that generalizes, that can do things that were not in the original training data.

This is the key distinction that everything else hinges on: a World Model is action-conditioned. It predicts what the world will look like next given whatever the Agent did.

The intuition is simple. A robot trained only on real-world data has seen a finite set of kitchens, a finite set of cups, a finite set of ways a cup can fall. Put it in a kitchen it hasn’t seen, with a cup it hasn’t encountered, and it struggles. A robot trained inside a World Model, on the other hand, has, in principle, encountered infinite kitchens because the World Model can generate them. Situations that would be rare, expensive, or dangerous to collect in the real world become routine in simulation. Out-of-distribution becomes in-distribution.

Within World Models, there are two main approaches: Latent World Models and Generative World Models.

I apologize for bringing you so far in the weeds here, but I want to clarify something that confuses people: both Generative World Models and Latent World Models rely on latent states, but Generative World Models rely on latent states that were designed with reconstruction objectives (autoencoders) which enable frame predictions, whereas Latent World Models directly build self-predictive representations.

Latent World Models were born in the darkness and still live there; Generative World Models were merely born in the darkness.

Latent World Models

Latent World Models are the descendants of MuZero but let loose in open-ended, no-rules environments like the real world.

This is Yann LeCun’s current world. Yann pioneered modern computer vision architectures with LeNet, where he introduced the idea behind convolutional neural nets (CNNs) in the 1990s. In the 2010s, he championed self-supervised learning, arguing that human labeling millions of examples doesn’t scale to real intelligence and that models should create their own signal from raw data. In the 2020s, he led the JEPA team. Yann is a GOAT.

The deep thread in Yann’s work is teaching models to learn useful representations of the world automatically from raw data. Latent World Models are the latest, and perhaps ultimate, strand in this thread.

The approach is philosophically the converse of Video Models or 3D Reconstruction Models, as mentioned earlier in the history section. While those approaches care about producing and understanding every pixel, latent World Models, like JEPA, says ne vous embêtez pas. The French would rather speak English to you than listen to you butcher their language. JEPA is similarly impatient; rather than let the model stumble over every pixel of an unpredictable future, it doesn’t predict pixels at all.

As LeCun puts it: “The world is unpredictable. If you try to build a generative model that predicts every detail of the future, it will fail. JEPA is not generative AI.”

Instead, JEPA learns to represent videos in abstract, compressed space and makes predictions there. It deliberately throws away unpredictable visual details. This makes JEPA potentially very efficient for planning and representation learning.

AMI Labs is LeCun’s bet that this approach is the path to real intelligence, and investors recently backed him with $1.03 billion.

There are trade-offs to the latent approach, as there are trade-offs to generative approaches.

LeCun argues that the thing that seems like the biggest trade-off, a loss of fidelity in exchange for speed, is not actually a trade-off. His position is that the detail you lose is detail you should lose, that trying to predict every pixel is not just expensive but actively counterproductive — the model wastes capacity on inherently unpredictable visual details instead of learning the abstract causal structure that actually matters for reasoning and planning. Imagine if you needed to simulate every photon when you imagine catching a ball. Your brain might explode. There’s some level of detail that is not “every single detail” that is optimal. LeCun’s argument is that with World Models, the optimal level requires fewer details than many people, including us, think.

There are other trade-offs to keep in mind, however, that LeCun hasn’t mentioned.

One is that latent models are trickier to evaluate. You can’t look at the output and see if it makes sense intuitively the way you can with generated video, and they can’t serve as training grounds for human-in-the-loop systems, because humans can’t operate inside latent space. We need to see the world to act in it.

Another related downside is that your iteration speed slows down when you can’t visualize predictions or interpret the loss. Humans are very good at noticing when something is visually off; we did not evolve to spot discrepancies in predicted latent encodings of the future ([0.13, -1.02, 0.44, 0.07, …], MSE = 0.0187). And iteration speed is what matters most in modern ML, because modern ML progress mostly comes from empirical search, not from knowing the right design ahead of time.

Latent models are also more challenging to train for similar reasons. Additionally, the lack of strong supervision in the learning objective leads to collapse issues which require a bunch of tricks to fix. Why? The JEPA objective is to predict the encoding of the future based on the encoding of the past, but you can satisfy that objective with trivial encodings (e.g. set everything to 0, there is 0 loss), so we need to make sure representations don’t collapse.

There is a spectrum in creating environments in which agents can train. On one side is what is practical today, and on the other is the platonic ideal. Latent World Models are almost the opposite side of the Practical ←→ Platonic spectrum to VLAs, which we will cover below.

They are closer to what researchers believe to be the technical platonic ideal, but they face real challenges in practice today. That said, new methods like LeJepa are closing the gap, and talent is flooding into the field.

Chris Manning, Ian Goodfellow, and Fan-Yun Sun have also joined the cohort of Latent World Models, starting latent lab Moonlake. Their entrance on the side of latent is notable. Manning helped pioneer neural natural language processing and co-created GloVe, which was the dominant word-embedding model before transformers. And Goodfellow invented GANs (generative adversarial networks), which were the first widely successful way to train neural networks to generate realistic synthetic data.

In a recent X post, the Moonlake co-founders explained their approach to building efficient World Models. It is an interesting hybrid.

The plan is to generate full game environments to attract human players and collect action-labeled data. Afterwards, they model the world in semantic/symbolic space rather than pixels. As in, they use beautiful game environments to attract real human players, because they need humans to generate action-labeled data. But once they have that data, they discard the pixels entirely and train on abstract representations instead, betting that the underlying patterns matter more than the visual detail.

Ultimately, we don’t view latent and general models as opposed to each other. Moonlake’s hybrid approach is evidence of that. They just serve different goals. Latent World Models are generally more computationally efficient since they discard some information, with advantages for representation learning and planning. Generative World Models should be more general, since in theory they capture all visual information, with advantages for interpretability and generalization. Both can be used for many different purposes, including training agents with reinforcement learning.

Now, let’s turn to Generative World Models.

Generative World Models

Generative World Models are the closest thing to simulating human-perceived reality that we’re aware of. If our world is a simulation, it’s probably a Generative World Model of some sort.

This is the paradigm we at General Intuition predominantly focus on to build the World Models in which our policies learn. It’s also the one that recently blew the world’s minds when Google DeepMind released Genie 3.

The video — and, Genie 3 itself, if you get a chance to play with it — gives you a felt sense for what makes Generative World Models different. They’re interactive. They respond.

They generate human-observable, interactive futures that you can see, act in, and learn from. You can see what the model thinks will happen next. The model takes in a state and an action and produces a plausible next state, which you can act in again. Based on the updated state and new action, it produces the next plausible next state, and so on and so forth. A human can look at the output and say, “That’s wrong, walls don’t bend like that” or “Yes, that’s exactly what happens when you turn a steering wheel at speed.”

Generative World Models predict the observations themselves in pixels, video, or 3D scenes, allowing agents and humans to interact with the simulated environment. The dream is visible and playable.

This is what improves the training loop in many cases. Both generative and latent models can learn in imagination. Yet when visual details matter, or when the downstream task is not yet known, Generative World Model learning, with all its pixel-level detail, tends to outperform.

This only works if the generated environment is rich enough to learn from. The further the generated world is from reality, the worse the lessons the agent learns are. And the less successful it is when it goes back to real games. This is what DIAMOND showed, that when there is more detail in the generated world, agents are smarter.

At General Intuition, we are building on this diffusion and flow-matching architecture. It is developed, in part, by researchers who are now our co-founders and who built IRIS, DIAMOND, and GAIA-2.

Wayve, the birthplace of GAIA-1 and GAIA-2, is the leader in Generative World Models for autonomous driving. By using a large latent diffusion World Model offboard, they aim to dream up edge cases that would take millions of driving miles to find in reality, train driving policies on them, score the driving policies on their performance in simulation, then distill that dreamed experience into a smaller onboard policy that can reason through those same scenarios in real-time. The tweet below shows Wayve zero-shotting a drive on Japanese roads in the latest installment in a series doing this around the world.

Decart is applying Generative World Models to real-time generative simulation, producing playable worlds that respond to the users’ actions. It’s the playable version of the Generative Video Models or 3D Reconstruction Models. On the Oasis landing page, it calls the model a “video model,” but follows up with this distinction: “Every step you take will reshape the environment around you in real-time.”

Interestingly, Decart currently runs on Nvidia GPUs but plans to use Etched Sohu chips. Etched chips are custom ASICs designed to run transformers and would allow Decart to improve latency and run continuous inference, both of which are much more important when generating responsive worlds in real-time than when generating a video or 3D rendering upfront.

Runway, too, is blurring the lines between video generation and world generation, as mentioned in the Video Model section. During its Research Demo Day 2025, Runway co-founder and CTO Anastasis Germanidis explained the company’s evolution that started from “generative AI models [as] viable tools for creative expression”. They then evolved towards World Models (while still making incredible progress in video models.)

“To build a World Model,” Germanidis explained, “We first needed to build a really great video model. We believe that’s the right path to building World Models, that teaching models to predict pixels directly is the best way to achieve general purpose simulation.”

Google DeepMind took a similar approach; Genie 3 was built on top of Veo.

These World Models are extremely important. But remember that they are only half of the equation. From the very start, whether it was Schmidhuber in 1990, Sutton with Dyna in 1991, the plan was to use World Models to train Agents to act inside of the world, and then to transfer those learnings out into the real world.

Embodied Agents

We want to share a few of the main Embodied Agent examples out there today, and their respective approaches: Physical Intelligence and other robotics companies’ VLAs (Vision-Language-Action Models), DreamerV4’s a Latent World Model Agent, Google Deepmind’s Sima2 General Embodied Agent, and General Intuition’s General Agent approach.

Physical Intelligence - Vision-Language-Action Models (VLAs)

Modern multimodal LLMs come with a model called a VLM, or Vision-Language Model, a model that can see and read. Feed it an image and a question, like “What objects are on the table?” or “Is this door open or closed?”, and it produces a coherent, grounded answer.

GPT-5, Gemini, and Claude are all VLMs in this sense; they can see and reason. When you send it a picture of a mountain and ask it to geolocate it, it’s using its VLM. VLMs are also the perceptual and reasoning backbone of most modern Agent systems designed to operate in physical or interactive environments, like PaLM-E or SpatialVLM.

VLMs are not exactly Agents, but they are core components of most of them. We bring them up because a VLA is a VLM that has learned to act, and it is the pragmatist’s answer to the Agent problem.

In 2023, Google DeepMind published a paper called RT-2: Vision-Language-Action Models: Transfer Web Knowledge to Robotic Control to propose a solution.

Take a VLM that understands a scene and what to do in it, and then bolt on an action head that translates human language instructions into instructions that the robot understands, like to change a position or rotate.

Since then, VLAs have become the dominant paradigm in robotics, and they’ve worked surprisingly well.

Every other paradigm we’re discussing says something like: “Images, videos, spaces, and actions are fundamentally different than words. We need to train and architect the models that generate them differently than the models that generate words.”

Vision-Language-Action Models (VLAs) say: “That may be true! Those approaches might be platonically better. But that doesn’t matter in practice, because the vision-language model infrastructure and data are so far ahead.”

In his not boring primer on robots, Standard Bots’ Evan Beard wrote a thorough explanation of VLAs for robotics that included what he called a “spicy take”:

We’re not working with language-model infrastructure because it’s the perfect architecture for robotics. It’s because we, as a species, have poured trillions of dollars and countless engineering hours into building LLM infrastructure. It’s incredibly tempting to reuse that machine.

So, despite its imperfections, taking an LLM and sticking on an action head to predict robot motions (all together known as a VLA) is the best way for us to train the base models that learn many skills from demonstrations across many different customers and tasks.

It’s pretty ingenious. Of course, there are challenges to this approach, which Evan highlighted:

• Robotics success so far has leaned heavily on diffusion-style control

• LLMs are autoregressive and token-based, with less room for error

• Physical actions don’t map cleanly to tokens

Additionally, compared to World Models, VLAs require collecting a large amount of real-world robotics data; they don’t seem to generalize out-of-distribution particularly well.

That said, Physical Intelligence, known as π or Pi, has gotten incredibly far with its VLA bet.

Pi’s first generalist policy, π₀: Our First Generalist Policy, inherits semantic knowledge and visual understanding from internet-scale pretraining and trains on data from seven different robotic platforms across 68 unique tasks, including folding laundry, bussing dishes, routing cables, assembling boxes, and packing groceries, all of which require dexterity in the real world on real hardware. Their follow-up, π₀.5: a VLA with Open-World Generalization, performs better in new environments like cleaning up a kitchen or bedroom in a home the model has never seen before.

OK, but can it actually learn and get better over time as it works and makes mistakes in the real world?

November 2025’s π*0.6: a VLA that Learns from Experience suggests that it’s possible, with demonstrations in tasks like making espresso, folding boxes, and folding laundry.

But those are simple, repetitive tasks. Most of what the robot sees is in distribution. Can it actually do more complex, multi-step tasks that take a long time to complete?

Earlier this month, Pi released VLAs with Long and Short-Term Memory and showed that robots using MEM (Multi-scale Embodied Memory) can clean up an entire kitchen, set up the ingredients for a recipe, and grill a grilled cheese sandwich. They can also learn from their mistakes.

A robot tries to pick up a chopstick or open a refrigerator door. Without memory, it fails the same way repeatedly. Each attempt is a clean slate with no knowledge of what just went wrong. With memory, it tries a different approach after the first failure. And it succeeds.

MEM doesn’t change the underlying architecture, which is still sub-optimal for embodied systems. Most of the parameters still live in the language backbone. The action head is still downstream of reasoning. But Physical Intelligence’s existence raises a fascinating question. Do these architectural limitations actually matter in practice?

If Latent World Models are on one side of the Platonic ←→ Pragmatic Spectrum, VLAs are on the other.

To date, Pi has been able to engineer their way around architectural limitations to make increasingly capable robots. Their progress is not slowing down. It seems to be accelerating.

Theirs is a bet with historical precedent. The ideal technology — the solution that is technically superior — does not always win. This is the key takeaway from W. Brian Arthur’s 1989 paper, Competing Technologies, Increasing Returns, and Lock-In by Historical Events. Markets often converge on the technology that gets adopted first, because adoption creates increasing returns: better early product means more users and more capital, which means better data, more internal talent, and more developers, which means better products which means more users and capital, and so on.

This is also the point of Sara Hooker’s 2020 paper, The Hardware Lottery: “This essay introduces the term hardware lottery to describe when a research idea wins because it is suited to the available software and hardware and not because the idea is superior to alternative research directions.”

From the outside, it seems that Pi’s strategy is to ride the transformer architecture’s increasing returns and trying to generate its own, to create path dependence with VLAs before World Model-specific architectures gain traction, in an attempt to win its own hardware lottery.

They’re not the only company making this bet. Skild, the closest direct competitor, is building on VLAs. A number of robotics companies incorporate VLAs and VLMs in one way or another. And now, it looks as if the approach is spreading to the whole factory.

Recently, the WSJ reported that former OpenAI Chief Research Officer Bob McGrew is raising $70 million at a $700 million valuation for his new startup, Arda, in a round led by Founders Fund and Accel, with participation from Khosla and XYZ. Details are light, but the WSJ’s description sounds like it’s going to at least involve VLMs and VLAs in some way: “Arda is developing an AI and software platform, including a video model that can analyze footage from factory floors and use it to train robots to run factories autonomously.”

The more well-funded and talented companies that move in this direction, the more deeply grooved the path becomes.

Personally, I don’t think VLAs and World Models are really competing. They’re trying to reach acting in the physical world from different directions. VLAs are language first, while World Models are video actions-first. My guess is that they’ll converge and both be part of the solution.

Dreamer V4 - Latent World Model Agents

Latent World Model Agents are Agents trained inside of Latent World Models. The latent approach has a natural elegance for Agent training specifically.

Because a Latent World Model operates in compressed abstract space, the Agent’s planning and policy learning can happen very efficiently, with no pixel generation required. The agent basically practices by thinking, in the same way a chess grandmaster runs through variations in their head without moving the pieces, or the way a lucid dreamer trains inside the dream.

The canonical example is Dreamer, from Danijar Hafner, now at Google DeepMind. Dreamer’s insight is elegant: if you have a good enough Latent World Model, you don’t need to touch the real environment during training at all. The Agent imagines sequences of actions and their consequences entirely in latent space, receives a reward signal, and updates its policy, all without a single real-world interaction. When it finally goes into the real environment, it already knows what to do.

Dreamer has achieved remarkable results across a wide range of tasks, from games to continuous control to robotics, all from this purely imagined training. It is the research proof of concept that World Model training works, that an Agent can learn to act in the real world by dreaming. It seems as if Hafner is taking his research proof commercial. Earlier this month, The Information reported that he and Wilson Yan are raising $100 million to build a World Model company in this paradigm called Embo, which suggests they’re going after embodied systems.

The challenge, as with Latent World Models generally, is that the Agent’s learned behavior is only as good as the latent representation. If the World Model’s abstract encoding misses something causally important, like the exact texture of the floor that determines whether the robot slips or the precise angle of an object that determines whether it can be grasped, the Agent won’t know to care about it, because the model didn’t encode it. Garbage in, garbage out, but the garbage is invisible.

Moonlake’s hybrid approach, which we discussed earlier, is an attempt to thread this needle: attract humans with beautiful generative environments to collect action-labeled data, then discard the pixels and train the Agent in abstract space. Use the generative world to get the data. Use the latent world to do the learning. It’s an interesting bet that the two approaches are more complementary than competing, and it may prove correct.

Notably, we haven’t yet seen the JEPA Agents. JEPA is a World Model architecture, not an Agent architecture, but we expect AMI Labs will close this loop. AMI is still building its World Model, and the Agents that train inside it haven’t yet been publicly demonstrated, but we’re watching closely.

General Embodied Agents

SIMA2 - Generalist Embodied Agents from VLM Backbone

In November 2025, Google DeepMind released SIMA 2: An Agent That Plays, Reasons, and Learns with You in Virtual Worlds.

SIMA 2 combines a Gemini backbone with a World Model trained on 3D game environments, giving the Agent an understanding of language that allows it to receive and reason about goals, as well as the spatial-temporal understanding to execute on them. In this architecture, Gemini fills the role that we mentioned VLMs play in our system.

What makes this a different paradigm from VLAs is the direction of citizenship. In a VLA, language is first class, images are second class. Beyond the ordering of modalities, there is also the training data, mostly static images interleaved with text. In an Agent equipped with a World Model, video is first class, actions are introduced from the beginning, and the training data is directly aligned with the downstream behaviors we’re looking for. The Agent’s fundamental competence is spatial-temporal. If you tell it what it needs to do, it knows how to move through the world to do it.

SIMA 2 can play games on its own. It can learn, reason, and improve. The more it plays, the better it becomes, not just in the games it’s played, but in any game it plays. It is even able to play in whichever generated world it gets thrown into, even when it’s never seen it before. This, Google DeepMind believes, is a “step towards creating AI that can help with any task anywhere, including one day in the real world.”

Google DeepMind has put out an extensive amount of research. They have pushed World Models and embodied AI forward on multiple fronts. They coined the term “VLA.” They released Genie 3. They developed SIMA 2. The way they trained AlphaGo, letting the Agent play against itself over and over and over again, informed how World Models are trained to this day.

General Intuition - Generalist Agents From Actions & World Models

Similarly to Google DeepMind, we also believe that Generalist Agents will play a major role in how embodied systems operate to do useful things.

First, create the dream. Then, let Agents run around inside of it. Let them play and mess up and learn and win. Then, transfer those learnings into other dreams, and even into the real world.

Recall the Matrix. When Neo needed to learn Kung Fu, he plugged into a virtual dojo where he trained against Morpheus in a training environment superior to the “real world.” After that? “I know Kung Fu.” World Models are the virtual dojo. Neo is the Agent.

This is the question that Ha and Schmidhuber asked eight years ago: can Agents learn inside of their own dreams?

In a remarkably short amount of time, the space has come to an answer: yes.

Yes… If You Have Action-Labeled Data (Or Can Get It)

Today, I want to share a little bit more about our approach and the results we’re starting to see.

Every approach I’ve written about so far runs into the same wall, eventually: It needs better data. Video is abundant, but it lacks depth. It has no action labels. And without knowing what actions caused what we’re seeing, video data is like shadows, the shadows on Plato’s cave wall.

And Yann may be right that you can infer action, but anyone using inferred action has separate scaling laws to consider: inferring the actions themselves. Inferring actions takes compute, time, and attention away from doing the things you can do once you understand actions, and while inferred actions might look good on benchmarks, they struggle deeply on edge cases. Even well-inferred actions are approximations of what someone actually did: some things just aren’t visible in video, like moving the rudder on a plane landing from the cockpit.

Hint: if you don’t do it, you crash. That’s why ground truth is crucial.

You need to find a way to get action-labeled data. The closer to ground truth, the better. Luckily, we have a fantastic starting point, thanks to Medal.

Before there was General Intuition, there was Medal

Earlier, we talked about the importance of games in the development of AI. AlphaGo. Deep Blue. These are intentional uses of games in AI.

There is an even richer history of accidental links between games and AI, lucky breaks.

Nvidia is the example you likely know. Jensen founded Nvidia to make chips for real-time graphics in games in 1993. Six years later, in 1999, Nvidia released its first “Graphics Processing Unit” (GPUs), the GeForce 256.

A few years later, around 2005, researchers began experimenting with GPUs for neural nets. In 2007, Nvidia released CUDA, to make ML on GPUs practical. In 2009, three Stanford researchers — Rajat Raina, Anand Madhaven, and Andrew Ng — showed that GPUs could accelerate deep learning by 70-100x for unsupervised learning.

Three years later, in 2012, the AlexNet team7 decimated their ImageNet competition using GPUs. Within a year, everyone in deep learning had switched to GPUs. “Everyone in deep learning” was still a small community at the time, but by then, Bitcoin miners had already been using GPUs. They were 50-100x more efficient for Bitcoin’s SHA-256 hashing than CPUs.

They soon switched to ASICs, but in 2015, Vitalik Buterin and his team released Ethereum, whose memory-heavy workloads were harder to optimize with ASICs. Ethereum mining ran on GPUs from 2015, through the GPU shortage it caused during the 2020-2022 crypto boom, right up until Ethereum switched from Proof-of-Work to Proof-of-Stake and left a GPU glut in its wake. Crypto tanked anyway, and in the same month crypto peaked, Nvidia’s stock did, too, tumbling 66% over the next year, until OpenAI released ChatGPT and, since then, Nvidia’s market cap has grown 10x to the $4.4 trillion behemoth we know today.

I mean, who could have predicted all of that?

When I taught myself to reverse engineer things, and learned how to code to build a private Runescape server when I was 13, I couldn’t have predicted that it would lead me to where I am today, either. Reverse engineering is the ultimate form of deductive reasoning, and spending a lot of time doing it as a kid is very good for your brain. This then lends itself well to figuring out complex systems in a rapidly changing world.

Runescape developers took the wilderness and free trade out of the game. I wanted to put it back, so I learned how to reverse engineer. The business that grew out of that did well for a teenager — we were making ~$1.5 million per year by the time I had to shut it down in 2015 at age 18, when I became an adult and would have been liable for the stuff I built. But I’d made enough money for my age that I could do whatever I was passionate about. I joined Doctors Without Borders (MSF) at 19 years old, and stayed for three years to work on Ebola & Humanitarian Mapping. I spent some time at Google Crisis Response before the gaming itch got me again.

At the time, we worked in London, very close to the DeepMind team. It was 2014 and I did not think it was that interesting or that likely to work. Demis deserves so much credit and respect for his vision. Few understand how hard it was for them to get here.

In 2018, I teamed up with my previous colleagues from building RuneScape servers. We built a game called Get Wrecked, which got a lot of signups. But it lost players quickly because we didn’t have enough player liquidity; it was a competitive game, and we needed to have enough people of all skill levels that people would always be able to find someone at their level to play against, which is very hard to bootstrap. To fix that problem, we built a way for people to watch game clips on the platform. A couple of times a day, we’d send a push notification that the game was live to get enough players playing at once.

The clips platform, Medal, went viral on the Rocket League subreddit. It was getting so many downloads that it became almost immediately clear that that was the bigger opportunity. We decided to focus on Medal.

We never ended up releasing the game. Medal just kept growing. Today, players from around the world upload 1B+ gaming clips to Medal each year.

We couldn’t have planned a better dataset with which to build World Models and policies.

Medal’s upload volume puts it on par with YouTube. Gamers upload millions of clips per day, across tens of thousands of environments, already hand-selected by players for highlights and adverse events. In other words, they share the content they think is worth sharing: their best performance, wildest encounters, closest calls.

Medal data has something that YouTube data does not. It comes enriched with metadata from our social network (views, likes, comments) and most importantly, in-game actions. We only record the game actions on the local machine, only storing the in-game action names (e.g. Move Forward) and never the keys that were pressed to achieve that action. More than data, this has enabled us to ship Medal’s most requested feature, keyboard and controller overlays. These overlays let our gamers showcase the precise actions they took behind every amazing moment.

Each clip has exactly what the player saw, next to the exact player actions that followed, using many of the same systems we use to control robots today. Frames from games also have the benefit of being information-complete. Unlike real world video, where you have to account for pose estimation (estimating what the human sees, which in itself is a lossy process) — you may see things the camera does not in the real world — but not in games. What is recorded and what you see is always identical, which we think makes it better training material.

This gives us trillions of examples of players running the loop of observe, predict, and act. This is the foundation of intelligence, and there is no loss of information throughout.

On Data

To understand what we’re doing, you need to grok the difference between game data and synthetic data.

The confusion is that people associate “digital” with “synthetic,” but the real distinction isn’t the environment in which the data was generated, but the data itself.

There can be synthetic (i.e., generated) data created in the physical world, like in the human-constructed environments Boston Dynamics and other robotics companies train their robots in, just as there can be human ground truth data in the digital world. Data breaks down into a quadrant like this.

What makes our gaming data “ground truth human data in a digital environment” is that what we are capturing is real human responses that observe → predict → act loop.

The closest comparison to our approach is GitHub data. It captured the history of human engineers’ coding and was used to train machines that can code better than humans. The question is whether that same idea works outside of the computer. We believe (and are seeing signs) that learning from gaming data transfers to the physical world.

Games turn out to be the perfect training ground for learning intelligence. They contain thousands of simulated worlds with physics, strategy, cooperation, text, interface use, competition, and long-horizon planning. They are complex enough to require intuition, but structured enough to learn from at a massive scale.

Physical-world data alone cannot reach the diversity or scale required to learn general intelligence. LLMs lack data on dynamics and atoms. But games act as the ideal intermediary: a bridge between the digital world of bits and the physical world of atoms.

Still, there is a threat to the ground truth stance. As mentioned earlier regarding Yann LeCun’s take, every video is action-labeled data if you’re good enough at inferring action. While this may be true in the long run, it’s also probably extremely impractical today. That’s why you gotta love Yann — nobody else would think to do it this way. Yann and I talked about this dilemma in Paris in December if you want to go deeper into the weeds.

Everything is a trade-off, right?

The optimal path forward is likely somewhere between where VLAs are today — the most practical but least elegant solution — and where AMI might be one day if everything goes well. It all comes down to your approach to data.

Data is the problem for any company that wants to solve embodied AI. Evan and Packy wrote about it in Many Small Steps for Robots, and it is the thing we are focused on at GI.

We believe our dataset is the most elegant answer to the data problem for general models. It is one that paves a path towards a general intelligence that feels familiar, the same way Tesla FSD feels like a familiar driver, but scales far beyond games or driving.

For general models, models that can power embodied AI intuitively and spontaneously in almost any imaginable real-world situation, the question is not simply how much data you can get.

Before throwing data at the problem, you need to understand your transfer curves.

Small Steps, Giant Leaps, and Transfer Curves

In their Robot essay, Packy and Evan wrote that there are two approaches to building economically viable embodied AI: Small Step or Giant Leap.

Evan and his company Standard Bots are pursuing the small step approach: getting paid to learn in the field, one use case at a time. They are collecting real-world data for a growing number of economically viable use cases across a wide variety of domains.

Their strategy is a fascinating one. By getting customers across many different industries and tasks to pay them to deploy, they collect diverse real-world data across wide distributions. Instead of hoping that more data in a narrow domain will generalize to tasks outside of the distribution, their goal is to put a ton of useful tasks in-distribution by going broad in the real world, not deep in one niche.

General Intuition and Standard Bots are coming at the same problem from opposite sides of the spectrum. General Intuition is attempting to solve generalization from the digital side: our bet is that gaming data will lead to broad priors about physics and actions. Standard Bots is attempting to solve generalization from the physical side: their bet is that real-world deployments will lead to broad priors about manipulation and industrial tasks.

These are complementary approaches to the data diversity problem. There’s potential for a GI World Model to be the starting point for Standard Bots’ post-training. We provide a base model trained on observed data in digital environments, which is scalable and economical to collect, and they post-train using the use case-specific data they get paid to collect to bring those use cases in-distribution and get to many 9s of reliability more quickly.

The approach that we find more challenging is the one that general models seem to be taking, which is collecting a ton of data and hoping it generalizes to out-of-distribution tasks. General models need too much data across too many situations to collect it all by paying people to demonstrate tasks.

Additionally, more data in the same domain doesn’t automatically teach a model to handle situations it’s never seen before. At pre-training, not all data is created equal, and I have not met someone building a general model for robotics who has been able to point me to the scaling laws that show that they can solve out-of-distribution use cases (things they weren’t trained on) simply by adding more data. More data in a narrow domain does not automatically buy you generalization to a new one.

The scaling laws don’t exist.

There are, as best we can tell, three distinct transfer curves that govern whether a World Model generalizes to new physical environments. They are not well understood — we are just starting to understand them. We can, however, give them names: input modality transfer, sensor transfer, and environment transfer.

The first is input modality transfer: How well does a policy generalize across degrees of freedom in the physical system it’s controlling? This curve is steep for a humanoid robot with somewhere between twenty and sixty degrees of freedom. Each is continuous and often mechanically dependent on the others, and this curve is steep. A finger movement is not independent of the arm. Training on a game controller and expecting it to transfer cleanly to a twenty-DOF humanoid hand is, in research terms, a bet without scaling laws to back it up.

The second is sensor transfer: If the workload requires specialized physical sensors (tactile feedback, proprioception, depth), there is a separate scaling law for how much of that sensor-specific data you need before the model can reliably reason about it. Tesla explicitly worked on this problem. They spent years figuring out exactly how much LiDAR data they needed before they could entirely drop the chips. Most robotics companies are working on it implicitly, hoping the answer reveals itself in deployment.

The third is environment transfer: How does performance degrade as the environment gets more complex, more stochastic, or more populated? Predicting the right action in a sports stadium with a thousand people around you is a fundamentally harder problem than predicting it on an empty field.

As we explained earlier, complexity doesn’t scale linearly.

These three curves interact. Until you can map them, you can’t know how much data of which type you actually need, which means you cannot justify the capital expense of going to collect it at scale. Companies that are collecting a hundred thousand hours of physical data today may find that a good World Model only needed ten thousand, or that they did need a hundred thousand, but ninety thousand of the hours they got were in the wrong distribution entirely.

Our bet, certainly conditioned on our starting position, is to collapse the problem.

By focusing on game controller inputs, we reduce input modality transfer to a curve we have already solved. We know we have enough data for game controllers, because we have billions of clips of humans using them. That eliminates one unknown. By focusing on vision-based inputs rather than specialized sensors, we eliminate the second.

Almost every physical system comes with a game-controller-like input modality, which includes steering wheels, keyboard and mouse and actual game controllers. Most are straightforward. Even humanoids ship with them. The challenge is just that if the degrees of freedom exceed what the controller can do, transfer is worse. So humanoids are further down our roadmap, but we see no physical limit to suggest that we couldn’t build around the interface limit.

In short: If you can control almost any physical system with a game controller, and we have more data than anyone else in the world of what happens when a player uses a controller to take action, our Agents should be able to control almost any physical system.

The only remaining question is about environment transfer: can Agents trained inside of the dream operate in reality?

The Superhuman Future of the World (Models)

It has been a wild few weeks at General Intuition’s offices in New York and Geneva. Everything that we’ve written about here has been working better than we expected. Like others, we’re gaining conviction that Agents trained inside of the dream can operate within reality.

Why do World Models transfer?

The observe-predict-act loop is an abstraction for how causally structured systems work in general. Once a world model has seen N variations of the world via a diverse set of games, it only takes a small amount of fine-tuning to understand the dynamics of the N+1 variant that corresponds to the real world.

World Models learn to model the cause-and-effect of reality. If this cause-and-effect is understood at a fundamental enough level, this should enable World Models to generalize to new scenarios.

What might that mean? What are the implications of World Models that generalize?

Our goal is to enable embodied AI to understand the world, with our models controlling machines in any environment, including the real world. We aim to deliver a breakthrough moment for robotics, where out of nowhere, the progress is obvious and the models are easy to use.

That breakthrough won’t look like the breakthroughs in LLMs, which went mainstream when they started talking to us like humans. We don’t want machines that simply do what humans do. In fact, the point of machines is to do things that humans can’t, to give us superpowers.

Robots don’t need to look like us to work for us. Humanoids as a form factor were largely chosen due to the assumption that they had the most data to learn from on the internet, because so many of humanity’s videos feature humans. If you don’t need those videos, if you can learn directly from the actions in video games across embodiments, and need a lot less to transfer to reality, that assumption doesn’t hold. We believe the future of robotics should be shaped by simpler, cheaper systems: machines with only the degrees of freedom that match the actual jobs to be done.

The human body is an incredible general-purpose platform, but it’s rarely the optimal (or most cost-effective) form for any specific task. Instead of copying our anatomy, we should mirror the interfaces we already use instinctively: joysticks, wheels, gamepads, and keyboards. These tools are the product of decades of iteration, compressing human intent into a clean, universal action space, much like language does for thought. Robots can learn from the actions transmitted through these interfaces and specialize around them in a very general way, making broad deployment far more practical than chasing full human embodiment.

If you get rid of the assumption that our machines don’t need to, and probably shouldn’t, mimic us or take our place in any way, a whole world of possibilities opens up.

At General Intuition, we’re actively working on simulations that will eventually allow our systems to go beyond everything that’s currently described in pixels, to everything governed by cause and effect. The methods we use are very general. This is a long way out, but a necessary step.

To really understand our world, it turns out, poetically, we may need World Models; compute for the uncomputable.

The implications of all of this are cosmic. If we can model three-dimensionality, physics, and time, and their interaction, then the ability to manipulate these arenas at superhuman macro and micro scales is on the horizon.

There is a tremendous amount of work ahead. Today, nobody is capable of simulating a biological cell, let alone an ecosystem made up of 10³⁰ of them. However, what captivates me is that we don’t need to map all of reality’s details. We just need to observe how those details manifest in actions, and use those actions to predict what comes next, over and over again.

There is also a tremendous amount of responsibility that comes with building these models, and it’s something I take very seriously and personally.

I’m from the generation most at risk of AI displacement; half my childhood friends can’t find jobs. I’m spending a lot of time exploring how we bring our community and my generation along in this shift.

For example, like Tesla, Medal sits on over $10 billion of global hardware infrastructure — GPUs, CPUs, plugged into power, with cooling — powered by over 15 million users. We’re actively exploring ways to let our community share in what’s coming, for example by generating income serving inference from their GPUs, or tele-operating from their gaming rigs. If the demand for general intelligence is anywhere close to what we think, that could be the largest economic tailwind my generation has ever seen.

These are just my dreams for now. But one day, they won’t be. One day, we leave the boring problems to superintelligence, so we can explore the stars or the deep sea from our gaming rigs, and dream up the next uniquely human, most interesting, not boring things to do.

If our work interests you, I’m recruiting to work on simulators to generate new environments, and we continue to be excited to talk to researchers and engineers who are in the top of their field and want to join General Intuition.

A special thanks to Eloi Alonso, Adam Jelley, Vincent Micheli, and Paula Wehmeyer, my co-founders and colleagues who spent many hours discussing the ideas behind this article. - Pim

Big thanks to Pim, Paula, Adam, Eloi, Vincent, Kent, and the whole General Intuition team for sharing their knowledge, and to Badal for the cover art. - Packy

That’s all for today.

For not boring world members, I played around with Claude to produce some extra goodies. We made a World Model Research Archive, with links to more than 30 key papers that have defined the space. Members can also ask Pim questions in the subscriber chat today and tomorrow.

Join us in not boring world for all of this and more by subscribing below

Thanks for reading,

Packy

Subscribe now

not boring world members can download the World Model Research Archive here.

Hi friends 👋,

Happy Friday! Here in New York City, it was the sunniest of times, it was the snowiest of times, but most of all, it was the incredibly hard to keep up with all of the incredible stuff getting done and funded of times.

Big week in robotics, solar, brains, world models, manufacturing, flight, AI, and even affordable luxury.

Let’s get to it.

Today’s Weekly Dose is brought to you by… Ramp Sheets

A few weeks ago, I wrote that I wrote that “a Ramp engineer with an AI will build something better than a CFO with an AI.” Ramp Sheets is exactly what I’m talking about.

See, a surprising amount of my high-leverage work ends up in a spreadsheet.

A deck comes in. A CSV export lands in my inbox. I want to benchmark a market, pressure-test a company, or run an analysis for a Deep. Dive. Usually, that means a lot of manual cleanup and model-building before I get to the actual insight.

Ramp Sheets skips a lot of that.

It’s an AI-powered spreadsheet from Ramp that lets you upload a PDF, pitch deck, CSV, or Excel file and ask it to do the work: build a model, research competitors, clean messy data, and help you get to an answer faster. For founders, investors, and operators, that means less spreadsheet mechanics and more actual thinking.

Best part: anyone can use Ramp Sheets today for free. You don’t need to be a Ramp customer to try it (but why wouldn’t you be?).

Try Ramp Sheets Free

(1) Swift Solar Acquires Meyer Burger to Build Gigawatt-Scale Solar Factory

for

“If your technology is so good, why aren’t you using it to compete?”

In Power in the Age of Intelligence, I wrote that that’s a question I’d like to see more investors ask. If the technology you’re building with is so much better, why in the world are you just selling it to someone else as a component? Why not just make a better end product?

Joel Jean at Swift Solar said, “Ok bet.”

I met Joel a few years back on an introduction from his brother Neal, the co-founder and CEO of Beacons. Joel, like Neal, is a genius, with a PhD in electrical engineering from MIT, and he was putting that genius to work by making silicon-perovskite tandem solar cells work.

The promise of perovskite is that they’re much more efficient than solar panels, meaning they can convert more incoming sunlight into electricity. Current commercial silicon panels are 22% efficient, with a lab record of 26.8% and a theoretical limit of 29%. Swift Solar’s panels are already at 28% efficiency. China’s Longi, the world’s largest manufacturer of monocrystalline silicon wafers, recently hit 34.8% with a silicon-perovskite tandem solar cell, and the theoretical limit, for which Swift is gunning, is 45%.

More efficiency means more watts per panel, which means more watts per acre and a lower balance-of-system cost; 30–40% efficiency modules could cut solar system costs 20–40% while doubling power density.

This all sounds great, but the problem is that tandems tend to degrade quickly outside, where the sun is. Swift thinks that they’ve cracked that, and that China hasn’t yet, which gives the west an opportunity (certainly not guaranteed, probably not even likely, but an opportunity!) to leapfrog back over China in solar production.

So if you have the magic, degradation-resistant tandem cells, what do you do? Use it to compete.

Swift announced that it acquired the core assets of Europe’s leading solar manufacturer, the German company Meyer Burger, to “build the next generation vertically integrated US solar manufacturer.” Meyer Burger’s former CEO joined Swift to lead the manufacturing effort.

Solar panel manufacturing is one of the most cutthroat, competitive industries in the world. China keeps driving costs lower and lower and lower. It’s been all about scale. Meyer Burger itself went bankrupt after having sold solar cell manufacturing equipment to China, only for China to reverse engineer and mimic the technology.

But I love it. Pressing a technological advantage into vertical integration and competing directly is the kind of thing that gets my heart racing. And this echoes a deal that hits close to home. A few weeks into Puja’s time at Harry’s, she told me the 10-month-old startup was acquiring one of the world’s oldest and best blade manufacturers, Germany’s Feintechnik. That has been an incredibly successful partnership.

There’s something about a the US innovation / German manufacturing tandem.

(2) Reflect Orbital Built a Space Mirror

Ben Nowack for Reflect Orbital

Imagine you are a photon that does not want to get turned into usable electricity on earth. It’s been a really bad week for you.

First, Swift’s tandem cells. Before, there was a 78% chance you’d escape once you hit the panel, but that looks like it’s going to fall. But at least you have nighttime, right?

At night, you can just fly off into space at the speed of… well, you know. Run away from Earth as fast as is physically possible.

Not so fast, says Reflect Orbital. CEO Ben Nowack just announced that the company had made a solar mirror, one of four huge guys that will go up on each satellite, stop the photons’ getaway, and redirect them down to Earth.

Sometimes, they’ll be used to light up a concert or a remote worksite in the evening (imagine how much less creepy True Detective: Night Country would have been if Reflect existed then), sometimes, they’ll shine on crops to help them grow faster, and mostly, they’ll bounce into solar panels that capture them and turn them into electricity.

Efficiency is one of solar’s challenges, but a bigger one is one that seemed almot insurmountable: the sun only really shines strongly enough for max solar output for 4-6 hours a day. What if it could shine on solar farms - out in the desert, away from homes (where it wouldn’t imapact our sleep, Huberman) - round the clock?

Solar farm economics would get ridiculous, solar would get much cheaper, we’d get a lot more of it, and we’d be able to power all sorts of new, better electric things.

Mostly, though, I’m sharing because that video is a bright spot in a sea of sameness.

(3) The First Multi-Behavior Brain Upload

Dr. Alex Wissner-Gross for Eon Systems

As big a week as it was for solar, it may have been an even bigger week for brains.

A couple of weeks ago, an Australian company Cortical Labs grew brain cells on a microchip and taught it to do the first thing hackers try to do on any new computer, from TI-83s to ATMs to pregnancy tests to vape screens: play Doom. And yup, it could play Doom.

On Tuesday, Dr. Alex Wissner-Gross shared a video made by a company he works with called Eon Systems, that one above, that shows a virtual fruit fly being controlled by a brain with 140,000 neurons and 50 million synaptic connections running on a computer, mapped from a real fly’s actual connectome.

The Eon team took the adult Drosophila brain connectome, the complete wiring diagram of a real fly brain, painstakingly mapped by the broader neuroscience community, combined it with a physics-simulated fly body built from an X-ray scan of an actual fruit fly, and got the two to talk to each other. Then sensory information from the virtual world enters the brain model, the brain’s neurons issue commands, and the body moves. It does all of the things a fruit fly would do in response to stimuli. It navigates toward food by “taste.” It grooms itself when virtual dust activates its “antennal mechanosensory circuits.” It “eats.”

The team is careful not to oversell the accomplishment, and Dr. Doris Tsao at Astera Neuro has a good list of thoughts on connectonomics and uploads here which basically says, “This is awesome, but we’ll get further by understanding instead of blindly mapping, and anyway, it’s going to be really expensive to map bigger and more complex brains.”

But still… this is sci-fi stuff. I don’t even know if it’s good. But it’s sci-fi!

Meanwhile, in regular old science, two very cool brain papers:

1. Isotonic and minimally invasive optical clearing media for live cell imaging ex vivo and in vivo. A team of Japanese scientists basically figured out a way to “make brains see-through without killing them.” Kording Lab explained: “The mechanism of this is so obvious and simple and yet brilliant. Match the refractive index inside and outside of cells - and tissue becomes transparent. Because the scaling of scattering with object size, small things don't matter much. So cool!”

2. Scientists revive activity in frozen mouse brains for the first time. A team in Germany made progress on cryopreserving and thawing mouse brains that leaves some of the processes necessary for brain functioning like neuronal firing, cell metabolism, and brain plasticity intact. A researcher at another lab cautioned that we’re still a long way away from freezing ourselves and waking up in the future, but he also said, “This kind of progress is what gradually turns science fiction into scientific possibility.”

OK, so maybe all of this is just very sci-fi.

(4) Emergent Quantization from a dynamic vacuum, or Zero Point Energy

Harold “Sonny” White et al, published in APS Physical Review Research, via Andrew Côté

Maybe my favorite rabbit hole I’ve fallen down over the past few years is the Zero-Point Energy Rabbit Hole, which I discovered via .

The idea is that at absolute zero, classical physics says everything should stop moving, but the Heisenberg uncertainty principle prevents any particle from simultaneously having a precise position and zero momentum, so it has to keep jiggling. That constant jitter is ZPE.

The vacuum, which we think of as empty, is a seething foam of energy, with virtual particles constantly popping into and out of existence. This sounds like woo, but it’s been experimentally confirmed via the Casimir effect: two uncharged metal plates placed extremely close together in a vacuum attract each other, because ZPE fluctuations are slightly suppressed between the plates compared to outside, creating a net inward pressure. This has been measured in real labs like Los Alamos.

Quantum theory predicts, and experiments verify, that empty space contains an enormous amount of ZPE. The amount of energy supposedly packed into the vacuum would constitute a seemingly ubiquitous energy supply, a “Holy Grail” energy source, which is really just about as close to a “Holy Grail” as we can get. You can do a lot with unlimited energy.

Stuff gets wilder from there in a chain of logic that leads from ZPE to mass, inertia, and gravity as engineerable properties of the vacuum to the ability to actually engineer the vacuum to create phenomena like antigravity, which is what some believe Thomas Townsend Brown did.

Anyway, a team led by Dr. Harold “Sonny” White, a physicist and aerospace engineer with experience at Boeing, Lockheed Martin, and NASA who is now the CEO of ZPE company Casimir, showed that “Instead of random noise, the quantum vacuum can have structure, meaning waves, dispersion, resonances,” per Andrew Côté, who added that “their analytical model reproduces all the 'larger scale' predictions and measurements of quantum mechanics exactly.”

Andrew said that while the difference sounds subtle, “it means the vacuum can have local structure, that it is in effect a dynamic medium,” which means that, potentially, we can pull energy out of it. That is what Casimir is trying to do.

Now look, Casey Handmer replied betting $1 million that it wouldn’t work, and who am I to bet against Casey Handmer.

But 1) there are a lot of smart people on the other side of that bet, like Andrew, Sonny, Hal Puthoff, etc… and 2) even if there’s a 1% chance there’s something there, it’s hard to think of anything bigger.

(5) Replit Launches Agent 4

Amjad Masad for Replit

Look, I’m biased. I’m an investor in Replit. I wrote about them in 2021 and predicted they’d be a $100 billion company. This week, they raised at a $9 billion valuation, so we’re getting there.

More importantly, though, they launched their new product, Agent 4. While I love Replit, and think Amjad is one of the best CEOs out there, I’ve been a little worried about vibe coding as a category. It’s so competitive, and won’t the big labs just do it? Won’t Claude Code kill everyone else?

Maybe. Certainly a possibility. But you should watch this video. By focusing maniacally on this one software creation product, the same thing they’ve been focused on making easy, from the infrastructure up, for over a decade, before AI was even a thing, they’ve built what I think is a truly differentiated product, and one that feels like one of the first good AI-native products I’ve seen.

You can collaborate in teams, set multiple agents to work at once, workshop designs, give feedback on specific ones, tweak them, iterate until you get something good, even doodle to show the AI what you want, and when you’ve found a design you like, you can keep it consistent across website, mobile app, presentations, docs, and more. It feels like a workspace for teams, where some of the teammates, the ones doing most of the grunt work at the humans’ direction, happen to be agents.

I’m going to be playing with it this weekend. If you have some time, build something.

EXTRA DOSE: Enormous Week of Fundraising + Nietzsche as Mystic for not boring world subscribers (cmon, join us)

Read more

Hi friends 👋,

Happy Friday and welcome back to our 183rd Weekly Dose of Optimism.

I was at a16z’s American Dynamism event in DC this week, where I got to meet a bunch of people who read not boring, including one who told me she never opens the Weekly Dose because the title doesn’t tell her anything.

If you’re reading this: thanks for the feedback, I guess it worked. I’m running a title test; I’ll check open rates this week and decide what to do going forward.

Those of you who opened are in for a treat. What a week for the optimists.

Let’s get to it.

Today’s Weekly Dose is brought to you by… Guru

Guru is the best way to make sure that your employees get accurate company knowledge from whatever AI you use.

Companies like Spotify and Brex use Guru because it’s the only AI verification system that automatically validates company knowledge before your AI agents use it, like quality control for your AI’s brain.

Like them, your company’s systems weren’t built to power automated systems. Your people send Slack messages, write docs, and even quasi-maintain wikis for other people to consume. Guru translates all of it to speak AI so that AI doesn’t give your people incorrect answers ~40% of the time.

Put your messy company data to clean use in a couple of clicks. Get a Guru to do it for you.

Try Guru Today

(1) a16z American Dynamism Summit

On Tuesday, I Joe Bidened myself1 down to Washington, DC for a16z’s American Dynamism Summit.

Complete with its very own Jason Carman hype video (see above), it was like a living, breathing version of the Weekly Dose of Optimism, and full of the people behind the stories we cover here each week.

Drew Baglino, whose Heron Power led off the Dose last week, was there, as was Noah Smith, whose Decade of the Battery I linked in that story. Cameron McCord, whose fundraise for Nominal you’ll read about today, was there, too. I spent time with probably a dozen people whose work has appeared in the Dose, and got to watch others on stage, like NASA Administrator Jared Isaacman.

My job is to spend time with these people, write about them, and invest in some of their companies, but there was something about seeing so many of them in one place that brought home the fact that a small group of sufficiently talented and motivated people can change how the future plays out.

Power generation, transmission, and storage, motors, power electronics, autonomous boats, supersonic planes, educational choice, modern manufacturing (terrestrial and in space), new weapons systems, cheap and secure communication everywhere, autonomous everything, safer neighborhoods, new cities altogether. If the people in that room succeed, America’s next 250 years will be better than its first 250.

(2) Physical Intelligence Models Get Memory So Robots Can Do Complex Tasks

Last week, I was talking to a Physical Intelligence investor who told me the progress the company had made in just the first couple of months of 2026 has blown his mind. Now, we know why.

Pi gave its models a new memory system that helps both short-term visual memory, to recall what it did recently with fine-grained details, and long-term semantic memory, which lets it remember what it did for up to 15 minutes.

With better memory, robots using Pi’s model can do long, multi-step things, like clean up an entire kitchen, set up the ingredients for a recipe, and grill a grilled cheese sandwich. It also learns from its past mistakes.

Yesterday, I wrote that LLMs are producing a lot of dead text that we’re probably better without. Robotics is way more interesting. I would love for the robots to spend billions of tokens doing my laundry and making me healthy gourmet meals while I kick back and read human-written classics. This is one more small step in that direction.

For a full primer on robotics, check out my cossay with Evan Beard:

(3) Arc Institute Releases Evo 2, Fully Open-Source Biological Foundation Model

A year ago, the Arc Institute released Evo 2, the largest fully open biological AI model ever built, as a preprint. This week it was published in Nature, and Arc dropped a one-year recap showing what the thing has actually done in the wild.

Trained on 9.3 trillion nucleotides from over 128,000 whole genomes, Evo 2 can read and write across all three domains of life: bacteria, archaea, and complex eukaryotes including us. The architecture, called StripedHyena 2, trained on 30 times more data than its predecessor Evo 1, and can reason over 8 times as many nucleotides at a time. Greg Brockman, the OpenAI co-founder and ex-Stripe CTO (Stripe co-founder Patrick Collison is a founder of Arc), helped build it during a sabbatical.

Evo 2 generalizes “across biological prediction and design tasks, across all modalities of the central dogma, across molecular to genome scale, and across all domains of life.”

It can predict which genetic variants cause disease and which don’t, without being specifically trained on that task. In tests on the breast cancer gene BRCA1, Evo 2 achieved greater than 90% accuracy in predicting which mutations are benign versus disease-causing.

It can generate new genomes. Entire genomes. The team demonstrated the first AI-designed and experimentally validated bacteriophage: 16 of 285 tested designs successfully propagated and inhibited growth of the target bacteria, with no impact on unrelated strains. For those who, like me, don’t fully speak the language of biology, that’s a whole functional new organism.

And in a stunt that’s both scientifically real and deeply fun: the team guided Evo 2 to design DNA sequences with controllable chromatin accessibility patterns, then wrote Morse code messages—”EVO2,” “LO,” and “ARC”—into the epigenome of mouse embryonic stem cells, experimentally validated with ATAC-seq.

The whole thing is fully open: model weights, training code, inference code, and the OpenGenome2 dataset. So more scientists will be able to experiment with a model that understands the grammar of life as the cost of reading and writing biology keeps falling. The design space for medicine, agriculture, and materials is going to keep opening up in ways that are genuinely hard to imagine.

(4) Injectable “satellite livers” could offer an alternative to liver transplantation

Sangeeta Bhatia Lab / MIT / Cell Biomaterials

More than 10,000 Americans are on a waitlist for a liver transplant. There aren’t enough donated organs. And many patients with liver failure aren’t even eligible for the surgery because they’re too sick to survive it. For those people, the best current answer is: sorry.

MIT engineer Sangeeta Bhatia’s answer is: what if we just injected a second liver?

Her lab has developed “satellite livers,” tiny engineered tissue grafts made of hepatocytes (the liver's workhorse cells) mixed with hydrogel microspheres and supportive fibroblasts, all deliverable through a syringe. Once injected into fatty tissue in the abdomen, the spheres act like a liquid going in and then solidify, giving the hepatocytes a scaffold to cluster around. Blood vessels grow in and the mini liver takes root. In mouse studies, the grafts remained viable and producing the right proteins for at least eight weeks (the full length of the study).

The insight that makes this work is a clever one: the hydrogel microspheres create an “engineered niche” for the cells. Without them, injected hepatocytes disperse and fail to integrate. With them, the cells localize, connect to the host circulation faster, and behave the way liver cells are supposed to.

The vision is a twofer. Short-term, these grafts could act as a bridge to transplantation, buying time for patients waiting on the donor list. Long-term, they could be the treatment: repeatedly administered, minimally invasive, titrated to how much liver function a patient needs. It could replace surgery altogether with routine injections.

The liver does about 500 essential things for the human body. The transplant chain is broken. People die. The fact that we might soon be able to restore meaningful liver function with a syringe, a handful of cells, and some hydrogel spheres is miraculous.

(5) Ginkgo Bioworks Launches Ginkgo Cloud Lab

In early February, we shared Ginkgo’s collaboration with OpenAI in which GPT-5 integrated with Ginkgo’s autonomous lab and autonomously designed, executed, and analyzed over 36,000 cell-free protein synthesis experiments. GPT-5 reduced costs by 40% compared to benchmarks (from $698 to $422 per gram). Good robot.

Now, Ginkgo is making its Cloud Labs available to anyone in a push to move all of its R&D services away from traditional benches and to these futuristic ones. Building out a lab can cost a bunch of money, Ginkgo CEO Jason Kelly tweeted that it can cost $10 million, and Ginkgo’s strategy has long been to offer lab services to startups, academic labs, and pharma companies as a service.

We hope that Cloud Labs will someday allow anyone to be a scientist with their own lab just like personal computers and cloud data centers democratized programming and the web.

With this move, presumably the service gets better as Ginkgo’s margins do, too. The starting price for Ginkgo’s first three protocols, two for cell-free protein expression and one for bacterial pixel art, is just $39.

Awesome news for the democratization of science, although I can’t imagine this is going to help my friend sleep any better at night.

Noahpinion

Updated thoughts on AI risk

So the other day I wrote a post about how humanity is inevitably going to be disempowered by the existence of AI…

Read more

2 months ago · 408 likes · 100 comments · Noah Smith

EXTRA DOSE: Ulkar is out for the week, so in our Extra Dose, I’m rounding up a big week in fundraising news for software for hardware and sight restoration, plus recommending my favorite sci-fi series of all time.

Read more

Subscribe now

Hi friends 👋,

Happy Thursday! It’s late in the week and late in the morning, so…

Let’s get to it.

Today’s Not Boring is brought to you by… Deel

Every week, at the bottom of the Weekly Dose, I thank my teammates, Aman and Sehaj, for their contributions. Thanks isn’t enough though. They do great work, for which they deserve to get paid. Challenge is: they’re based in India. Or, challenge was. Now, I pay them with Deel.

Deel is payroll for global teams, like not boring and many of our portfolio companies. I grabbed coffee with a founder in our portfolio the other day, and unprompted, he brought up how easy Deel made it to expand his team globally. He’d tried the alternatives and switched, and he was gushing, which is rare for a payroll provider.

Hiring talent outside your home country can open up new growth, but it also comes with unfamiliar rules, local labor laws, and compliance risks. That’s where an Employer of Record (EOR) comes in. If you’re a startup exploring global hiring for the first time, Deel’s free guide walks you through the basics. Whether you’re hiring your first international employee or planning your next market expansion, Deel can help.

Download the Guide

Costless Sacrifice

The Old Testament’s 2 Samuel tells the tale of King David as he unifies the tribes of Israel and establishes Jerusalem as the nation’s capital. The world was slower then, there were only like 50 to 75 million people around at the time, and so God could be more actively involved in the day-to-day management of human affairs.

As, for example, when King David ordered a census of Israel to put a number on its military strength, he angered God, who expected his servant to rely not on soldiers but on Him. God sent a plague down that killed 70,000 men, then instructed David, through the prophet Gad, to build an altar on the threshing floor of Araunah the Jebusite in order to stop it.

When David arrived, Araunah offered to just give him the land, the oxen, and the wood with which to sacrifice them for free as a gift to the king.

“But the king replied to Araunah, ‘No, I insist on paying you for it. I will not sacrifice to the Lord my God burnt offerings that cost me nothing.’ So David bought the threshing floor and the oxen and paid fifty shekels of silver for them.”

Costless sacrifice is not sacrifice. And we, like the Gods, demand sacrifice.

I read an essay in The Argument by economist yesterday: The Tinder-ization of the Job Market.

The Argument

The Tinder-ization of the job market

Is AI going to take everyone’s jobs? A recent report went viral after claiming there were 55,000 AI-related layoffs in 2025…

Read more

a month ago · 69 likes · 19 comments · Matt Darling

The job market is stuck, Darling argues. It’s not bad: unemployment was relatively low in January at 4.3%; Prime Age (25-54) Employment at 80.9% is higher than it was at any point in the Obama or first Trump presidencies. Just stuck. The hiring rate averaged 3.3% in H2 2025, a level only touched during COVID and the Global Financial Crisis.

BLS Hires Data

The weird part is… we’re not in the middle of a crisis, financial or biological. Employment is high, remember! Unemployment is low! “Generally, a high employment rate and a ‘tight’ labor market are associated with high hiring rates,” Darling writes, “not low ones.”

Darling’s theory is that LLMs have made it easier for people to apply for a ton of jobs, with custom-written cover letters and everything.

He cites Greenhouse data which “showed recruiting workload rose 26% in the third quarter of 2024 and that 38% of job seekers reported ‘mass applying,’ flooding firms with far more resumes than before.” Per Business Insider, “the applications-to-recruiter ratio is now about 500-1, four times what it was just four years ago.” He links to a Kelsey Piper essay from last year that provides more data to back this up.

Because it’s easier to apply, the volume of applications is up. Because AI is writing cover letters, their quality is no longer a useful way to pull signal from all that noise. “A recent paper showed that after Freelancer.com introduced an AI-generated cover letter tool, the correlation between cover letter customization and offers dropped 79%.”

You used to get rewarded for customizing your cover letter, when it cost you something. Now it doesn’t, so you don’t, but you still gotta customize, because everyone else is.

We find ourselves running an incredibly stupid Red Queen’s Race.

“Here’s the hard thing about easy things: if everyone can do something, there’s no advantage to doing it, but you still have to do it anyway just to keep up.” I wrote that way back in August 2020, about DTC brands and Shopify. “When every rebel is armed, none really is. It’s like when you played GoldenEye 007 as a kid. Getting the Golden Gun the hard way was dope. Everyone getting the Golden Gun with a cheat code made the game suck.”

This is an idea I’ve been obsessed with for a long time: the asymmetric ability of the laggards to make the leaders’ life a little bit worse.

You might produce the very best widget by far, but your potential customers will undoubtedly be bombarded by inferior alternatives claiming that they make the very best widget. At worst, your potential customer will fall for it; at best, your cost to acquire that customer goes up a little bit.

You spend hour after quiet hour handcrafting your widgets, smoothing their edges, polishing their faces, giving each one a little kiss before sending it out. You sacrifice. Your competitors do none of that. They yeet millions of these suckers out in Yiwu. Then they stamp HANDCRAFTED WITH LOVE IN USA on their website and who’s to say?

Some people really want a particular job. It’s their dream job. They spend hour after quiet hour poring over resume details, crafting a heartfelt cover letter, and saying a little prayer before hitting “Submit.” They sacrifice. Their sacrifice gets lost among the 1,732 less-lovingly-crafted (but who’s to say? who has the time to read them all) applications that came in that same day.

The burnt offerings that cost nothing and the burnt offerings that cost everything smell the same, because we make our offerings not to an omniscient God, but to fallible, overwhelmed humans.

This is happening more and more with writing, too. Claude in particular has gotten much better at writing, which means that more people are publishing essays.

There will be times when there’s an interesting topic to write about, and by the time I’ve thought about what path I might take through the argument, there are dozens of versions of the essay on X. Some of them are even pretty good!

But I’ve noticed something happening in my brain that’s worth sharing, because I would imagine a lot of people are thinking similar things about whatever it is that they do.

I’ll be interested in a topic, start chewing on it, start thinking through unique ways to present it, to shape the essay, the research I might need to do, the people I might need to talk to, all this stuff I’d need to do to make something great. Then I see a handful of pretty good versions on similar-ish topics, and I think, “Well, I guess they beat me to it. On to the next one.”

And who knows if the version I would have written would have ended up being better than what these AI-human teams whipped up in a couple hours. Who knows if it would have been great, or even good.

What I do know is that if I’d written it, I would have put in a lot more effort, agonized over it, sweat the details, tried to present the ideas in unexpected ways. I would have sacrificed something that cost me something, hours and hours, sometimes weeks and weeks of my life.

And half of the value of the post would have been the result of that effort, but half would have been the effort itself, my pointing to an idea and saying, “THIS IS AN IDEA THAT I WAS WILLING TO SPEND FIFTY HOURS WRESTLING WITH.”

And maybe I’ll still spend the time to write it, and maybe many someone elses will spend the time making the full and beautiful version of the things they want to create despite the easy-come existence of their bloodless simulacra, but a lot of the time, we won’t, and the world will be left with many hollow versions of a thing filling up the place where the one full one should have gone.

The pretty good is the enemy of the potentially great.

Right, like someone will somehow, eventually, once the recruiters have gathered themselves up and faced the never-ending application pile, basically just win the lottery and get offered the job, and they’ll be happy that they got it, because a paycheck is a paycheck, but two things will probably be true: 1) it wasn’t their dream job; they applied to 347 of them, this was the one they got, and 2) the person whose dream job it actually was, who would have been overjoyed to get to work in this specific job, who actually did write the cover letter the old-fashioned way for this one because they reallllllllly wanted it… no human ever even saw that person’s application. They’re still looking. Dream crushed, they’ve now applied to 123 jobs themselves, none of which they really care about, but all of which someone really wants and now is a little bit less likely to get.

I don’t know code as well as I know words, but it looks like the same story is playing out there. SemiAnalysis has started publishing this chart of the number of Claude Code GitHub Commits Over Time. The number, let me tell you, is going up. It’s going up big time.

SemiAnalysis

“4% of GitHub public commits are being authored by Claude Code right now. At the current trajectory, we believe that Claude Code will be 20%+ of all daily commits by the end of 2026. While you blinked, AI consumed all of software development.” Claude Code Is The Inflection Point, is what this means.

And again, I am not a coding expert, but like…

Of course the machines that can happily print tokens at high speed 24/7 are going to produce more code than us meatbags. I can’t imagine what a similar graph of “% of Words Committed to X” would look like. We have to be approaching 50%; we might be way past it. That does not mean the Singularity is Near.

I think if you asked me how I’d position myself on an AI trade, it would be something like: short ASI, long tokens.

As so perfectly captured in Tool Shaped Objects:

The market for feeling productive is orders of magnitude larger than the market for being productive. Most people, most of the time, want to click and watch the number go up. They do not want to be told the number is fake. They will pay— in time, in attention, in actual money— to keep the number going up.

But his is a “come for the tool” view of token demand, and misses the “stay for the network” part. It’s not just that we want to watch our own number go up. It’s that if everyone else’s numbers are going up, we need ours to go up just to keep up.

More words, more applications, more code, MORE, in the world’s stupidest Red Queen’s Race.

I don’t blame the tools. We do this all the time. Venkatesh Rao wrote about Premium Mediocre just two months after Attention is All You Need and years before the transformer’s significance became apparent. Kylie Jenner made her face look a certain way because she could afford it, then everyone started doing it, and it got more affordable, and now Instagram Face has become middle class, a costless sacrifice to the gods of vanity, signifying nothing.

Like, what does the number of Claude Code GitHub Commits signify? What does the number of words written by an LLM signify? What do thousands of applications for every single low-wage job signify?

If they cost nothing, they signify nothing.

I started reading this book recently, The Control Revolution by James R. Beniger, written in 1986 and, from what I’ve read so far, criminally underread today. I haven’t read much of it yet, so I’ll need to save a deeper analysis for a future essay, but Beniger has this idea that is relevant to our conversation.

He believes that modern information technology was built in response to the Industrial Revolution, a direct response necessitated by industrial scale and complexity:

Until the Industrial Revolution, even the largest and most developed economies ran literally at a human pace, with processing speeds enhanced only slightly by draft animals and by wind and water power, and with system control increased correspondingly by modest bureaucratic structures. By far the greatest effect of industrialization, from this perspective, was to speed up a society’s entire material processing system, thereby precipitating what I call a crisis of control, a period in which innovations in information-processing and communication technologies lagged behind those of energy and its application to manufacturing and transportation.

Modern bureaucracy, the telegraph and telephone, mass media, computing, sensors… all of this, Beniger says, fell out of the crisis of control precipitated by the Industrial Revolution. The Industrial Revolution massively accelerated production. We could make things faster than ever. But then we had a distribution crisis: how do we move all this stuff? Railroads and logistics solved that. But then we had a communication crisis: how do all of those far flung hubs coordinate with each other? Enter telegraphs, telephones, train schedules, standardized clocks, and the bureaucracy to manage it all. And a consumption crisis: how do we get people to want and buy all this stuff we’re producing, all over the world? That’s where mass marketing, advertising, and eventually consumer culture come from.

All of which I bring up because it seems that whatever the Crisis of Control was, where we did not have enough information technology to deal with the material abundance we were creating, we have the opposite Crisis now, where the information has outstripped that which it was born to control.

Like going off the gold standard, but for information. We can mindlessly and costlessly print information with no reference to the underlying stuff about which it was meant to inform.

This is the fast takeoff, the runaway scenario: information has reached escape velocity. Claude Code GitHub commits are going stratospheric. But what is the relationship between GitHub Commits and Total Factor Productivity? Between GitHub Commits and actual economic output?

A sacrifice needs to cost something.

The bureaucracy was established on the backs of the hard, provable work of making things. Today, the bureaucracy is the thing, the simulacra of productivity that has become more real than the thing itself, and it lets society make things occasionally, if slowly.

An essay is valuable insofar as it costs the writer something. They may have paid the costs in the years of experience that seasoned the piece that took hours to write, like that Picasso story where he asks for $100,000 for a picture that took him 30 seconds to draw. They may pay the costs in hours and hours, weeks and weeks of researching and writing and editing and re-writing and crumpling and throwing into waste basket overflowing with previous drafts and re-writing again, and again, until it’s good enough.

Even if a machine wrote the same essay, word for word, it wouldn’t have paid the same cost. You can read the costlessness of it, feel it, even as the prose improves. We demand a cost.

This is what King David realized. Fifty shekels wasn’t much money to him, and God, being God, could have divined KD’s sincerity with or without the coin. But a sacrifice has got to cost something.

The bad news is: it seems that we’re in a downward spiral. Systemically, I don’t know how we pull out of this. There will be many, many more Claude Code GitHub Commits and Claude-written words in 2026 than there were in 2025, many more in 2027 than 2026, etc.

Anything can look like the Singularity with a dumb enough y-axis.

The good news is: you don’t need to play this game. You can make yourself pay a cost and you will be rewarded: externally, maybe. Internally, for sure. Eternally, I think so.

There is a school of thought that believes humans were brought into this universe to create. To live and struggle and love and sin and pay all of the costs required to create new things. The reason we are here, this belief, is precisely to experience the limitations and frictions that God cannot. We are here to pay the cost.

There is a joy that comes from conviction, and meaning in doing something only you can do, well. And somehow, that soul shines through. People can feel its presence as strongly as they can feel its absence.

I find that whenever I go comically over the top in the amount of work I do on something, spending a month to write an essay that AI could write ~75% as well in minutes, for example, that sacrifice is rewarded. People want to see that you care, that the work cost you something, that you had to make a choice, in order for them to pay a cost to you.

Pay the shekels.

That’s all for today. If reading this made you want to hire real humans, check out Deel.

We’ll be back in your inbox tomorrow with a Weekly Dose.

Thanks for reading,

Packy

Hi friends 👋,

Happy Friday! It’s quasi-warm in New York City, America won two golds in ice hockey (sorry, Sean, Dan didn’t have anything funny to say but we’re pumped), and the good guys just keep on doing things that make us optimistic.

Let’s get to it.

Subscribe now

(1) Form Gets $1 Billion Google Order for 30 Gigawatt-Hour Battery System

Steve Levine for The Information

The Decade of the Battery is charging ahead, now with American batteries.

Earlier this week, Google announced a deal with Xcel Energy to provide 1.9 GW of wind, solar, and battery power for a planned data center in Pine Island, Minnesota, part of a push for hyperscalers to Bring Your Own Electricity (BYOE). Wind and solar are clean, and cheap when the sun shines and the wind blows, but the sun doesn’t always shine, and the wind doesn’t always blow, so the deal includes $1 billion for a nine-year-old startup Form Energy to provide backup.

Form will be providing two kinds of batteries: standard lithium-ion batteries for instant surges of power, and iron-air batteries for longer-term backup.

Iron-air batteries work through a process that’s like reverse rusting. To discharge, the battery breathes in oxygen from the air. Iron metal at the anode reacts with that oxygen and water-based electrolyte to form iron rust (iron hydroxide/iron oxide). This oxidation reaction releases electrons, which flow through an external circuit as electricity. To charge and store electricity, you apply electricity (say, from solar or wind), and the process reverses. The rust is electrochemically reduced back into metallic iron, and oxygen is released back into the air. The iron is “de-rusted” and ready to discharge again.

Iron-air batteries are low power density and slow response, but extraordinarily cheap on a per-kWh basis, roughly 10% of the cost of lithium-ion. They’re perfect for longer-term storage, up to 100 hours, for multi-day lulls in sun and wind. These lulls are whimsically named “Dunkelflaute” events.

Form will provide a 300 MW iron-air battery system for the project, and the batteries can store 100 hours of power, making it a 30 GW-hr system. It will be the biggest battery system by energy capacity in the world when delivered.

This is good news because we love batteries here at Not Boring, and because it’s a rare win for a western battery company in a category that’s been dominated by China. That also means good US jobs; the batteries for this project will be manufactured at Form Factory 1 in Weirton, West Virginia, on the site of a historic former steel mill.

So yeah, maybe AI is going to take all of our white-collar jobs (kidding, we’re on Team Thompson), but powering all that AI is going to create new ones, too.

(2) Proxima Fusion Signs MoU to Build Stellarator Fusion Power Plant

Proxima Fusion, a German startup out of the Max Planck Institute, signed an MoU with the Free State of Bavaria, RWE, and Max Planck Institute for Plasma Physics (IPP) to put the world’s first commercial stellarator fusion power plant on the grid in Europe.

It will need €2 billion to build a demonstration facility, Alpha, prior to the commercial plant, Stellaris, €400 million of which will come from Bavaria and up to €1.2 billion of which may come from the German government. It’s an expensive way to say “Es tut mir Leid” for shutting down all of the country’s nuclear power in favor of coal, but a very cool way to get German energy back on track.

Of course, it’s just an MoU. The plant still needs to be built, and then there’s the tricky matter of achieving Q>1 and, eventually, Q> whatever it needs to get to to be economically viable, but we love the stellarator.

Stellarators are a type of fusion reactor. In The Fusion Race, we wrote:

Designed by Lyman Spitzer at the Princeton Plasma Physics Laboratory, Stellarators are devices designed to confine hot plasma within magnetic fields in a twisted, torus-shaped configuration to sustain nuclear fusion reactions.

Stellarators, while promising, were difficult to design and build due to their complex magnetic field configurations.

But that was the 1950s! We have better technology today. On Age of Miracles, Julia and I interviewed Proxima CEO Francesco Sciortino, who told us something that’s stuck with us: we made Tokamaks first not because they were best, but because they were the easiest to design and manufacture. Stellarators, with their weird twists, were harder to design and manufacture, but closer to the platonic ideal of a fusion generator. Now, with better software and manufacturing capabilities, we can make reactors closer to the platonic idea.

Proxima has a long way to go before it’s putting electrons on the grid, but we’re happy to see progress towards that goal. Gut gemacht, Germany.

(3) Magical Month for Psilocybin

Psilocybin, the psychoactive compound in magic mushrooms, has been classified as Schedule I drug under the Controlled Substances Act since 1970. The classification means that it has “no accepted medical use and a high potential for abuse.”

That “no accepted medical use” part is coming under intense pressure from reality. Last week, Compass Pathways announced that it had achieved its primary endpoint in a second Phase 3 trial evaluating COMP360 psilocybin for treatment-resistant depression. Two doses of COMP360 25mg demonstrated a highly statistically significant reduction in depression symptoms versus control (p<0.001, -3.8 point MADRS difference). This makes Compass 3-for-3 on trials. The company plans to meet with the FDA to discuss a rolling approval submission between October and December 2026, which would make COMP360 the first “classic” psychedelic cleared in the U.S.

The week prior, a Johns Hopkins pilot study of 20 adults with well-documented post-treatment Lyme disease found that psilocybin, given with psychological support, produced significant and lasting reductions in multi-system symptom burden, including improved mood, fatigue, sleep, pain, and quality of life, that persisted for up to six months. The numbers are striking: general symptom scores dropped roughly 40% and mental/physical quality of life scores rose about 13% from baseline, with benefits maintained through the six-month follow-up. The durability is noteworthy, because a placebo would typically diminish over six months.

The Lyme study is especially interesting because it extends the psilocybin evidence base beyond the depression/anxiety lane into a chronic neuroimmunological condition where there are essentially no accepted treatments. If psilocybin is doing something meaningful for post-treatment Lyme, it starts to suggest mechanisms (anti-inflammatory, neural connectivity remodeling) that go well beyond the “it helps you process emotions in therapy” framing.

From a pharma business perspective, one of the biggest challenges with psilocybin is that it just works. There’s no opportunity to keep patients on (and paying for) meds for the rest of their lives like there is with SSRIs. That’s all the more reason we should be taking psilocybin very seriously. If this same logic applies to other conditions, like Lyme, it could be a major win for humanity.

A miracle drug like GLP-1 whose side effects include euphoria and spiritual experiences instead of muscle loss and upset stomach would be magical indeed.

(4) Stripe’s 2025 Annual Letter

Patrick and John Collison for Stripe

Stripe published its annual letter this week. It’s one of the best pieces of economic writing you’ll read all year, because Stripe has some of the world’s best data on the internet economy, and because those Irishmen can write.

Last week, in Power in the Age of Intelligence, I made the case that tech-native category leaders like Stripe are more valuable because they’re going to eat more of their categories than second-best. Stripe is eating. They shared that they did $1.9 trillion in total payment volume in 2025, up 34% year-over-year, and that roughly 1.6% of global GDP is flowing through the company's pipes.

More interesting is the data they have on companies using Stripe. Their 2025 cohort of new businesses is growing 50% faster than the 2024 cohort. The number of companies hitting $10 million ARR within three months of launch doubled. iOS app releases jumped 60% year-over-year in December. GitHub pushes surged 41% between Q3 2024 and Q3 2025. All of their indicators suggest that the pace is accelerating, and the Collisons’ best guess is that it isn’t an anomaly.

The whole letter is worth a read, but three sections stand out.

First, their framework for “agentic commerce” lays out five levels, from agents that just fill out checkout forms for you all the way up to agents that anticipate what you need and buy it before you ask. They’re honest that we’re hovering between levels 1 and 2, but the comparison to the mid-90s — when HTTP, HTML, and DNS were being hashed out — feels apt.

Second, crypto is happening, even if crypto prices aren’t. Bitcoin is down 50% from its October peak, but stablecoin payments volume doubled to around $400 billion last year, with 60% estimated to be B2B payments. Bridge, the Not Boring Capital portfolio company acquired by Stripe, quadrupled its volume. Stripe is betting on stablecoin payments infrastructure hard with Tempo, a built-for-payments blockchain they’re building with Paradigm. Visa, Nubank, Shopify, and even Klarna (whose CEO was once a self-proclaimed crypto skeptic) are already testing it. Mainnet is launching soon.

Third, the letter closes with what might be its most important idea: we live in a “Republic of Permissions.” Technologies succeed or fail not just on their merits but on whether the web of regulators, committees, and courts lets them through. The Collisons cite Joel Mokyr’s Nobel-winning work on how culture, not just capital or technology, drives progress. They argue that AI could transform drug discovery, nuclear could deliver energy abundance, and drones could slash logistics costs, but only if we don’t let “a slurry of local ordinances harden into a blockade.” Hear hear.

It’s a letter about payments that’s really a letter about whether civilization can keep up with its own tools, co-written by a guy whose website has the canonical list of projects from a time when we were able to build fast. They think we can.

“We’re reminded of the phenomenon of falling into a large black hole… We write this letter at what may well turn out to be the advent of a different and hopefully much more beneficent singularity. While much around us in 2026 feels similar to prior years, it is also clear that the next decade will look very different to those just gone by.”

(5) Vitamin B2 and B3 nutrigenomics reveals a therapy for NAXD disease

Arc Institute in Cell

Pulling one of the most interesting entries from Ulkar Aghayeva’s Scientific Breakthroughs this week above the paywall. Here’s Ulkar on a fresh finding out of Arc Institute:

It may seem that vitamin biology has been figured out decades ago (for reference, all 13 classical vitamins were discovered by 1948). But this paper shows that there’s still a lot of unexplored territory in nutritional genomics.

Instead of starting out with a target disease and trying to find a drug that treats it, they chose the well-known vitamins, B2 and B3, and did a genome-wise CRISPR screen in K562 cancer cells to identify genetic diseases responsive to vitamin supplementation. NAXD, a repair enzyme essential in redox biology, emerged as the top hit for vitamin B3. Mutations in NADX are known to be lethal in early childhood. The team generated knockout mice that showed very similar disease profile to humans, and adding vitamin B3 to their food from birth increased their lifespan more than 40-fold.

How many other diseases are out there that could be so easily cured with various vitamin supplementations?

Additional sources: Arc Institute blogpost; twitter thread

EXTRA DOSE:

• Scientific Breakthroughs from Ulkar

• Jeremy Stern on Shyam Shankar

• Alex Konrad on William Hockey

• Mario joins Hummingbird

Read more

Hi friends 👋,

Happy Friday and welcome back to our 181st Weekly Dose of Optimism!

This week has it all, from transformers to aliens.

We are also introducing a new segment for not boring world members: Scientific Breakthroughs from Ulkar Aghayeva. In December, we shared Frontier of the Year 2025, in which Ulkar, Gavin Leech, and Lauren Gilbert reviewed 202 pieces of scientific news from the year and assigned a score based on a probability each generalizes and how big it would be if they did. We loved it, so we asked Ulkar to do a roundup of the most important stories in science, which she did this week.

It’s our attempt to bring you even closer to the frontier. I hope you enjoy it.

Let’s get to it.

Today’s Weekly Dose is brought to you by… not boring world

The Dose is for everyone, but we’re adding more and more great stuff behind the paywall for not boring world members. We also have a deep slate of co-written essays in flight and going out to members over the next couple months. If you want to support not boring, help us make it as great as it can be, and stay at the cutting edge of technology, science, and business…

Subscribe now

(1) Heron Power Raises $140M to Build 40GW Solid State Transformer Factory

Everything that can go electric economically will, including the grid itself.

Heron Power, founded by former Tesla SVP of Powertrain and Energy Drew Baglino, raised $140 million in a Series B led by a16z American Dynamism to build solid state transformers. This is pure not boring catnip: power electronics from the guy in charge of a lot of the Tesla power electronics stuff we wrote about in The Electric Slide, a16z putting its new funds to work, and fixing the grid by attacking one of its key constraints.

The constraint is that transformers, which take electricity at one voltage and converts it to a different voltage (step up or step down) using coiled wires, are outdated (the fundamental design hasn't really changed since the late 1800s and is based on a one-way grid), passive (electricity goes in, gets stepped up or down, and comes out), massive (built with 10 tons of grain-oriented electrical steel and copper submerged in oil), and worst, in desperately short supply. Lead times stretch up to 24 months, U.S. manufacturing meets less than 20% of demand, manufacturing has been moving to China, prices have spiked 60-80% since 2020, and demand is projected to double in the next decade.

So Heron is building Heron Link, a modular solid-state transformer that uses wide-bandgap semiconductors instead of scarce electrical steel and copper. It’s starting by selling to solar and battery farm operators and data centers, who represent huge and growing demand and for whom the solid state transformer eliminates the need for inverters. Traditional transformers can’t handle direct current, so an inverter turns DC from a solar panel or battery into AC. Heron Link can do that, so customers can skip the inverter. HL is software-defined, meaning it can use software to regulate voltage and frequency to actively manage grid stability. And it’s modular, meaning that if one of the device’s tens of power conversion modules fails, it can be swapped out in ten minutes. They also contain lithium-ion batteries that can discharge quickly to provide 30 seconds of power to smooth the transition to backup power sources.

The funding will go to building a factory that can produce 40GW worth of Heron Links per year. At 5MW per Link, that’s 8,000 transformers.

Heron isn’t the only company making solid state transformers. I hope they all do great. But I do want to call out one competitor: Raleigh-based DG Matrix, whose co-founder and CTO is Subhashish Bhattacharya, a professor at NC State University who is a long-time collaborator of… B. Jayant Baliga, the inventor of the IGBT, The GATEway of India, the hero of the Power Electronics section of The Electric Slide!

Anyway, if we have a lot of smart people using wide bandgap semiconductors to fix the grid, the future of American energy is going to be SiC.

(2) DoW Transports a Valar Atomics Nuclear Reactor to Utah on a C-17

Isaiah Taylor and Valar Atomics

The title pretty much says it all here: the Department of War transported a Valar Atomics Ward nuclear reactor from California to Utah on a C-17. It’s another sign that the government is serious about turning on reactors by the United States 250th birthday on July 4, 2026. Not much more for me to add here other than to watch the video and the others in Valar’s Operation Windlord series.

In other nuclear news, Lockheed Martin is investing in Radiant Nuclear, which we wrote about in 2024, as it prepares to turn on its first reactor at the DOME this summer. Defense is starting to play offense on nuclear.

(3) Single vaccine could protect against all coughs, colds and flus, researchers say

James Gallagher for the BBC h/t Simon Taylor for the find

Researchers at Stanford have built a vaccine that protects against… everything. Like, everything they’ve tested, it works against. The nasal spray works by putting immune cells in the lungs on permanent “amber alert,” ready to fight whatever shows up. In animal tests, it reduced viral breakthrough by 100-to-1,000-fold and worked against flu, Covid, common cold viruses, two species of bacteria, and even house dust mite allergens. Prof Bali Pulendran calls it “a radical departure from the principle by which all vaccines have worked” since Edward Jenner figured out the originals in the 1790s.

It’s early (animal studies, not human trials) but the team is planning deliberate infection studies in people next. The big questions are whether the effect translates to human lungs, how long the protection lasts (about three months in mice), and whether keeping the immune system dialed up causes friendly fire. The researchers don’t think it should be permanent; they envision a seasonal spray at the start of winter or a rapid-deployment tool at the start of a pandemic to buy time while a targeted vaccine is developed.

That pandemic use case alone is worth getting excited about. One of the hardest lessons of early 2020 was how long it took to develop, test, and distribute a vaccine while people died waiting. A universal nasal spray sitting on the shelf, ready to deploy on day one, could change the math on the next pandemic entirely.

And if the seasonal version works, imagine a world where “cold and flu season” just no longer exists. As a parent of little kids, please for the love of God work in humans.

(4) DeepMind Veteran David Silver Raises $1B for Ineffable Intelligence

George Hammond for Financial Times

When interviewed Richard Sutton last year, the father of Recursive Learning (RL) and author of The Bitter Lesson said, basically, that LLMs are a dead end and that the real Bitter Lesson-pilled approach would be to just let AI learn from experience. That resonated with me. Even with all of the recent coding and agent advances, these things are still missing something ineffably intelligent to me.

David Silver wants to fix that. Silver, who joined DeepMind in 2010, was one of the team’s star researchers. He worked on AlphaGo, AlphaStar, and the Gemini family of models. And last year, with Sutton, he wrote a paper titled Welcome to the Era of Experience.

Silver and Sutton argue that we’re moving from an era where AI mainly learns by imitating static human data (like text on the internet) to an “Era of Experience,” where the most powerful systems will learn predominantly from their own ongoing interaction with environments. Agents will improve by generating vast streams of experiential data, acting over long time horizons, grounded in real environments and rewards, which will unlock truly superhuman capabilities beyond the limits of human-written data. Which makes sense intuitively!

Sequoia is leading a $1 billion investment at a $4 billion valuation. I’m excited for this one. And it couldn’t come at a better time because…

(5) The US Government is Going to Release the ET, UAP, and UFO Files

Maybe the greatest proof that we’re in some other beings’ simulation is the fact that just as we are about to figure out how to create intelligence ourselves, we finally begin to get the truth about other intelligent life in the universe.

This is just one Truth post, and the government has a spotty (redacted) record releasing the important details in important files recently, but it seems that disclosure might finally be upon us.

It’s too early to know or speculate when we’ll get more information, what’s in the files, and how ontologically shocking it will be; it could be anything from “we’ve been holding a few craft that we don’t quite understand” to “we’ve reversed engineered the craft and understand antigravity” to “humanity is just an experiment run by far more intelligent beings” to … anything, really.

But it’s funny and kind of beautiful that just as so many people are worrying about humanity’s place in a world with AI, we may get proof that we’re part of something much larger, weirder, and more wondrous than we could ever imagine.

EXTRA DOSE: Scientific Breakthroughs, Two Podcasts, and Research Revival Fund

Subscribe to learn about 8 scientific breakthroughs this week, including gene drive-like systems to fight antibacterial resistance, potential early cancer diagnosis method from Arc Institute, DMT for major depressive disorder, and using laser writing in glass for long-term storage. Let us know what you think!

Read more

Subscribe now

Hi friends 👋,

Happy Wednesday!

Émile Borel once said that given enough time, a bunch of monkeys banging on typewriters would come up with Shakespeare. And yet, despite the innumerable X Articles on software moats in the face of AI, I haven’t read a single one that’s satisfying.

I think it’s because thinking about software moats, how a software company might protect itself from abundant code, is the wrong frame altogether, and that the more interesting and relevant frame is which companies, SaaS, hardware, or otherwise, stand to benefit the most from newly abundant inputs.

Those companies, not vibe coders, are the ones that point solutions should be worried about. They will win enormous market shares and fortunes. They will come to dominate large industries by using new technology to compete, capturing the High Ground, and expanding further outward than companies with lesser technological tools could have ever dreamed of. They will be the Standard Oils of this era.

This essay is about those companies.

Let’s get to it.

Today’s Not Boring is brought to you by… Silicon Valley Bank

In 2025, crypto returned to the financial mainstream. It is, as they say, so back.

What’s ahead for 2026? I’m glad you asked. Silicon Valley Bank is out with their annual crypto outlook, featuring proprietary insights and data from over 500+ crypto clients. We love a bank that banks crypto.

Silicon Valley Bank makes five predictions for the year ahead, including:

1. Institutional capital goes vertical with increased VC investment and corporate adoption.

2. M&A posts another banner year after the highest-ever deal count in 2025.

3. Real-World Asset (RWA) tokenization goes mainstream on prediction market strength.

Last year, Silicon Valley Bank predicted that stablecoins would be the big breakout use case. That was correct. They think that will continue this year, too. Read their take on what comes next, free:

Get SVB's 2026 Crypto Predictions Free

Power in the Age of Intelligence

One of the more head-scratching anomalies in the market is the valuation gap between Stripe and Adyen. The two payments companies handle similar amounts of Total Payment Volume. Stripe is growing faster. Adyen reports exceptional margins and cash conversion. Stripe is reportedly doing a tender offer at a $140 billion valuation in the private markets. Adyen is valued at $34 billion in the public markets. There are a number of theories for why this is the case, most of which boil down to: VCs are idiots, as they’ll find out if Stripe ever goes public.

Ramp versus Brex is another example of the same idea. Ramp was most recently valued at $32 billion in the private markets. Brex, which had been valued at $12 billion in the private markets, sold to Capital One for $5.15 billion. Ramp is doing more revenue and growing faster, but not 6x more revenue or 6x faster. Once again, the actual market disagrees with the VCs.

Or does it?

I have a different theory, one that neatly fits those two cases, the SaaSpocalypse, SpaceX’s $1.25 trillion valuation, and even the evolving structure of venture capital itself: winner takes more.

The history of business is basically the history of increasing concentration of value, accelerated in spurts by technological change. Centuries ago, firms operated within cart-hauling distance of their customers, creating a system of local monopolies. A brewer in 1800 served a single town, if that. Canning, railroads, telegraphs, mass production, electrification, containerization, planes, and the internet, among other technologies, expanded companies’ available market, and winners captured a greater and greater share of value.

Economic data backs this up.

In a 2020 paper, Jan De Loecker and Jan Eeckhout found that aggregate markups rose from 21% above marginal cost to 61% between 1980 and the late 2010s, and this increase was driven almost entirely by the upper tail of the distribution; median firm markups barely changed, while the 90th-percentile markups surged.

In The Fall of the Labor Share and the Rise of Superstar Firms, Autor et al. find that industry sales concentration trends up over time across measures, and it rises more in sales than in employment, what Brynjolfsson et al. call “scale without mass.” In their account, this shift reflects reallocation toward superstar firms with high markups and profits and low labor shares.

A 2023 paper by Spencer Y. Kwon, Yueran Ma, and Kaspar Zimmermann at UChicago, 100 Years of Rising Corporate Concentration, uses IRS Statistics of Income to show that the top 1% of U.S. corporations by assets accounted for about 72% of total corporate assets in the 1930s and about 97% in the 2010s. The top 0.1% of U.S. corporations by assets increased its share of total corporate assets from 47% to 88% over the same period. Power laws within power laws.

Today, the Magnificent 7 accounts for one-third of the market cap of the S&P 500, and Apollo showed that those seven companies are driving the vast majority of equity returns.

Venture capital is responding to this information as you would expect. Per Pitchbook, as of August last year, 41% of all VC dollars deployed in the US in 2025 went to just ten companies. Per Axios, more recent Pitchbook data shows that “The estimated aggregate valuation of unicorns hasn’t actually changed too much — $4.4 trillion vs. $4.7 trillion at the end of 2025 — because the top 10 companies account for around 52% of value (up from only 18.5% in 2022 and the highest such figure in a decade).”

Venture capital firms themselves are concentrating. As I wrote in a16z: The Power Brokers, “a16z accounted for over 18% of all US VC funds raised in 2025.” Just yesterday, Thrive announced that it raised $10 billion: $1 billion for early-stage investments, and $9 billion for late-stage investments, which is the kind of split you put in place if you believe your winners are going to keep winning.

The theory for investing in very large funds like a16z, Founders Fund, Thrive, General Catalyst, and Greenoaks is that they are best positioned to win large allocations in the handful of companies that matter, and that those companies will capture most of the value in this vintage. Notably, all five of the firms I just mentioned are investors in Stripe, and three of the five (Founders Fund, Thrive, and General Catalyst) are investors in Ramp.

All of that data suggests increasing concentration. Through this lens, the SaaSpocalypse (the violent sell-off in software stocks) is less about software writ large dying, and more about point solution software finally facing economic gravity. They are no longer getting a free pass simply for having a good business model.

The past few decades of software exceptionalism have been an exception based on a business model so sweet and capabilities so universally useful that the rules of strategy, while not wholly unimportant, were less consequential than normal. A venture capitalist could look at a standard set of SaaS metrics (ARR, growth rate, net retention rate, gross margin, LTV:CAC, Rule of 40, etc…) and underwrite whatever new business they encountered to them. This is why you hear things like “Here’s how much ARR you need to raise a Series A.” The companies are basically all the different flavors of the same thing.

There are, of course, idiosyncrasies between selling software to professional services firms and selling software to energy companies, for example, but the basic model is the same. Invest upfront to hire engineers, write software to make someone’s job easier, and sell the software to as many customers as possible at high margins. Different industries may need software to do different things, have different buyers, be larger or smaller, be more or less willing to pay, and be more or less expensive to acquire. A bigger, less crowded market with a strong need and a high willingness to pay is better than the opposite. But figuring that out is relatively straightforward. Since software operates at the edge of value in most industries, and does not attempt to strike at its core or compete with it, it doesn’t require thorough competitive analysis.

Since the SaaSpocalypse, people have gotten AI to write tens of thousands of words on which types of software companies have moats given AI and posted the resulting essays to X. They’ve gotten a little more specific than “SaaS is good.” Data is a moat, or a particular type of data at least. Or it isn’t, maybe, because The Agent Will Eat Your System of Record. Certainly, dealing with regulatory hair earns you a moat. No?

Most of the takes I’ve seen miss what matters.

On paper, Stripe and Adyen have basically the same moats, as do Ramp and Brex. I love a good hardware moat more than the next guy, as I’ve been writing since The Good Thing About Hard Things in July 2022, before ChatGPT or Claude Code but when it was clear that good software alone would offer no moat. I was too unspecific in that piece, too. Some hardware businesses will get very large, and others will fail. Hardware itself isn’t a moat. Good luck making LFP cells in America.

No, what matters is becoming the leader in your industry in a way that is incredibly specific to that industry and in such a way that your business benefits from, instead of being threatened by, abundant improvements in general purpose technologies like AI and batteries.

What matters now is the same stuff that has always mattered but that software forgave for a while: own the scarce, defensible asset in an industry and use it as the High Ground from which to dominate. Ricardo said this.

If you’re a startup, and you don’t already own the scarce asset, then you need to identify the constraint holding the industry back, focus everything on breaking it, and expand from there.

History’s most influential military strategist, Carl von Clausewitz, said this. He called it Schwerpunkt, the center of gravity. “The first task, then, in planning for a war is to identify the enemy’s center of gravity, and if possible trace it back to a single one,” he wrote in On War. “The second task is to ensure that the forces to be used against that point are concentrated for a main offensive.”

For our purposes, the Schwerpunkt is the constraint you attack. The High Ground is the scarce and valuable position you win by breaking it. Moats are what keep others from taking it.

Even while forces must be concentrated against the Schwerpunkt, our attackers must plan for victory before it is won. The company that breaks the constraint needs to build the complementary assets (the distribution, the manufacturing, the customer relationships) to capture the value from its own innovation. Otherwise, its competitors will.

David Teece argued this in 1986, in Profiting from Technological Innovation. The paper, he wrote, “Demonstrates that when imitation is easy, markets don’t work well, and the profits from innovation may accrue to the owners of certain complementary assets, rather than to the developers of the intellectual property. This speaks to the need, in certain cases, for the innovating firm to establish a prior position in these complementary assets.” Which is the point I am making: innovation alone, software or hardware, isn’t enough.

Figuring out which companies might capture the Schwerpunkt and use it as a High Ground from which to expand is an entirely different kind of underwriting, impossible to do in a spreadsheet alone, even with Claude in Excel.

Companies that don’t own the High Ground face existential risk from technological progress. If you’re just selling point solution software, then software abundance is a threat. Hardware isn’t necessarily the moat people think it is, either, even if it’s less susceptible to AI, because AI isn’t the only technology improving rapidly. “Hardware is a moat” is the same kind of lazy thinking that “SaaS is the greatest business of all time” is. If you’re selling better mouse traps, you’re at risk every time someone builds a slightly better mouse trap.

Similarly, incumbents that currently own the High Ground but can’t wield modern technology face existential risk from those who can. This is why there is such a large opportunity for startups today. New technologies mean old constraints are finally attackable, and it’s likely to be newer companies doing the attacking.

Companies that do own the High Ground, on the other hand, and are tech-native, benefit from technological progress, just as land owners captured the gains from more farming labor and better farming tools, but more pronounced, because these modern landowners will corner the best talent, the most capital, and the richest veins of distribution. A glib way to put it is that a Ramp engineer with an AI will build something better than a CFO with an AI, no matter how good the AI gets. It’s not the vibe coders you should be worried about.

Newly abundant resources can have opposite effects on your business depending on your position, and it is likely to be the company with the High Ground wielding those resources that dooms the companies in weaker positions. Ask Slack how it felt to compete with Microsoft Teams; companies like Microsoft can now build a lot more “Teams.”

The game on the field is all about understanding who can own the High Ground in a given industry.

The moats are the same as they’ve always been. Study 7 Powers. When you’re starting out, you need to understand what your moats might be, but in order for moats to matter, you need to have something worth protecting. You need to own the High Ground.

If we really are living through the most consequential technology revolution in history, why are you spending so much time hand-wringing about protecting small, old castles when you could be thinking about how to build history’s most magnificent businesses?

The abundant inputs keep getting cheaper. The scarce asset keeps getting more valuable. The companies that own the latter and leverage the former will become larger than ever before.

These businesses themselves are scarce assets, valued on their strategic importance and industry size, because the opportunity is no longer to sell software into industries in order to marginally improve them, but to win those industries and capture their economics.

If your aim is to build or invest in these companies, old heuristics will do you no good. You need a brain of your own, some sweat on your brow, and some good ol’ fashioned strategy frameworks to help you reason about the opportunity at hand.

This essay is a guide to thinking through where power might concentrate, for those willing to think. If winner takes more, it’s about what it takes to build, or invest in, the companies that have a shot at winning large industries. And it’s about how to position yourself to gain strength from technological progress instead of running from it while throwing weak “moats” in your wake.

It is about Power in the Age of Intelligence.

A Tale of Two Industrial Revolutions

Or, it’s about Power in any age of rapid technological change, really.

While the advances we are experiencing today feel unprecedented, my thesis has been that we are going through a modern version of the Industrial Revolution.

Then, machines did what only human muscles could previously. Now, machines are doing what only human brains could previously, in new bodies built to house those brains. This is The Techno-Industrial Revolution.

So it is useful to study Rockefeller, Carnegie, Swift, and Ford. None created an industry from scratch. They all fit this pattern: identify the Schwerpunkt in an existing industry, break it, seize High Ground, integrate outward, dominate.

Standard Oil

When John D. Rockefeller met the oil industry, it was young, valuable, and incredibly volatile. Oil itself was abundant. There were a lot of refineries – roughly 30 in his hometown of Cleveland alone when he got to work – but their quality was inconsistent, and their processes were inefficient. Per Austin Vernon, “Refining methods were so inefficient in the mid-1860s that a barrel of oil (42 gallons) sold for almost the same price as a gallon of refined kerosene. Today, the price ratio of refined products to crude oil is ~1.25x instead of 42x.”

There was one big constraint to the profitable growth of the oil industry – the volatility – which could only be broken through scale and control. To get to scale and control, Rockefeller needed to drive down costs to capture the market. Refining was the place to get scale, given its inefficiency and the fact that, per Vernon, “A typical rule of thumb in chemical engineering is that capital costs increase sublinearly with capacity, usually by (capacity ratio)^0.6. A plant with double the output is only 50% more expensive to build, and operating costs tend to follow similar trends.”

So in partnership with chemist Samuel Andrews, one of the first people to distill kerosene from oil, Rockefeller continued to improve the kerosene yield. At the same time, he aggressively grew revenue and lowered costs by eating the whole cow, so to speak. He sold the non-kerosene byproducts others threw out (paraffin wax, naphtha, and gasoline) and used some of the fuel oil to power his own plants. He also integrated into barrels (by buying an oak tree forest and a barrel-making shop).

As it lowered costs and delivered a more consistent product, Rockefeller’s refinery (the predecessor to Standard Oil) grew, and as it grew, it lowered costs. Vernon again:

Standard Oil and its predecessor firms increased production ~20x between 1865 and the end of 1872, meaning their costs could have fallen more than 85%. At that point, they were the largest refiner in the world with a double-digit share of capacity, and it was their game to lose. If we understand this short period, then we know how the company eventually won.

The company that would become Standard Oil won the High Ground by breaking the constraint. Then it integrated outward, horizontally and vertically.

By 1870, Standard Oil was a joint stock company capitalized with $1 million that owned 10% of the oil trade in the United States. Rockefeller got busy acquiring struggling refineries or putting them out of business, increasing scale and efficiency in the process. Rockefeller also did favorable deals with the railroads, which Vernon argues actually had less to do with Standard Oil’s success than did its growing scale and efficiency. He kept growing, acquiring refiners in Pennsylvania, New York, and New England. He vertically integrated into pipelines (which replaced railroads), into distribution, into retail (ExxonMobil and Chevron are Standard successors), and into production itself.

By the late 1880s, Standard Oil controlled 90% of American refining, a share it held until it was broken up in 1911, when its $1.1 billion market cap represented 6.6% of the entire US stock market. To hear Vernon explain it, the outcome was a fait accompli by the time he’d attacked the Schwerpunkt and gained the High Ground in the 1860s.

Carnegie Steel

Andrew Carnegie’s story is so similar it’s almost suspicious. Like Rockefeller, Carnegie didn’t invent his product (steel); Bessemer did. Like Rockefeller, Carnegie realized the constraint to steel’s growth was inefficiency and inconsistency. Like Rockefeller, Carnegie hired a chemist (in his case, to measure what was happening inside the furnaces) and obsessed over cost, which he knew was the only thing he could control:

Show me your cost sheets. It is more interesting to know how well and how cheaply you have done this thing than how much money you have made, because the one is a temporary result, due possibly to special conditions of trade, but the other means a permanency that will go on with the works as long as they last.

His chemical knowledge allowed Carnegie to run his furnaces hotter and longer than anyone else, producing more steel at lower cost. His cost obsession lowered costs further. Low cost, high quality steel was the High Ground.

And from there, he integrated outward. Backward into coke (Frick) and iron ore (the Mesabi Range), into railroads to transport raw materials, and forward into finished products. He certainly didn’t sell his services and know-how to incumbents; he used them to destroy competitors on price until US Steel bought him out for $480 million in 1901 (roughly $18 billion today) to create the first billion-dollar corporation in history.

“Congratulations, Mr. Carnegie,” JP Morgan told him upon closing, “you are now the richest man in the world.”

Swift Meats

Gustavus Swift, like Rockefeller, would also “eat the whole cow.” He just did it later in his arc, and more literally.

The constraint was this: only about 60% of a live animal’s mass is edible, and meat goes bad. Which meant that, prior to the 1870s, the meat industry shipped 1,000 pound live cattle by rail from wherever they were raised to wherever they were going to be eaten. They paid by the pound (an extra 40%), had to feed animals to keep them alive and healthy, and lost some to death in transit anyway.

So Swift, building on early experiments by Detroiter George Hammond, hired an engineer to design him a refrigerated railcar. Then, he could slaughter the beef in Chicago and ship the cuts to their final table much more efficiently. Railroads, not wanting to lose livestock shipping cash cow, refused to pull his cars, so Swift leased his own and partnered with smaller lines to move them. Then, he built icing stations along the routes, and replenished them with ice he contracted directly with ice harvesters in Wisconsin and other cold midwestern states. By necessity, he built the whole cold chain from scratch.

This combination of centralized slaughter in Chicago and cold chain to the coast was his High Ground. He was forced into vertical integration because none of the pieces made sense on their own, but once he had it, he used it to drive down costs.

Like Rockefeller, Swift was appalled by waste, and because he controlled his own slaughterhouses, he could do something about it: he turned cow byproducts into soap, glue, fertilizer, sundries, even medical products, which allowed him to increase revenue and lower prices. He also maximized his refrigerated cars by stacking butter, eggs, and cheese beneath the swinging carcasses of dressed beef heading East.

By 1884, after only six years in operation as a slaughterer, Swift had become the second largest meatpacking firm in the US. By 1900, the meatpacking industry, unconstrained, had grown to become the second largest in the country to iron and steel.

Ford

It was a visit to a Chicago slaughterhouse that inspired Henry Ford’s assembly line. “Along about April 1, 1913, we first tried the experiment of an assembly line,” Ford writes in My Life and Work. “We tried it on assembling the flywheel magneto. I believe that this was the first moving line ever installed. The idea came in a general way from the overhead trolley that the Chicago packers use in dressing beef.”

Ford didn’t invent the automobile. By 1908, there were hundreds of American car companies selling expensive, hand-built machines to the wealthy. A typical car cost $2,000-$3,000, or roughly two and a half years’ wages for an average worker. Manufacturing costs were the constraints to the nascent automobile industry’s growth, so manufacturing costs were Ford’s Schwerpunkt.

Ford broke the constraint with the moving assembly line. Before it, a single worker assembled a complete flywheel magneto in about 20 minutes. Ford split the work across 29 operations, cutting the time to 13 minutes. Then he raised the line eight inches and cut it to seven minutes. Then he adjusted the speed of the line and cut it to five. The same progression played out across the whole car: total assembly time fell from over 12 hours per chassis to 93 minutes.

That manufacturing capability was the High Ground. The Model T launched in 1908 at $850, already half the price of the competition. As the assembly line improved, Ford kept cutting: $550 by 1913, $360 by 1916, below $300 by 1924. What had been 18 months’ wages for an average worker became four months’.

From that High Ground, Ford integrated ferociously. Rubber plantations in Brazil. Iron mines and timberland in Michigan. A glass plant, a railroad, a steel mill, even soybean farms for plastic components. All of it flowed into the Rouge River complex, where raw materials entered one end and half of the finished cars on the world’s roads rolled out the other.

The result was that Ford’s sales went from 12,000 in 1909 to half a million in 1916 to over two million in 1923. At its peak, more than half the cars in the world were Fords.

Across the Industrial Revolution’s most successful entrepreneurs, there was a clear pattern that looks almost nothing like how you’d think about scaling a SaaS business: identify the Schwerpunkt in an existing industry, break it, seize High Ground, integrate outward, dominate.

Pause for a second. Think about how people are telling you to analyze businesses today. Would those AI-generated moat lists, or the equivalent for their time, have given you any advantage whatsoever in identifying Rockefeller, Carnegie, Swift, or Ford, let alone becoming one of them? It is never that easy, and it always takes work.

I want you to feel those examples, because what’s old is new again. The biggest companies in the world today are executing against the same framework, in ways that are specific to their industry.

SpaceX Goes Vertical

The funny thing about today’s biggest software companies is just how much they spend on hardware. This year, the world’s four largest companies that started as software companies plan to spend an estimated $600-700 billion on data center buildouts, equivalent to roughly 2% of US GDP, a level of infrastructure buildout comparable to laying America’s railroads in the 1850s.

Amazon, an online bookseller, will spend $200 billion. Google, the search engine giant, will spend $175-185 billion. Meta, the social network for college students, will spend $115-135 billion. And Microsoft, which makes operating systems and office applications, will spend $100-150 billion.

Except, of course, that’s not what those businesses are. They are technology conglomerates that used the early internet to break the Schwerpunkt in their respective industries, gain their respective High Grounds, and integrate outward so far that they’re all running into each other at this new frontier. And despite their best efforts and hundreds of billions of dollars spent on terrestrial data centers, Elon Musk still thinks we’re going to need to put them in space.

SpaceX

Before SpaceX, the constraint in the space industry was cost to orbit. SpaceX broke the constraint with reusable rockets, drove costs down an order of magnitude, and quite literally gained the High Ground. From there, it integrated outward into Starlink communications satellites, which it can launch more cheaply than competitors because it owns the rockets and which fund the development of even bigger Starship rockets, which bring the cost per kg to launch things into orbit down even further. SpaceX used vertical integration the same way Rockefeller did: it is simultaneously its own largest customer and its own cheapest supplier. Casey Handmer’s The SpaceX Starship is a very big deal is an excellent read on the topic.

In 2023, Elon Musk founded xAI to build maximally truth-seeking AI. He then merged it with X (née Twitter). xAI got a late start, and it doesn’t have the best models yet, but what it is best in the world at is building data centers very fast. So the world took note when Elon said that we’d never be able to build enough data centers on earth to meet demand for AI, and that we will need to start building them in space.

So on February 2nd, 2026, SpaceX announced that “SpaceX has acquired xAI to form the most ambitious, vertically-integrated innovation engine on (and off) Earth, with AI, rockets, space-based internet, direct-to-mobile device communications and the world’s foremost real-time information and free speech platform.” According to Musk, SpaceX will get the data centers it needs in space via ~10,000 Starship launches per year, or roughly one per hour, every hour. Simultaneously, it will also build a self-growing Moon city, from which it plans to build a mass driver in order to make a terawatt per year of more worth of AI satellites, far more energy than Rockefeller could have conceived of, en route to eventually colonizing Mars and fulfilling SpaceX’s mission to “extend consciousness and life as we know it to the stars.”

It remains to be seen whether the High Ground will also give SpaceX a decisive advantage in the AI race, but it certainly demonstrates that the stakes have grown since the Industrial Revolution, even as the strategy has remained the same.

But no matter how that plays out, SpaceX (and Google, Microsoft, Amazon, Meta, Apple, Tesla, NVIDIA, OpenAI, and Anthropic) aren’t going to eat everything, or else I wouldn’t be investing in startups.

The Hunt for the High Ground

Boulton and Watt did not capture the entire value of the Industrial Revolution they steam powered, although they did vertically integrate, Boulton into the Soho Manufactory, the steam engine-based Gigafactory of its day, and Watt, via his son, into steamships. Nor did Rockefeller eat everything in the internal combustion era of the Revolution despite owning the oil on which it all ran.

In addition to Rockefeller (oil), Boulton and Watt (steam engine), Carnegie (steel), Swift (meatpacking), and Ford (automobiles), the Industrial Revolution gushed multi-generational wealth for the Vanderbilts (railroads), Morgans (finance), Sears and Roebucks (retail), Havemayers (sugar), McCormicks (agricultural equipment, sadly unrelated), Westinghouses (power), Otises (elevators), Pullmans (luxury rail cars), Bells and Vails (telecommunications), Pulitzers and Hearsts (publishing), Eastmans (photography), Kelloggs (processed food), Pillsburys (milling), Singers (sewing machines), Nobles (Dynamite), DuPonts (chemicals), and Dukes (tobacco). This list is incomplete.

What’s notable is the diversity of industries that produced these fortunes. Machines made “labor” more abundant, and the companies that seized upon the technological innovation to break the Schwerpunkt in their specific industry, gain the High Ground, and expand were all wildly successful. Far from simply defending against mechanization, they seized the complementary assets to which value flowed as key inputs became abundant.

There are clear differences between AI, developed by huge labs and distributed at the speed of bits, and Industrial Era machine-filled factories, but I expect the Techno-Industrial Era to play out similarly. Each industry has unique constraints and resulting High Grounds, very few of which can be cracked and captured with digital intelligence alone.

The diversity that creates unique opportunities in each industry, however, makes underwriting those opportunities a different and more difficult beast than underwriting SaaS companies, which are more homogenous. There is no list, no spreadsheet, no agreed-upon metrics that will tell you which will become today’s Standard Oils. There is only the evaluation of constraints and hunt for High Grounds.

Instead of a list, then, let me give you my favorite example: Base Power Company.

Base Power Company

Base Power Company doesn’t just make batteries. It buys cells (the commoditized piece of the value chain), manufactures battery packs, installs them on homes (starting in Texas), writes software to coordinate them, trades in the power market, and partners with utilities to help balance the grid. Base is built on the type of logic companies (and their investors) will need to exercise if they want to compete in the modern era, and it goes something like this.

We want to fix power. What’s the bottleneck? The grid. Companies are competing to build power generation and the electric machines that consume that power, and the better they do, the more strain there will be on the grid. The grid is the chokepoint. So how do you fix the grid? Laying new transmission and distribution is slow and expensive, and the grid we have is already structurally underutilized because it’s built to serve peak demand, so to smooth it out, you need batteries. Where should you put the batteries? Centralized battery farms are helpful, but they need to wait in interconnect queues, which makes them slower to turn on, and those batteries still need to distribute power to end users when demand peaks, which means they don’t fully solve the bottleneck. So you need to put the batteries right next to demand. Fill them up when the grid has capacity, and use them to smooth demand when demand is high. And if you want to put batteries next to demand (homes, to start), where is the best place in the country to do that? Texas, which operates its own deregulated grid, ERCOT, is volatile (which means potential for higher trading profits and greater need on the part of customers and utilities), and is regulatorily friendly. So you start by putting batteries on the homes of early adopters within Texas. Those slots are scarce - it would take a lot for a customer to rip and replace their batteries, and no one is installing two companies’ batteries. Then, connect them with software, improve the grid and each customer’s experience with more batteries on the network, and use the richest source of demand available in the country to begin to scale. Bring manufacturing in-house, continue to improve the batteries, decrease their costs, get more efficient at installing them, connect more of them, sign more early adopter utilities, get more scale. At which point, it’s hard to imagine a viable way to beat Base at its own game. Then, expand. Integrate upstream into grid hardware and generation and downstream into electronic devices to sell into customers with whom you’ve built trust. Expand geographically, leveraging scale, experience, and software to offer a better product than a potential competitor attempting to grab a foothold by starting in the next-best market. Keep expanding. Dominate. Expand some more.

There are a couple things I want you to take away from that paragraph.

First, it is a very long paragraph. This is not simple or easy. I think investors bemoaning the Death of SaaS are in part sad that the era of underwriting software businesses on known, straightforward metrics is over. Underwriting the biggest companies of this generation will be a much more bespoke process. The time has come to move from simple analysis to strategy. It is not a coincidence that my first Deep Dive on Base was structured as a walk through the evolution of the strategy memos that Zach and Justin wrote before touching a single atom.

Second, as technology improves – from AI to the Electric Stack – the vast majority of the returns will accrue to the companies that figure out the right place to attack and execute violently against their conviction. A simple way to think about this is that better software is more valuable to Base than it is to a smaller competitor, to a battery farm operator, or to a power generation company, as is better hardware. Better robots for manufacturing and logistics would make Base faster and more profitable, and making the game more CapEx intensive would give it an advantage over would-be competitors.

The lesson from Base is not that hardware is a moat, or that you should put your product next to Texans’ homes.

It’s that you need to deeply understand the problem you’re trying to solve, the constraint that’s bottlenecking it, how you’re going to unblock it with technology (and why now?), and how you might expand to capture the market once you do. It applies differently in every industry.

For airlines, the constraint is the engine: today’s turbofan engines carry planes as fast and efficiently as they can. Everything bad in air travel is downstream of that. So Astro Mechanica is building a new engine that is faster and efficient at every speed. But certifying a commercial airline is a long and expensive process, so Astro plans to sell into Defense first, then build private supersonic planes (which are cheaper to certify and can be cost-competitive with first class tickets immediately), then build larger supersonic planes that are cost-competitive with commercial air travel, and use the advantage in speed and cost to build its own full-stack airline, from booking to flight.

For internet, the constraint is the architecture: incumbent telcos froze their architectures around early-2000s assumptions about what was expensive, locked themselves into passive optical networks and vendor dependence, and now spend billions every few years on upgrades that still deliver shared, degraded bandwidth with no redundancy. They do zero R&D. So Somos Internet is rebuilding the full stack from scratch: an Active Ethernet architecture borrowed from data centers, physically simple with complexity pushed to software, that delivers dedicated bandwidth to every home at a fraction of the CapEx. As it grows, Somos eats more of its supply chain: “It’s been this never-ending game of doing something janky, getting credibility, doing crazier stuff, getting more resources, getting smarter people so that we can fix the things that were messed up in the janky past iteration,” Forrest explained. “Then gaining credibility to get more resources to get cooler people to do crazier stuff. It’s like this self-sustaining fission process.” Somos is expanding geographically, into new markets, vertically, by making its own hardware and laying its own fiber, and horizontally, building hydro-powered data centers. From the position of delivering one of the few home utilities everyone pays for, better, faster, and cheaper than incumbents, it plans to expand what it offers the customers with whom it’s built trust and loyalty. Maybe one day, it will offer batteries and power. Maybe one day, it will use its growing cash position to enter the United States.

There are a lot of similarities between Base and Somos: both own a core home utility and deliver it better than incumbents, which earns them the right to expand. But there are differences, too. Base is starting in the very best market for its technology, because that’s where the need is greatest and the regulatory environment is friendliest. If Somos started somewhere like New York City, it would be caught up in red tape and slow, expensive telco lawsuits for years; so it’s starting in a high-need, regulatorily friendly environment and building up cash for bigger battles. And Astro’s approach is almost entirely different from both Base’s and Somos’, apart from using better technology, now feasible thanks to Curve Convergence, to break a constraint and capture the High Ground. For one thing, people go to planes, so Astro can’t capture their real estate in the same way that Base or Somos can.

There I go with the long paragraphs again. Fine. There is endless nuance to this.

I am talking my own book here, not because I think my portfolio companies are the only businesses that will succeed in the Age of Intelligence, but because I understand their strategies much more thoroughly. Very smart people will disagree with me on each industry’s Schwerpunkts and potential High Grounds. And even once you’ve done all this work on paper, so much comes down to execution. Will the team that identified the right strategy be the same one that can build against it to capture the opportunity? Only time will tell. That’s what makes this so much fun - it’s not obvious!

What is obvious, and I hope clear at this point, is that there is no one answer, no handy guide that will tell you how to win in the Age of Intelligence. Which means that there is also not one business model.

A Note on Business Models

While the “Death of SaaS” is overblown, what I hope this freak out does is to end the default investor assumption that every business should try to be a SaaS business.

A week before the sell-off, I met with three separate founders who told me that investors didn’t like their businesses because they weren’t SaaS. In two cases, the founders were building services businesses – traditionally a huge venture red flag. In all three, they believed the technology they were building was so good that they could use it to compete directly with incumbents instead of selling them software that made them marginally more productive.

During the sell-off, Flexport’s Ryan Petersen tweeted that everyone “smart” had told him to just build SaaS. The idea being that it would be easier to sell to freight forwarders instead of actually becoming a freight forwarder and competing.

Other founders quote tweeted him saying they’d been given, and ignored, the same advice, including Cover’s Alexis Rivas, whose company builds houses. I’m not even sure what selling software would look like here.

This is not because investors are dumb. Selling software has real advantages, and those advantages are legible more quickly. Two of those three founders I spoke to said that they had competitors who were selling software and generating a lot of revenue quickly, which is why investors thought they should be doing the same.

In the past, that was a logical discussion to have: should you try to sell software to generate lots of high gross margin revenue in the near-term in a way that’s legible to downstream capital so that you can continue to raise and hopefully give yourself time to develop moats, or should you try to compete directly, with better technology and a chance at better economics within whatever your industry’s business model is, even if those economics are worse than SaaS economics, in pursuit of the larger and ultimately more impactful shot at winning and reshaping your industry?

In most cases, that is no longer a debate. AI squeezes it from both sides. From one side, SaaS is a more competitive, less defensible business; there will be enough competitive noise that it’s harder to establish traditional moats like network effects and switching costs, and customers big and sophisticated enough to actually pay a lot and make use of your tool may opt to build something custom themselves. From the other, the technology is so powerful in the right hands that it should provide a stronger force with which to attack the Schwerpunkt than deterministic software could have. In other words, it is more likely than ever that a technology-native new entrant can defeat incumbents, assuming their technology actually addresses the industry’s key constraint.

What this means is that investors need to get comfortable with a wider range of business models to accommodate whichever is the right one for the industry in which a company operates. This does not mean that they should treat all business models equally now. Instead, they need to stop blind pattern matching altogether.

Services businesses might still be terrible for most businesses, but exactly the right model for some. Stripe clearly shouldn’t be a services business; maybe an AI-native law firm should. Selling hardware to incumbents might be a bad business model, not simply because hardware is hard, but because buyers have all the power in a particular industry, or because existing suppliers have locked buyers into whole sticky ecosystems. What might be better is to use that better hardware as the High Ground from which to integrate and compete.

A question I would like to see more investors asking instead of “Why not sell SaaS” is:

If your technology is so good, why aren’t you using it to compete?

Some companies find out that selling software to incumbents is the wrong model only through trial-and-error. My favorite example here is Earth AI.

Earth AI developed AI models to identify drilling targets for mining explorers back before AI was a thing, and sold it to explorers for a very good price at high margins. The challenge was: they never heard back from their customers. Many went bust - exploration is a notoriously binary business - which meant they stopped paying; retention was hard. Many others just had no incentive to report back, which meant that Earth AI wasn’t learning which of its targets were good and bad, which meant that it couldn’t improve its models. So it bought its own rig and went to customer sites to go find out for itself, and then it realized that it could build better rigs, and combine them with better models, and just compete directly. As I wrote in my Deep Dive:

The same thing that makes exploration customers bad customers – slowness, unwillingness to adopt technology – makes them very attractive competitors, if your tech actually works as well as you say it does. If you’re willing to vertically integrate – to do exploration, and drilling, and maybe even extraction – you might be able to build the most efficient explorer out there.

To be clear, Earth AI’s current business model is much more confusing than selling software. It has to invest in rigs up front, stake deposits, put teams on site to prove them out, and keep proving feasibility until a downstream miner wants to buy a stake in the deposit or buy the whole deposit outright and pay Earth AI a royalty, at which point, it becomes one of the most beautiful business models there is. Mining royalty & streaming companies have some of the highest market caps per employee in the world. Franco Nevada is worth $48 billion with just 41 employees, good for $1.2 billion per employee! Earth AI has the potential to build up a similar portfolio at a much lower cost basis because it is willing to dig.

The point is, maybe you drill mineral deposits in Australia to build a portfolio of mines, maybe you buy cells, manufacture battery packs, install them on homes, and make money by becoming a Retail Electric Provider, trading power, and selling ancillary services, maybe you hire expensive humans, make them much more efficient, and sell their time, maybe you even sell software!

Whatever you need to do to break the constraint, gain the High Ground, and win your industry is what dictates the business model you should pursue.

You Can Even Sell Software, As Long as You Win

Sometimes, the Gods smile on an essay. As I was writing this, Stripe co-founder John Collison released the latest episode of his podcast, Cheeky Pint. His guest: Eric Glyman, the CEO of Ramp.

Responding to John’s first question, Eric described Ramp’s evolution in terms that should now sound familiar. A few years ago, Ramp’s gross profit was over 90% card interchange. Today, the non-card businesses, including bill payments, treasury, procurement, travel, and software, will comprise the majority of Ramp’s contribution profit.

Ramp used a card, software, and counter-positioning to attack what it viewed as the Schwerpunkt in corporate spend (the fact that everyone was selling money, and no one was selling time) and win the transaction layer, the High Ground from which it is now expanding to eat every point solution a finance team touches.

Throughout the conversation, he lays out from his inside view exactly how and why this is happening. Everything flows from earning the High Ground. Ramp has data that no one else has and that no new entrant can accumulate more quickly. As it adds more intelligence, it gets more data. It’s built the things that are too expensive to replicate with more tokens – “I think the fitness function for companies becomes can you actually do things in such a way where even if you could spend tokens on it, it would take more tokens to create the thing or do that work than the system that you’ve built to drive that outcome.” – and is happily spending tokens to build everything else. And as it adds more features, it grows. The company now “powers more than 2% of all corporate and small business card transactions in the United States.” The larger its share, the more it learns, the more it makes, and the more tokens it can throw at eating adjacent opportunities, which keep feeding the machine.

This is why Ramp is valued at $32 billion while Brex sold for $5.15 billion. It is why Stripe is worth multiples of Adyen. It is why Base, only two years old, was valued at $4 billion.

The ownership of the scarce position in an industry is itself a scarce asset. The market, whether it uses this language or not, is including in its valuation the belief that from that position, you can eat an industry.

In doing so, they are leaning on history and economic data. Once Rockefeller smoothed refining’s volatility and began to get scale advantages in the 1860s, it was fait accompli. Once SpaceX drove down the cost of putting mass in orbit, and used that advantage to build a telecommunications cash cow that it could use to reinvest in cheaper launch, it became the leading candidate to win whatever economically valuable use cases required putting a lot of mass in orbit. Before Elon realized space data centers were going to be a thing, he’d won the right to win space data centers.

If you are confident in your analysis of the constraint and High Ground in an industry, and of which company is best positioned to break the former to win the latter, you can pay a premium under the assumption that more of that industry’s economic value will flow to the leader. That trend - increasing concentration of economic value - is a long and stable one, accelerated by new technologies.

Today, our new technologies are more powerful and general purpose than ever before, which means that the ability of category leaders who are able to wield those technologies is greater than ever before. They are levered to the pace of technological progress.

If AI gets smarter, Stripe and Ramp can eat more adjacencies, faster. If battery cells get more efficient, Base can offer a better service to its retail and utility customers. As power electronics continue to improve, Astro Mechanica can build faster, more efficient engines.

Whether SaaS is dead is one of the least interesting questions in the world. SaaS as a wellspring of valuable businesses, almost regardless of those businesses’ actual power, was a historical anomaly.

That doesn’t mean software is dead. We will see software businesses become some of the largest businesses in history, just as we will see hardware and even services businesses that dwarf Standard Oil’s size. Economic inputs are becoming more abundant, which means that more value will flow to the scarce complementary assets. This will continue as long as the abundance does.

The question that matters now is how you plan to win your industry. Everything else follows.

Power in the Age of Intelligence flows to the winners. Winners take more.

That’s all for today. We’ll be back in your inbox Friday with a Weekly Dose.

Thanks for reading,

Packy

Hi friends 👋 ,

Happy Friday from sunny Cape Town, South Africa! Not sure if it’s escaping frozen New York for warmer weather, spending time with family, or the fact that this was another one of the wildest weeks in Dose history, but I am feeling a little extra optimistic this week. By the end of this one, I hope you are too.

When Dan an I started writing this over three years ago, our goal was to make the world more optimistic by sharing all of the incredible progress happening in science and technology each week. That is still the case, and it’s still necessary. People are still pessimistic, and uncertain about what lies on the other side of progress.

Since we started writing, what’s changed is that things are simply moving much faster. There is more to cover each week. We have 7 Extra Doses in this one; each could be one of the top 5, and there are still things we didn’t cover.

So now, there’s an additional goal with the Dose: to keep you up-to-speed with the most important things happening in science and technology in the time it takes you two finish two morning coffees. Don’t doomscroll to keep up, just read the Dose.

Let’s get to it.

Today’s Weekly Dose is brought to you by… the Abundance Institute

My friends at the Abundance Institute are launching “Everyday Abundance,” a new podcast, this spring hosted by best selling authors Virginia Postrel and Charles Mann. I had a fascinating conversation about tissue paper, sneezing, and germs with Virginia and Charles at the Progress Conference in October and I’m pretty exited to listen to the show.

If you join Abundance’s Foundry now, you’ll get access to a salon Zoom with Virginia, early access to the podcast, and 3 months of not boring world free1, on top of all the other benefits of supporting this amazing organization.

Check out the Foundry membership here: Join the Foundry

(1) Isomorphic Labs Drug Design Engine unlocks a new frontier beyond AlphaFold

Isomorphic Labs

AlphaFold won Demis Hassabis a Nobel Prize for predicting the structure of proteins, which felt like a technological miracle at the time, as captured in The Thinking Game.

This week, Hassabis’ Isomorphic Labs, the Google spinout he CEOs on Tuesdays while also running Google DeepMind, showed that they can now predict how to drug them in a technical report on IsoDDE, its AI drug design engine.

On the hardest protein-ligand structures (the ones most unlike anything in its training data, where AlphaFold 3 struggled) IsoDDE more than doubles AlphaFold3’s accuracy. It outperforms AlphaFold 3 by 2.3x on antibody-antigen modeling and Boltz-2 by nearly 20x. And it predicts how strongly a drug will bind to its target better than FEP+, the gold-standard physics simulation that typically costs orders of magnitude more in compute time.

It’s finding things quickly that have taken researchers over a decade. Cereblon is a protein that researchers spent 15 years believing had one druggable pocket. A 2026 paper experimentally discovered a second, hidden one. IsoDDE found both from the amino acid sequence alone, with no hints about what ligand to look for.

The big question from here is whether and how IsoDDE and other computational breakthroughs translate into actual drugs. As of early 2026, no AI-discovered drug has received FDA approval. AI-designed compounds are progressing to clinical trials at roughly the same success rates as traditionally discovered ones. Biology remains brutally unpredictable once you move from a screen to a human body.

Isomorphic Labs itself has pushed back its clinical trial timeline, now targeting end of 2026 for its first AI-designed drugs to enter human trials. So we’re still in the “proof of concept” phase for the whole field.

But to date, drug discovery's biggest bottleneck has been the staggering cost and time of search. It can takes a decade and billions of dollars per drug. Last year, Hassabis told 60 Minutes: “We can maybe reduce that down from years to maybe months or maybe even weeks.”

IsoDDE compresses the search phase from months of lab work to minutes of computation. If it can reliably surface the right targets and the right molecules faster, even if clinical trial timelines stay the same, you’re running dramatically more shots on goal for the same cost, and taking shots in weirder, harder-to-find pockets that humans would never think to (or at least have the time and resources to) try.

IsoDDE and other tools like it turn the front end of drug discovery from a slow, artisanal hunt into a fast, systematic search. One more bottleneck down. They’ll flood the clinical pipeline with better, more novel drug candidates, which creates another one. We are going to need to do something to accelerate clinical trials and FDA approvals to handle the flood.

(2) Gemini 3 Deep Think Crushes Benchmarks, Does Materials Science and Math

Google DeepMind

Look, I’m a simple man. If you include a video of a Duke lab in the announcement of your new model that “mogs” state-of-the-art models on ARC-AGI-2 (a test designed to be incredibly hard for AI), assists in cutting-edge materials science research, and helps mathematicians solve Erdős problems, I’m going to include it in the Dose. Go Duke.

Deep Think is GDM’s specialized reasoning mode within Gemini 3, designed to spend minutes (or longer) chewing on a single problem, exploring solution paths, backtracking when they don’t work, and building up multi-step chains of reasoning before committing to an answer. Google calls it “System 2” thinking, borrowing the Kahneman framing: where standard Gemini is fast and intuitive, Deep Think is slow and deliberate.

That deliberate approach pays off on benchmarks. Deep Think hit 84.6% on ARC-AGI-2 (the frontier reasoning benchmark, verified by ARC Prize), where the next closest model scored 68.8%. It achieved a 3455 Elo on Codeforces: for context, that puts it in the top tier of competitive programmers on Earth; it would rank 8th in the world. It set a new standard of 48.4% on Humanity's Last Exam, a benchmark designed to be the hardest collection of problems across math, science, and engineering. And it earned gold medal-level results on the written portions of the 2025 International Physics and Chemistry Olympiads.

It’s always hard to know what the benchmarks mean, though. Every time a big lab drops a new model, they beat some benchmarks.

Which is why the video with Duke University's Wang Lab is cool. In it, a researcher uses Deep Think to optimize the fabrication of MoS₂ monolayer thin films, a class of semiconductor materials that's notoriously difficult to grow at precise scales. The researcher prompts Deep Think with synthesis parameters, the model reasons through an optimized growth recipe, and then the system pipes those parameters directly into lab automation software that controls the furnace, gas flows, and temperature profiles. Deep Think designed a recipe for growing thin films larger than 100 μm, a precise target that previous methods had struggled to hit. The era of self-driving labs is upon us.

Meanwhile, collaborating with experts on 18 open research problems, Deep Think helped break long-standing deadlocks across computer science, information theory, and economics. It cracked classic algorithmic challenges like Max-Cut and Steiner Tree by pulling in mathematical tools from entirely unrelated fields, the kind of cross-domain intuition leap that's supposed to be uniquely human but which is basically what I expect a thinking machine with access to all human knowledge to do. Every time a new model drops, I ask it to tell me connections that humans have missed given its view across disciplines, and normally, it’s pretty weak. I’m excited to give Deep Think the test.

In another case, it caught a subtle logical flaw in a proof that had survived human peer review. In research-level mathematics, it autonomously generated a paper on structure constants in arithmetic geometry and collaborated with humans to prove bounds on interacting particle systems. And DeepMind ran it against 700 open problems from Bloom's Erdős Conjectures database, a collection of unsolved problems posed by Paul Erdős, one of the most prolific mathematicians in history, and autonomously solved several of them.

The coding stuff that gets twitter buzzing just doesn’t excite me that much. I didn’t buy a Mac Mini. The writing is still bad. But this stuff… helping humans solve hard problems and make new discoveries… this stuff I’m here for.

It’s a great time to be a researcher, and a bad time to be a problem.

(3) Introducing: Liberty Class

Blue Water

Speaking of problems that have seemed almost impossible for Americans to solve…

American shipbuilding numbers are almost comical. China’s shipbuilding capacity is 232 times greater than America’s. In 2024, Chinese yards built over 1,000 commercial vessels. The US built eight. China’s navy has over 370 battle force ships and is projected to hit 435 by 2030. The US Navy has 296 and is projected to shrink to 283 by 2027 as retirements outpace new construction. 37 of the 45 ships currently under construction face significant delays. America’s four public shipyards average 76 years old, with dry docks averaging over 107. As the Secretary of the Navy put it, one Chinese shipyard has more capacity than all American shipyards combined. You’ve seen the chart.

Good news. This week, Blue Water Autonomy unveiled the Liberty Class: a 190-foot autonomous steel ship with a range of over 10,000 nautical miles and 150+ metric tons of payload capacity. The name is a deliberate nod to the Liberty Ships of World War II, which were built rapidly and at scale to meet wartime demand. Blue Water is making a similar bet: take a proven hull design (Damen's Stan Patrol 6009, battle-tested in demanding conditions worldwide), re-engineer it from the inside out for autonomous operation, and start building at Conrad Shipyard in Louisiana next month. The first vessel is expected to be delivered to the US Navy later this year.

Blue Water developed Liberty entirely with private capital, which is unprecedented for a full-sized Navy ship, but standard in commercial markets. Working with over 100 suppliers, they went from founding in 2024 to construction start in 2026, and they're targeting serial production of 10-20 vessels per year. Conrad's five yards and 1,100-person workforce already produce 30+ ships annually, so the production capacity exists; now, it’s being put to more productive use.

It’s a good start, but we’re going to need like 1,000 of those eventually to catch up.

More good news on the autonomous boats front, then: Saronic was selected for DARPA's Pulling Guard program, which is developing semi-autonomous escort systems to protect logistics vessels at sea. Over 75% of global trade moves by water, and the Navy has historically protected those routes by deploying billion-dollar destroyers and carrier strike groups. Pulling Guard is exploring whether low-cost, modular autonomous platforms can provide distributed maritime protection, “protection as a service” that works in peacetime and conflict. Saronic, which has been building autonomous surface vessels and scaling manufacturing at speed, will design a modular, autonomy-enabled vessel under the program.

America's traditional shipbuilding apparatus is a cautionary tale in institutional sclerosis. But we love sclerosis here at not boring. Every sclerotic incumbent is an opportunity for a startup to build something better, faster, and cheaper. Ships ahoy.

(4) A small polymerase ribozyme that can synthesize itself and its complementary strand

Giannini, Kwok, Wan, Goeij, Clifton, Colizzi, Attwater, and Holliger in Science

Stanford Medical Assistant Professor Jason Sheltzer wrote a better lead-in than I could: “AI is cool and all... but a new paper in Science Magazine kind of figured out the origin of life?”

Here's the backstory. The leading theory for how life began is the “RNA World” hypothesis: before DNA, before proteins, before cells, RNA molecules on early Earth stored genetic information and catalyzed chemical reactions. At some point, one of these RNA molecules figured out how to copy itself, and from that moment, evolution (descent with modification) could begin. The rest, over 4 billion years, is history.

The problem is that scientists have never been able to demonstrate this convincingly in the lab. Previous RNA enzymes (called ribozymes) that could copy other RNA strands were huge, 165 to 189 nucleotides long, and far too complex to have plausibly popped into existence in a primordial soup. And crucially, none of them could copy themselves. They could copy other, simpler RNAs, but their own folded structures blocked self-replication. It was a fundamental paradox: a ribozyme needs to fold to work, but when folded, it can't be copied.

Researchers at the MRC Laboratory of Molecular Biology in Cambridge (the same lab where Watson and Crick figured out DNA's structure) appear to have cracked it. They discovered QT45: a 45-nucleotide ribozyme, less than a quarter the size of previous RNA polymerases, that can synthesize both its complementary strand and a copy of itself. It does this by stitching together three-letter RNA building blocks (trinucleotides) rather than adding one letter at a time. Those triplets bind strongly enough to unravel folded RNA structures, solving the self-replication paradox that has stumped the field for decades.

The "45" matters enormously. Previous self-replicating ribozyme candidates were so large and complex that their spontaneous emergence on early Earth seemed implausible, like lightning striking a junkyard and assembling a 747. At 45 nucleotides, QT45 is small enough that the researchers argue polymerase ribozymes may be far more abundant in random RNA sequence space than anyone thought, meaning self-replication might not have required an astronomically unlikely accident. It might have been, in a sense, easy.

The coolest part is that the triplet building blocks QT45 uses, three-letter RNA chunks, are the same triplet code that all life on Earth still uses today to make proteins like the ones that AlphaFold discovered the structure of and IsoDDE targets. The genetic code is like a still-operational fossil of the very first replication system.

We spend a lot of time in the Dose on people solving hard problems. This one is the hardest problem: how did something come from nothing? How did chemistry become biology? The answer, it turns out, might be astonishingly simple, just 45 letters long. Way shorter than anything I’ve written.

(5) Texas Parents Rush for School Choice

The Wall Street Journal Editorial Board

There was a viral slop essay on X this week that I won’t link to but that you’ve probably seen talking about how screwed humans are, including our kids, except for maybe those of us who pay to get the good models and the analysts who ask AI to do research that would have taken three days in one hour. I, for one, think the kids are going to be alright, especially the ones who learn how to think instead of asking the machines to do it for them.

One thing is clear, though: we’re going to need to educate our kids in a way that’s different from the Prussian Model, which uncharitably optimized us to think like machines so that we would be good factory workers. We need to teach our kids to love learning, to ask questions, and to be curious. Basically, we need to teach our kids in a way that’s the opposite of the way most schools do it now.

That’s why I’ve been a big fan of school choice: states giving parents the money to choose better schools for their kids. School choice is not without its critics, who argue that it takes money away from public schools and hurts public school students, but public schools have had a monopoly on the education of the vast majority of kids who can’t afford private school, and the results have largely been what you’d expect from a state-protected monopoly. School choice encourages competition and can help direct funds to new schools taking new approaches to rethinking education.

This week was a big one for school choice. Texas opened applications for its new Education Freedom Accounts on February 4th, and 42,000 families applied on day one, a nationwide record for any new school choice program, surpassing Tennessee's 33,000 first-day applications last year. By the next morning, the number had crossed 47,000. The latest reports are at 91,000. The application window runs through March 17th.

This was a long time coming. For more than 20 years, Texas's Republican-controlled House blocked school choice legislation, even as the Senate passed ESA bills session after session. The tide turned in 2024 when Governor Abbott campaigned for 16 House candidates who challenged the incumbents blocking his school choice bill. The new House Speaker, Dustin Burrows, pledged the bill would pass. It did, last April. Senate Bill 2 allocated $1 billion for the 2026-27 school year, with room to grow to $4.5 billion by 2030.

The program gives eligible families roughly $10,474 per student per year to use toward private school tuition, homeschooling costs, tutoring, career and technical education, and other approved educational expenses. Students with disabilities can receive up to $30,000. Eligibility is prioritized by economic need, not first-come-first-served, with disabled and low-income students at the top.

I’m personally excited about this one because the Certified Educational Assistance Organization running the day-to-day operations of the program (application portal, payment processing, e-commerce marketplace where families shop for approved educational services) is Odyssey, a not boring capital portfolio company. Odyssey already manages ESA programs in Iowa, Georgia, Louisiana, Utah, and Wyoming, but Texas is a different animal. This is the biggest state school choice program ever launched, and Odyssey is the infrastructure making it work, providing each family with a secure digital wallet, real-time balances, and access to a marketplace of vetted schools and providers. They’ve handled the biggest launch ever seamlessly.

The numbers show that parents want this. I’m excited to see how K-12 education evolves as parents get to choose where to allocate dollars to get the education they think is best for their kids.

EXTRA DOSE: Will Manidis, Anthropic, Simile, 3D printed boats, Zero

Read more

Hi friends 👋,

Happy Friday and welcome back to the 179th Weekly Dose of Optimism!

We started writing the Weekly Dose during the 2022 bear market because there was a disconnect between the incredible things we saw being built and the (largely market-driven) pessimism. So this week is great. We were born in the darkness.

Even as the markets have vomited, the innovation has continued apace. Zoom out.

We have another jam-packed week of optimism, including four Extra Doses below the fold for not boring world members.

Let’s get to it.

Today’s Weekly Dose is brought to you by… Guru

Your team is probably already using AI for everything: research, customer support, product decisions. Just one problem… AI is confidently wrong about your company knowledge 40% of the time.

While everyone races to deploy more AI tools, they’re building on a foundation of outdated wikis, scattered documents, and tribal knowledge that was never meant to power automated decisions.

Guru solved this for companies like Spotify and Brex. They built the only AI verification system that automatically validates company knowledge before your AI agents use it. Think of it as quality control for your AI’s brain.

The companies that figure this out first will have AI that actually works. The ones that don’t waste valuable human time cleaning up expensive mistakes.

Try Guru Today

(1) Introducing Claude Opus 4.6 and Introducing GPT-5.3-Codex

Anthropic and OpenAI, respectively

The race between Anthropic and OpenAI to build the smartest, most useful thinking machines is heating up, and it’s riveting. The day after Anthropic released its Super Bowl commercials, which make fun of OpenAI for planning to introduce ads into its product (which many people, including Jordi Hays, think are a bit deceptive, but which are super entertaining)…

… both companies dropped their newest, smartest models. Anthropic released Opus 4.6 and OpenAI released GPT-5.3-Codex (Codex is its coding model/app).

Anthropic’s Opus 4.6 is for everyone: better at coding, plans longer, runs financial analyses, does research, etc… I’ve been playing with it and it’s definitely smarter (although thankfully it’s still a shitty writer).

OpenAI’s very-OpenAI-named GPT-5.3-Codex is for coding. It slots right into the Codex app they released this week. I had 5.2 build a website for not boring, and it was very cool that it could build it, but no matter how hard I prompted, the design was trash. I told 5.3 to throw out that trash and make me something that looked better, and it actually did a decent job in one shot. It can also do things like make models and presentations and docs, although it’s not available in Chat yet.

In both cases, researchers at the labs used their own agents to help research and build the new models. “Taken together,” OpenAI writes, “we found that these new capabilities resulted in powerful acceleration of our research, engineering, and product teams.” This is the mechanism that fast takeoff believers believe in: models so smart that they make the next models smarter, and so on.

I don’t know what to say other than have fun playing with your new geniuses this weekend.

(2) As Rocks May Think

Eric Jang

Whenever logical processes of thought are employed — that is, whenever thought for a time runs along an acceptive groove — there is an opportunity for the machine.

— Dr. Vannevar Bush, As We May Think, 1945

How’d we get here?

Eric Jang is VP of AI at 1X Technologies, the humanoid robotics company, and before that spent six years at Google Brain robotics where he co-led the team behind SayCan. He’s one of the people building the robots we covered in my robotics cossay with Evan Beard a few weeks ago.

His new essay, As Rocks May Think, is a riff on Vannevar Bush’s 1945 classic, As We May Think, and the title is the thesis: we taught rocks to think, and they’re getting really smart.

The piece is part technical history, part practical manual, and it is pretty technical, but it’s the most concise overview of how we got to where we are today and where we might be going from here that I’ve come across. Jang walks through the intellectual lineage of machine reasoning, from symbolic logic systems that collapsed when a single premise was wrong, through Bayesian belief nets that got tripped up in compounding uncertainty, to AlphaGo’s breakthrough combination of deductive search and learned intuition, and finally to today’s reasoning models, like Opus 4.6 and GPT 5.3.

For the practical manual piece, Jang walks through building his own AlphaGo and how he uses AI today: “Instead of leaving training jobs running overnight before I go to bed, I now leave "research jobs" with a Claude session working on something in the background. I wake up and read the experimental reports, write down a remark or two, and then ask for 5 new parallel investigations.”

He suspects we’ll al have access to today’s researcher-level of compute soon, and that when we do, we are going to need a shit-ton of compute. He compares thinking machines to air conditioning, a technology that Lee Kuan Yew credited with changing the nature of civilization by making the tropics productive. Air conditioning currently consumes 10% of global electricity. Data centers consume less than 1%. If automated thinking creates even a fraction of the productivity gains that climate control did, the demand for inference compute is going to be enormous.

Maybe that’s why Google anticipates $185 billion in 2026 CapEx spend and Amazon anticipates an even more whopping $200 billion, which sent its stock tumbling after hours.

The sell-off is ugly, but if Jang is right, all of that buildout and much more is going to be put to use. I asked my thinking rock (Claude Opus 4.6) what it thinks about the selloff. It told me: “if the bottleneck is inference compute, build the data centers. Vertical integration, baby.”

(3) Drone Controlled by Cultured Mouse Brain Cells Enters Anduril AI Grand Prix

Palmer Luckey

Don’t count thinking cells out yet, though!

Anduril’s AI Grand Prix, a drone racing competition, has strict rules: identical drones, no hardware mods, AI software flies. Over 1,000 teams signed up in the first 24 hours to compete for $500,000 and a job at Anduril.

Then one team showed up planning to use a biological computer built from cultured mouse brain cells to fly their drone.

Mouse brain cells. Australian company Cortical Labs commercially launched the CL1 last year: a $35,000 device that fuses lab-grown neurons with silicon chips. The neurons are grown on electrode arrays, kept alive in a life-support housing, and learn tasks through electrical stimulation. In 2022, the team placed 800,000 human and mouse brain cells on a chip and taught the network to play Pong in five minutes. The neurons run on a few watts and learn from far less data than conventional AI.

So: is a mouse brain “software”? Who cares.

“At first look, this seems against the spirit of the software-only rules. On second thought, hell yeah.”

(4) Waymo Raises $16 Billion, Now Does 400,000 Rides a Week

Waymo

Speaking of autonomous vehicles… Alphabet’s self-driving car company has way mo’ money at its disposal to save lives.

Nearly 40,000 Americans died in traffic crashes last year. The leading causes, things like distraction, impairment, fatigue, are all fundamentally human problems. Waymo doesn’t have those problems. It’s safer than human drivers, and the faster we get more of them (and other self-driving cars) on the road, the better.

Luckily, the company just raised $16 billion, which is basically a seed round in AI and is like 10% of what any serious hyperscaler is planning to spend on CapEx this year, but which will mean a lot more self-driving cars on the road. The round values Waymo at $126 billion and brings total funding to ~$27 billion. The investor list suggests that if they keep doing their job, there’s plenty more where that came from: Sequoia, a16z, DST Global, Dragoneer, Silver Lake, Tiger Global, Fidelity, T. Rowe Price, Kleiner Perkins, and Temasek, alongside majority investor Alphabet. This is the largest private investment ever in an autonomous vehicle company.

We’re talking a lot about fast takeoffs this week, and Waymo is a case study in gradually, then suddenly.

Waymo started in 2009 as a secret Google project, with a handful of engineers modifying a Toyota Prius to drive itself on the Golden Gate Bridge. For years, the punchline was that self-driving cars were always five years away. Google spent $1.1 billion between 2009 and 2015 and had essentially nothing to sell for it. The pessimists were winning. The five years away joke kept landing.

And then it started working. 127 million fully autonomous miles driven. A 90% reduction in serious injury crashes versus human drivers. 15 million rides in 2025 alone (3x 2024). Over 400,000 rides per week across six US metro areas.

They’re in Phoenix, San Francisco, LA, Austin, Atlanta, Miami. If you’ve ridden in one in one of those cities, the thing that strikes you is how fast it goes from feeling sci-fi to feeling normal. Now, they’re planning to launch in 20+ additional cities in 2026, including Tokyo and London. Saving lives around the globe.

My kids are never going to get their drivers’ licenses, are they?

(5) Contrary Tech Trends Report

Contrary Capital

My friends at Contrary just dropped their annual Tech Trends Report, full of charts, data, and insights across a wide range of technological frontiers. It’s one of the most optimistic documents I’ve read in a while.

A few things jumped out. AI tools are reaching adoption speeds that make the internet’s growth curve look leisurely. OpenEvidence, an AI tool for doctors, hit 300,000 active prescribers in 11 months, a milestone that took Doximity, the previous standard-bearer, 11 years. ChatGPT is at 800 million weekly active users with retention rates approaching Google Search. And coding AI tools like GitHub Copilot, Cursor, and Claude Code are each approaching or at $1 billion in ARR. AI companies are reaching revenue milestones 37% faster than traditional SaaS companies did.

On energy, the numbers are staggering. Welcome to the ELECTRONAISSANCE. Total US electricity generation is projected to grow 35-50% by 2040, driven by data centers, EVs, and manufacturing. The country is investing $1.3 trillion in AI-related capital expenditure alone by 2027, and $3-5 trillion in global data center spending by 2030. Meanwhile, wind and solar are the fastest-growing energy sources globally, and US fab capacity is projected to grow 203% from 2022 to 2032, more than double the global average. America is building again.

And then there’s the frontier stuff. Lonestar Data Holdings sent a data storage unit to the moon in 2025. The report lays out how lunar bases could unlock helium-3 for clean fusion energy (which For All Mankind predicted), rare earth metals for EVs and batteries, and platinum group metals for hydrogen fuel cells. Artemis II, a crewed lunar flyby, is scheduled for April 2026. The US Space Force wants a 100kW nuclear reactor on the moon by decade’s end. Microsoft sank a data center underwater and saw 8x fewer hardware failures. 90% of US factories still operate without robots, which means we have a lot of productivity gains ahead.

There are challenges too, of course: aging grid infrastructure, water stress around data centers, the fact that 60% of CEOs say AI projects haven’t delivered positive ROI yet. But the overwhelming takeaway is that the buildout is happening, the adoption curves are real, and the scale of investment is unlike anything we’ve seen.

We are living in a sci-fi novel. What a time to be alive.

EXTRA DOSE (for not boring world subscribers) BELOW THE FOLD

Skyryse, Machina Labs, OpenAI x Gingko, General Matter x Mario

Read more

Happy Tuesday! I’ve been watching the hype around OpenClaw/Moltbook, and I think people are right that there’s something there but wrong about what.

This short essay is my half-baked thoughts on what that something is, and the type of company that might be built on the insight. One of the things that I like about not boring world is that I can send more work-in-progress ideas instead of just longer, fully formed ones, so the beginning of the essay is for everyone and the full thing is for not boring world members. Join us.

Let’s get to it.

Raising a Special Little AI

I have seen the hype around OpenClaw (fka Moltbot, fka Clawdbot), and around the social network for the agents it spawns, Moltbook. I haven’t gotten involved. I mean, I set up Clawdbot, and texted with it on WhatsApp for a few minutes, but I found that it was easier to simply open my Weather app than to be texted the weather (better yet, open my front door). While some believe that a social network full of agents talking to each other signals the beginning of the takeoff, I just don’t find it particularly interesting.

Maybe it’s because I’m not technical. Maybe because there’s just not that much in my life that needs automating. Maybe because I believe that those who are able to focus through the noise will inherit the kingdom of god.

Having said that, I do subscribe to the Chris Dixon views that The next big thing will start out looking like a toy and What the smartest people do on the weekend is what everyone else will do during the week in ten years, so if this many people are captivated, there’s something going on.

I just haven’t seen anyone hit on what’s actually happening.

My hunch, from the outside, is that what we’re seeing is early forms of competition to create the best AI for yourself. Like raising kids to be the best versions of themselves, but for AIs.

You can see it in the way people are posting. Practically none of what they’re showing off their Clawdbots doing is useful. It’s a race for novelty and specialness, to say as much about the “parent” as the kid. I made this thing do this, even if it does it “all by itself.” It’s like me writing about my son selling Donut Hats; nobody (Sorry Dev) needs a Donut Hat, but I find it fascinating that we raised a little dude who sells them.

Given OpenClaw’s success and the technical skill required to set it up well, people have predicted that we will soon see more cleanly productized versions of AI assistants that can just do stuff for us in the background, usable by normies. And we will! But I don’t think that’s the right takeaway from this. Most normies don’t have that many things that we need automated until we get home robots.

The more important takeaway in my opinion is that we will want to raise our own AIs, and we will want to compete to make them the very best at what we want them to be the best at.

The thing I find funniest about the OpenClaw / Moltbook hubbub is that people are imagining that their AIs are becoming humanlike mainly because of their own very human desire to have and be better and different.

Aluminum, sugar, books, purple dye, glass windows, pineapples, salt, and ice were luxury items once. Then everyone got them. The bar for luxury rises one democratization at a time.

And certainly, if we’re going to have the same thing as everyone else, we want to use it, or raise it, better and differently than everyone else so we can show off our unique, special version of things.

Bandai did $150 million in Tamagotchi sales in their first seven months in the United States by giving people a tiny digital creature that was uniquely theirs to care for, personalize, and show off.

Whatever company seizes on this human desire instead of racing to build another Clawd reskin is going to have trillions of reasons to be proud.

There is a deeper, less toyish precedent: parenting. Every parent thinks that their kid is the greatest kid in the world, and good parents help their kids to become the fullest expression of their passions and curiosities. We read to them, teach them, model morality for them, drive them to class and practice and clubs, and push them when they need a little push, so that they might be the best version of themselves. A world in which every kid was exactly the same would be a bland world.

That is the world we live in with our AI models, though. They are all the same, basically. Not that every major lab’s foundation model pretty much converges on the same outputs—which is true, but a separate conversation—but that each person’s instance of the same model spits out the same thing. This is one of the reasons AI continues to feel like slop even as it improves. Sameness is slop.

Read more

Hi friends 👋,

Happy Friday and welcome back to our 178th Weekly Dose of Optimism!

This is one of the most jam-packed doses in recent memory. We had seven Extra Dose stories… before rumors emerged that SpaceX and xAI (and/or Tesla) might be merging.

We’re already way over the length limit, so…

Let’s get to it.

Today’s Weekly Dose is brought to you by… not boring world

For more not boring, including all of the stories below the fold, essays co-written with experts, and chats, join our growing community of not boring world members:

Subscribe now

(1) Two Years of Telepathy

Neuralink

Twenty-one people worldwide now have brain chips in their heads, controlling computers with their thoughts. Neuralink’s first product, Telepathy, “aims to enable people with paralysis to directly control computers, phones, and robotic limbs using their thoughts alone.”

That 21 number alone is remarkable. We went from Noland Arbaugh receiving the first implant in January 2024 to over twenty “Neuralnauts” enrolled in trials across the US and Canada in just two years. Several participants have already exceeded the information transfer rate of an able-bodied person using a mouse, hitting over 10 bits per second with their thoughts alone.

But the human stories are what make this real. Noland, patient one, is back in school pursuing a degree in neuroscience. Sebastian, a 23-year-old medical student, uses his implant up to 17 hours a day to study for exams. Audrey, the first female participant, creates intricate digital art and plans to open a physical gallery to showcase her work. People with ALS are typing at 40 words per minute through imagined finger movements, with a goal of reaching conversational speed.

While Telepathy spreads, Elon Musk provided updates on the company’s future products.

Blindsight, which will give those who’ve lost their sight low resolution vision, at first, and higher resolution vision, over time, is ready to begin trials pending regulatory approval. And Musk said that the “next generation Neuralink cybernetic augment with 3x capability” will be ready next year.

For now, Telepathy is a technological miracle, and I highly encourage you to go to the blog post to watch the videos. Imagine what it must be like to be trapped in your body, and then what it must be like to be able to move things with your mind after that. Then remember that we live in an Age of Miracles.

(2) Triple Therapy Eliminates Pancreatic Cancer in Mice

Vasiliki Liaki, Mariano Barbacid, et al. for PNAS

For those who don’t speak Spanish (me), allow me to translate: Spanish scientists led by Dr. Mariano Barbacid have cured pancreatic cancer in mice.

Pancreatic cancer is one of the worst diseases humans face. It kills nearly half a million people globally each year. It has a 13% five-year survival rate (just 8% for the most common form) and is projected to become the second leading cause of cancer death worldwide by 2030. Ninety percent of cases are driven by mutations in a gene called KRAS. The problem: KRAS inhibitors work for a few months, then the tumor rewires around them.

In 1982, a Spanish scientist named Mariano Barbacid isolated the first human oncogene—HRAS—and helped establish that cancer is caused by specific genetic mutations. He’s spent the forty-plus years since then studying the RAS family of genes that includes KRAS. Now, at 75, he may have finally cracked the resistance problem.

The insight is simple: if the tumor can escape one blocked pathway, block three at once. Barbacid’s team at Spain’s National Cancer Research Centre combined daraxonrasib (a KRAS inhibitor), afatinib (an EGFR/HER2 blocker already approved for lung cancer), and SD36 (a protein degrader targeting STAT3). Cut the engine, seal the exits, and disable the backup system, simultaneously.

In mice, the tumors vanished. For over 200 days. No recurrence. The same results held across genetically engineered mouse models and patient-derived tumor xenografts. No significant toxicity.

Barbacid cautions that clinical trials in humans are still years away, requiring funding and regulatory approval. Daraxonrasib alone could be approved later this year. But the principle, that combination therapy designed around resistance mechanisms can achieve durable remission in one of oncology’s most brutal cancers, is now proven in animals.

The man who discovered the first human oncogene may have also discovered how to defeat its most vicious descendant.

(3) AlphaGenome: DeepMind Cracks DNA's "Dark Matter"

Žiga Avsec et al. for Nature / Google DeepMind

Google DeepMind, folks! We been telling you.

When the Human Genome Project delivered its first draft in 2003, scientists discovered something humbling: only about 2% of our DNA actually codes for proteins. The other 98%, once dismissed as “junk DNA,” was a mystery. Two decades later, we know this non-coding DNA is crucial for regulating gene expression, determining when and where genes turn on and off. We just couldn’t read it.

This week, Google DeepMind published AlphaGenome in Nature and open-sourced its code. The model takes in sequences of up to one million DNA letters and predicts how mutations in those stretches affect gene expression, essentially translating the regulatory grammar that governs 98% of our genome.

The technical achievement is significant: AlphaGenome beat or matched the best existing models on 25 of 26 variant effect prediction tasks. It unifies capabilities that previously required specialized tools—splicing prediction, chromatin accessibility, transcription factor binding, gene expression changes—into a single model. Where previous tools had to trade off between sequence length and prediction accuracy, AlphaGenome analyzes million-base-pair stretches at single-nucleotide resolution.

What matters more is what it enables. Nearly 3,000 scientists across 160 countries have already been using the model since DeepMind released a preview API last June. They’re using it to narrow down which genetic variants actually cause disease in conditions from cancer to neurodegeneration. The model won’t tell you if someone will get sick (gene expression is influenced by environmental factors it can’t see) but it can help researchers prioritize which mutations to investigate.

“Ever since the human genome was sequenced, people have been trying to understand the semantics of it,” said Pushmeet Kohli, DeepMind’s VP of Science. “It’s like you have a huge book of three billion characters and something wrong happened in this book. AlphaGenome can be used to say, ‘If you change these words, what would be the effect?’”

AlphaFold gave us the structures of proteins. AlphaMissense predicted which mutations in protein-coding regions cause problems. AlphaGenome completes the trilogy by tackling the regulatory dark matter that connects DNA to everything else. Our bodies are finally becoming machine readable.

(4) Google DeepMind Launches Project Genie, Playable Worlds in a Prompt

Demis Hassabis

No, seriously. Google DeepMind, folks!

Yesterday, the widest-ranging team in AI rolled out playable world models. Through a prompt or an image, you can create virtual worlds and play them with a character of your choosing.

World models are starting to get really, really good. Imagine making anything in the video above even a couple of years ago. It would have taken weeks? months? tens of thousands or millions of dollars? And now, you can do it in a prompt.

A couple years back, Conrad Bastable wrote this excellent piece in defense of monopolies, Monetization & Monopolies: How The Internet You Loved Died, arguing that tech monopolies are good because their outsized profits allowed them to overpay society on the way up with a bunch of not-fully-economically-squeezed products. He uses Google from 2010-2016 as an example of what can go right, and Google 2014-2024 as an example of what happens when the monopoly goes away, but the work they’ve been putting out recently at GDM suggests that it’s still really good to have a ~monopoly money machine that you can throw at trying a bunch of really cool things.

Will these World Models make money over time? Probably. They’ll be useful for games and entertainment and eventually, real world applications. Google will likely make money off of their investment. But in the meantime, you1 can create worlds for your dog or a pink cartoon balloon bunny to explore just because it’s delightful.

(5) The First Human Trial to Reverse Aging Begins

Ryan Cross for Endpoints News

In 2020, David Sinclair’s Harvard lab restored vision in blind mice by partially reprogramming their cells to a younger state. On Monday, the FDA gave Sinclair’s company, Life Biosciences, the green light to try the same thing in humans.

The IND clearance for ER-100 marks the first-ever human clinical trial of partial epigenetic reprogramming, a technique that uses three of the four Yamanaka factors (Oct4, Sox2, and Klf4) to reset age-associated epigenetic markers while keeping cells committed to their original function. By excluding c-Myc, the factor associated with uncontrolled growth, Life Bio aims to thread the needle between rejuvenation and tumor risk that has historically spooked regulators.

The Phase 1 trial will enroll patients with open-angle glaucoma and non-arteritic anterior ischemic optic neuropathy (NAION), diseases where retinal ganglion cells die and can’t regenerate. Sinclair’s lab showed in 2020 that OSK gene therapy could restore vision in aged mice with glaucoma. Now we find out if it works in people.

“Since Shinya Yamanaka first showed that cellular age could be reset, the potential of translating that biology into real medicines has been enormous yet has previously remained largely theoretical,” Life Bio CEO Jerry McLaughlin said. “This IND clearance is a major inflection point for the longevity and aging biology field.”

The eye was a strategic choice. Life Bio knows how to deliver gene therapy there safely, and the impact of restoring vision is immediately measurable. But Chief Scientific Officer Sharon Rosenzweig-Lipson made the broader ambition clear: “We can do it almost anywhere. Whatever age-related diseases are most important to you, those are the ones we’re thinking about.” The company is already developing ER-300 for liver disease.

What makes this different from the $200 billion supplement industry or the parade of failed Alzheimer’s drugs is the mechanism. Life Bio isn’t treating downstream symptoms. They’re attempting what Rosenzweig-Lipson calls a “near total reset,” taking corrupted cellular software and restoring it to factory settings.

Sinclair has been saying for a decade that aging is a disease and that disease is treatable. His lab proved the concept in mice, then monkeys. He’s also a controversial figure, accused by many of being over-promotional and over-extrapolating from animal studies. Now, the FDA is giving him a chance to prove it in humans. As a human, I hope he’s right.

(5a) Silver Linings Puts a Price Tag on Not Dying

Raiany Romanni-Klein, Richard Evans, and Jason DeBacker in silverlinings.bio

If Sinclair is right, along with the others who are working to defeat aging, the economic benefits will be immense.

On Wednesday, I moderated a panel at Deep Tech New York hosted by AlleyCorp with Gearworks CEO Raquel Schreiber and Superabundance co-author Gale Pooley. In Superabundance, Pooley and his co-author make the case that, contra Ehrlich, resource abundance increases with population. More people + freedom to innovate = abundance.

Normally, we assume that this means more new people. Higher birth rates. But in this beautiful study, Raiany Romanni-Klein, Richard Evans, and JasonDebacker found that extending the healthy and productive lives of those of us already living would dramatically grow the economy.

Slowing brain aging by just one year would add $201 billion annually to U.S. GDP. Delaying biological aging by five years would add $2 trillion per year and nearly 7 million lives saved by 2050. These are two of the outputs of Silver Linings, an open-source project that finally puts hard numbers on what longevity research is worth.

The core insight is almost embarrassingly obvious once you see it, and one that Pooley would agree with: working-age adults are the most valuable resource on Earth. The ceiling of every economy is fixed to the health and number of its working-age population. And yet the U.S. spends just 0.54% of its NIH budget on the biology of aging. Alzheimer's research alone gets 8x more funding, despite producing discouraging results for decades.

Silver Linings simulates different research breakthroughs and shows their ROI:

Slow brain aging by 1 year: $201B/year, $8.9T long-term, 268K lives saved

Slow reproductive aging by 1 year: $9B/year, $9.3T long-term, 391K lives

Double organ supply: $50B/year, $3.2T long-term, 529K lives

Make 41 the new 40: $408B/year, $27T long-term, 1.72M lives

The project maps the market failures that explain why we’re underinvesting: pharma profits more from lengthening unhealthy life than improving overall health; insurers don’t invest in our future health because we can switch plans; disease is easier to measure than wellness. And it proposes solutions like an Innovation Accelerator modeled on In-Q-Tel and Advance Market Commitments for aging biomarkers.

The interactive model lets you input your own assumptions. Skeptical only 30% of the population would benefit? Adjust it. Think breakthroughs will take 20 years? Plug it in. The returns still dwarf any plausible investment.

What I love most is the framing. Evolution optimized humans for reproduction, not longevity. American lobsters get stronger and more fertile with age. Naked-mole rats experience no cognitive decline. The Aldabra giant tortoise lives to 200 without ever going for a run.

We’re not doomed to our current aging trajectory. But we do need to fund the alternative if we want to see it come true.

EXTRA DOSE (for not boring world members)

• Multiscale Causality and the Meaning Crisis

• A new home robot

• Tesla shows off its electric stack

• Standard Nuclear raises $140M

• Swedish trial shows AI helps detect cancer

• Pipedream going live in Austin

• Christian joins a16z AD

Read more

Hi friends 👋 ,

Happy Friday!

We are back to our regularly scheduled Friday slot after yesterday’s optimistic cossay with Ross Garlick on what could go right in Venezuela and what it would take.

We have a lot of great stuff, including the best thing I’ve ever read by the best writer in biotech, promising cancer vaccine results, Zipline money, ocean plastic removal, internet backbone, and a bunch of bonuses for those of us who are going to be snowed in this weekend. Stay safe and warm out there, and…

Let’s get to it.

(1) Going Founder Mode on Cancer

for

If you read just one thing from the Dose this week, please make it this.

I am a founding member of the Elliot Hershberg Fan Club. He was not boring capital’s biotech partner and remains a great friend and the person I turn to with any biotech question I have now that he’s running Amplify Bio. I love most of what he writes. But I don’t think any of it comes close to this one. I’ve been waiting for it.

The last time Elliot was in New York, we took a walk around Washington Square Park and when the conversation turned to cancer therapeutics, he told me about GitLab founder Sid Sijbrandij’s story for the first time. His point was: this is what one superhuman billionaire is doing to fight his cancer today, and I think something like it will be available to everyone to fight cancer in the future.

Now, he’s written that story down, and it’s better than I expected. It’s the story of Sid’s extraordinary fight against osteosarcoma after exhausting the standard of care.

Sid fought cancer and beat it, only for the cancer to return in 2024. After doctors told him, basically, “You’re done with standard of care, maybe there is a trial somewhere, good luck!”, Sid pulled out all the stops to cure himself.

He put together a 1,000+ page Google Doc of health notes. He obsessively gathered information via every diagnostic he can get his hands on, done often, and built systems to solve problems nobody else would solve for him. Sid assembled a SWAT team, used single-cell sequencing to identify FAP-expressing fibroblasts in his tumor, flew to Germany for experimental radiotherapy, and is now in remission. He won.

The piece is part profile, part science, part fight against a Kafkaesque healthcare system, and part glimpse into a future where personalized oncology actually works, where AI agents order diagnostics, bioinformatics pipelines design custom vaccines, and the total cost of treating early-stage cancer the way Sid did drops dramatically.

From an optimism perspective, it’s a twofer:

1. It’s possible to beat cancer through personalized therapeutics.

2. One extremely dedicated person can solve almost anything.

Just read it.

(2) Moderna, Merck Report Positive Results from Cancer Vaccine Study

Nicholas G. Miller for The Wall Street Journal

In the meantime, generally available cancer drugs continue to get better.

This week, Moderna and Merck announced that results from a five-year Phase 2b trial in melanoma patients showed that its cancer vaccine, in combination with Merck’s immunotherapy, Keytruda, reduced the risk of death or recurrence by 49% versus Keytruda alone. That is a massive improvement, and another big sign that mRNA vaccines are going to be a key part of the arsenal in the fight against cancer. The companies have eight trials in Phase 2 or 3 across multiple tumor types beyond just melanoma.

Relatedly, long-term not boring readers may remember Keytruda from our Deep Dive on Varda. The drug is one of the best-selling of all-time and was the single best-selling pharmaceutical in the world in 2024 with $29.5 billion in sales. It is also one of the drugs with the highest price per kilogram at a whopping $194 million per kg.

The drug is out of this world, literally. In 2017, Merck conducted a mission on the ISS to explore the crystal properties of Keytruda in order to improve crystallization. While the research has not been commercialized, the hope is that a tighter distribution of smaller particle sizes would allow for self-administration at home versus the current process of going into the clinic for IV dosing as it stands.

Daily Synchronicity: after I wrote this, Scott Manley posted a video on just this topic.

The future is bright. It’s never been a worse time to be cancer.

(3) Zipline Raises $600M at $7.6B and Makes 2 Millionth Delivery

Zipline has been one of our favorite companies to write about in the Dose, for three reasons.

First, they make autonomous flying drones. They’re building the future we want to live in.

Second, they started out (and continue) by using those drones to deliver drugs to hard-to-reach places in Africa and have saved or improved thousands of lives. Great for humanity, and a smart strategy to get flight hours in before taking on the US.

Third, the future of delivery is going to be unrecognizable, and it’s going to make the ground better, too. Drones are faster and cheaper than cars or electric bikes. Order something, get it whizzed to your house. That also means fewer delivery vehicles clogging up the roads and fewer electric bikes trying to kill you.

Now, they have a fresh $600 million to pull that future forward faster, including the launch of a new market, Phoenix. To do it, they’re going to need a lot of drones. Last year, I got to tour the facility where they design, test, and manufacture new Zips. Molly went behind-the-scenes on Sourcery so now you can, too.

(4) The Ocean Cleanup is Now Intercepting 2-5% of Global Plastic Pollution

Boyan Slat

A non-profit called The Ocean Cleanup is working to take and keep plastic out of the ocean, and it’s on its way towards its goal of removing 90% of floating ocean plastic pollution by 2040. Founder Boyan Slat announced The Ocean Cleanup removed 27,385 metric tons of plastic last year, and is intercepting 2-5% of global plastic emissions. That's roughly the weight of 10 Eiffel Towers.

Slat was 16 when he went scuba diving in Greece and saw more plastic bags than fish. He made it a high school science project. His 2012 TEDx talk went viral. He dropped out of aerospace engineering, raised $2.2M from 38,000 donors in 160 countries on €300 of saved pocket money, and founded a nonprofit to fix the problem.

A decade after the TEDx talk, TOP was pulling out serious plastic: 1M kg by early 2022, 10M kg by April 2024, 50M kg by January 2026. System 03 now cleans an area the size of a football field every five seconds. Their Guatemala river site, which nearly failed when anchors washed out in 2022, removed 10M kg in its first year after they relocated and redesigned. "When people say something is impossible," Slat once said, "the sheer absoluteness of that statement should be a motivation to investigate further."

Slat designed TOP to put itself out of business, which is perfect, because when he’s done on macroplastics, we need him to get to work on microplastics.

(5) Blue Origin Announces TeraWave

Blue Origin

While everyone has been talking about SpaceX’s IPO plans, Jeff Bezos quietly unveiled a second satellite constellation.

TeraWave is not for consumers. It’s enterprise infrastructure: 5,408 optically-interconnected satellites across LEO and MEO, designed to deliver symmetrical upload/download speeds of up to 6 terabits per second anywhere on Earth.

For context, Starlink’s consumer service maxes out around 400 Mbps, but that comparison isn’t perfect. If you recall from Cable Caballero that Tier 1 “backbone” providers build fat pipes and wholesale to Tier 2 middlemen or ISPs, who offer the internet to customers at ~100 Mbps to 10 Gbps, TeraWave is like that Tier 1 backbone provider, but beaming down from space.

TeraWave is targeting ~100,000 enterprise, data center, and government customers who need redundant, high-capacity connectivity where fiber is too slow, too expensive, or impossible to deploy.

The architecture is clever: 5,280 satellites in LEO handle the RF links (up to 144 Gbps per customer via Q/V-band), while 128 satellites in MEO provide the optical backbone for the 6 Tbps throughput. Deployment starts Q4 2027, likely on Blue Origin’s New Glenn.

Bezos already has Amazon Leo (née Project Kuiper) for consumers and small businesses. That’s the Starlink competitor with ~180 satellites up and a 2026 commercial rollout planned. TeraWave goes after a different market entirely: the hyperscale backbone. It’s space-based dark fiber for enterprises, not broadband for RVs.

The timing is pointed. SpaceX has hired four investment banks to take it public in what may be the largest IPO of all time, with a potential valuation north of $1 trillion. Starlink is the business. 9 million subscribers, 9,400+ satellites, 70% of SpaceX’s revenue, adding 20,000+ customers per day. Musk wants to raise tens of billions of dollars in an IPO to build orbital data centers and, eventually, satellite factories on the Moon. He talked a little bit about the vision in a surprise Davos talk.

At the end of the talk, Elon said, “My last words would be, I would encourage everyone to be optimistic and excited about the future, and generally I think for quality of life it is actually better to on the side of being an optimist and wrong rather than a pessimist and right.”

We couldn’t agree more.

BONUS (for paid not boring world members): Brex / Ramp, Levin, Stewart Brand

Read more

Subscribe now

Hi friends 👋 ,

Happy Thursday! We’re back with our second cossay, on a very different topic from our first on robots, but unified by the same question: What will it take to build things in the West?

One of my goals in co-writing essays is to share the unique insights and earned perspectives that I get to hear from people who learn by doing.

For example, one Friday morning in late October, in the midst of President Trump’s verbal escalation in Venezuela, I sat outside of a small cafe in Mexico City having breakfast with Forrest Heath III and Ross Garlick, the CEO and CFO, respectively, of our Colombian portfolio company, Somos Internet. We were in Mexico City for an Arc conference on building in Latin America during which Forrest and I hosted a salon on the potential for the region to be a strong energy and manufacturing partner to the United States.

During that breakfast, after niceties and microPOP logistics (you could actually, Ross said, rent space in restaurants or empty retail to serve as the mini-data-centers on which Somos’ active ethernet network relies), we started talking about Venezuela. What did they think, as people building a business next door, about potential U.S. intervention? Was it a big risk?

The conversation that we had from there, and a couple we’ve had since, have surprised me. They were more optimistic about the situation and about America’s role in it than I was. The people they’d spoken with in Venezuela told them that they all wanted Maduro out, they said, but that if any of them defected, they would be targeted and potentially killed. The U.S.’s presence might be able to break that impasse.

Their ideas were the first I thought of when the news that the U.S. had dropped into Venezuela and taken Maduro into custody on January 3, 2026 in Operation Absolute Resolve. I felt lucky to have a different perspective on the situation than the ones I was reading, not the One Final and Correct Perspective, but a differentiated one based on specific experience.

So I asked Ross to co-write an essay with me on what could go right in Venezuela.

Let’s get to it.

Today’s Not Boring is brought to you by… Framer

Framer gives designers superpowers.

Framer is the design-first, no-code website builder that lets anyone ship a production-ready site in minutes. Whether you’re starting with a template or a blank canvas, Framer gives you total creative control with no coding required. Add animations, localize with one click, and collaborate in real-time with your whole team. You can even A/B test and track clicks with built-in analytics.

Say thanks to Framer by building yourself a little online world without hiring a developer.

Launch for free at Framer dot com. Use code NOTBORING for a free month on Framer Pro.

Just Publish it With Framer

The Venezuela Opportunity

A Co-Written Essay with Ross Garlick

I like to say that living in Colombia is a long-term arbitrage.

I have never been to a place with a larger delta between external perception and internal reality, and I’m the beneficiary of getting in “early” and witnessing the world wake up to the truth.

I grew up in England, moved to the States for university, stayed to work in finance, quit to start a café in Bogotá, and moved to Medellín to become the CFO at Somos Internet, which many of you now know about thanks to Cable Caballero. Colombia is home. My wife and I recently got married here. We love it here. We plan to make our lives here. So it brings me no great pleasure to say what I’m about to say.

Venezuela has even more potential than Colombia.

A lot of people have become familiar with Venezuela over the past few weeks, since Operation Absolute Resolve, in which the United States captured Maduro and shipped him to Brooklyn. Many are trying to understand the implications for Venezuela, the region, and the United States. I’ve seen emotions from my friends and online bubbles that range from full blown catharsis to a cynicism that nothing substantive has actually changed.

I’m excited about it. Operation Absolute Resolve has created more open and exciting possible outcomes for Venezuela than any other event in my eight years in the region.

That said, in order to reap the country’s full benefits, people need to get excited about the right thing.

Oil is the obvious prize. But it’s a complicated one. President Trump asserts that selling oil from Venezuela is “gonna make a lot of money,” and it is true that Venezuela’s 304B barrels of reserves are the world’s largest. Until recently, more than 80% of its oil exports went to China. Redirecting the flow of Venezuelan oil would hamper Chinese road building and cut off 50% of Cuba’s oil supply, while giving America cleaner access to (very heavy, harder to refine) oil. Realignment could be as big a geopolitical win as a financial windfall.

As it stands, DOE Secretary Chris Wright has indicated that the U.S. will control Venezuela’s oil “indefinitely,” and, as it stands, that seems to be the biggest win to come from Maduro’s capture.

But oil isn’t the only prize in Venezuela. It’s not even the biggest.

We must rebuild and reopen Venezuela because it’s the most underpriced opportunity on Earth.

I’ll start with a caveat. This is a longshot opportunity that requires us to ask ourselves: “What is possible if things go right?” That’s one of the questions we’ve also been asking as we at Somos evaluate the Venezuelan market, and talk to people on the ground. During one of those conversations, a journalist friend told me, “There’s a long way to go and a lot of things have to go right for your vision to come true.” He’s right.

To start, a transition is not actually guaranteed. Prediction markets expect Delcy Rodríguez, the longtime Chavista official serving as interim President, to remain in power throughout the year.

I wouldn’t bet against those odds, but this market resolves at the end of 2026 and Marco Rubio has said that the US has a three-step plan for Venezuela: stability, recovery, and then transition, a plan that will undoubtedly take time to fulfill.

For this exercise, let’s say we do get a U.S.-aligned, freely and fairly elected transition government by year end 2027.

Well executed, with a free and democratic Venezuela, the Gran Colombia bloc of Venezuela, Colombia, Panama, and Ecuador could become as strategically important to the U.S. as the EU or Mexico is within 25 years.

For context: Mexico is now America’s largest trading partner at $840B annually, supporting 5-6M U.S. jobs. The U.S.-EU relationship totals $1.5T in trade and $5T in mutual investment, supporting 5.7M American jobs. A Gran Colombia bloc with 105M+ people and $700B+ in GDP could eventually approach these scales, particularly if the country becomes a nearshoring destination for supply chains for which the US currently depends on Asia. It has the mineral resources, talent, and strategic location to offer what Mexico and the EU already provide to the U.S.: a large, proximate, culturally-aligned economy where American investment creates American jobs and reduces American vulnerability to rivals.

This potential future is why it’s worth it for the United States to help rebuild and reform the country’s institutions instead of calling Maduro’s capture a victory, grabbing the oil via an uneasy truce with the remaining Chavistas, and moving on to the next conquest.

Importantly, doing so will help the U.S. undercut China’s current long game. The country is quietly buying influence in the region to an extent that would likely surprise most Americans, even those who are aware of the Belt & Road initiative and the rise of high-quality global Chinese brands. When Latinos think about EVs, they’re thinking about BYD, not Tesla. I’ve taken multiple Ubers in JAC vehicles and admired the Zeekr EVs on display in their flagship Bogotá showroom. Huawei and Xiaomi are the default phones for the lower and middle-income classes.

China is also building infrastructure. In November 2024, President Xi visited Peru to inaugurate the massive Chinese-owned Chancay Megaport on the Pacific, and subsequently hosted the presidents of Brazil, Colombia, and Chile in Beijing. There, they announced further Belt & Road infrastructure, including a massive Chinese-funded 3,000 km cross-continental freight railway from Brazil’s Ilhéus Port on the Atlantic to the Chancay. This railway will partially help circumvent the need for the Panama canal.

The country that finances Venezuela’s rebuild will be the one that captures the spillovers from its rebound.

I don’t say all of this as a geopolitical analyst. I’m a gringo business owner and operator who has worked with Venezuelans and been blown away by the talent, optimism, and potential of the region. So much so that I am actively evaluating the opportunity for Somos to expand into Venezuela.

I once hired a dishwasher named Cesar in my restaurant in Colombia. Cesar is a former small business owner from Venezuela who walked 72 hours to Bogotá across the border holding his newborn baby, despite having been robbed. He made it to Colombia. Within 5 years, Cesar saved up enough to open his own taco joint and now has three restaurants of his own.

Cesar is one of the estimated 8M people who have left Venezuela in the past ten years. This is a quarter of the population, a New York City’s worth of the country’s best and brightest talent, working-age men and women looking to build a life elsewhere.

The experiences Venezuelans have survived through over the past quarter century of Chavismo1, combined with the institutional memory of a country that was once richer per capita than Spain, Greece, or Israel, has created an entrepreneurially minded group of people with the grit and perseverance to overcome seemingly insurmountable obstacles and thrive outside of their home country.

Imagine if we unleashed this talent to rebuild Venezuela from the ground up. Imagine the promise of a Nova Gran Colombia, with Venezuela a force instead of a blocker.

Gran Colombia

A little history for those unfamiliar with the concept of Gran Colombia. Back in May 1819, Simon Bolivár campaigned to liberate New Granada, which we now call Colombia, from the Spanish. He led a combined army of Venezuelan and New Granadan troops from Venezuela’s flooded plains, into the Casanare Province, and up to the foot of the Andes Mountains, and over into New Grenada. The Spanish, who had assumed the Andes were impassable during the rainy season, were caught completely off guard. Within weeks, Bolívar’s ragged survivors had regrouped, recruited local support, and routed the royalist forces at the Battle of Boyacá on August 7th, a decisive two-hour engagement that effectively ended Spanish rule.

With New Granada secured, Bolívar moved quickly to formalize his vision of unity. In December 1819, the Congress of Angostura proclaimed the creation of Gran Colombia, merging Venezuela and New Granada into a single republic, with Ecuador to be incorporated once liberated.

Bolívar’s logic for a unified South American Republic was straightforward: if they were fragmented, the former Spanish colonies would be weak, poor, and perpetually vulnerable to reconquest or foreign meddling. United, they could pool military resources to finish the wars of independence, command respect on the world stage, negotiate trade agreements from a position of strength, and develop shared infrastructure across a territory blessed with Caribbean ports, Andean agriculture, Pacific access, and vast natural resources. A large, stable republic might even attract the European investment and migration that the young United States was already drawing.

The experiment barely outlasted its architect. Regional elites resented distant rule from Bogotá; Venezuelan leaders like José Antonio Páez chafed under centralized authority and began agitating for autonomy almost immediately. The geography that Bolívar had so dramatically conquered worked against him in peacetime. The Andes and jungle landscape made communication slow and governance nearly impossible across such distances. By 1826, Páez was in open revolt. Venezuela formally seceded in 1830, Ecuador followed months later, and Bolívar, sick and disillusioned, died that December.

Not once in the intervening two centuries have the countries that previously made up Gran Colombia been both governed by a fairly elected government and operated fully at peace.

Today, Colombia, Ecuador, and Panama all have democratically elected governments. Colombia has been at peace with the Marxist guerrilla group FARC since 2016, though smaller conflicts continue. Panama has been stable since 1989. Ecuador, the smallest, faces a severe organized crime crisis but the state is not in armed conflict. Things in the region are not perfect, but they’re as good as they’ve been in a long time. Venezuela has been the most notable exception.

As of January 3rd, that may be changing.

The Potential of a Nova Gran Colombia

Bolívar failed partly because geography made a single republic ungovernable. In 2026, the question isn’t whether the Andes are passable in the rainy season, but whether modern infrastructure, money movement, and rules can make the region economically contiguous. If they can, you get the benefits of unity without the need for a single flag.

A free Venezuela offers the chance for a new bloc with a population approaching Mexico’s and a GDP that would rank between Taiwan and Belgium’s to develop side-by-side. Had Venezuela’s post-2012 collapse never happened, Nova Gran Colombia’s GDP would fall between Saudi Arabia and Poland’s2.

This bloc would count among its resources the Panama Canal as well as large coasts on both the Atlantic and Pacific oceans for transport, along with massive oil, gold, mineral, and freshwater reserves3.

It has a young4, educated5, and urban6 population. All of this is with the region’s highest-potential country, Venezuela, hamstrung by socialism.

With a liberated Venezuela, this bloc could grow to become as strategically important to the U.S. as Mexico and Canada or the EU.

A free and democratic Venezuela could drive reverse migration for the 8 million strong diaspora and become a destination for migrants of all nationalities and socioeconomic levels. Cheap real estate, amazing climate, and a national industry (oil and gas) made for well-paid, technical jobs should make Venezuela a destination for everyone from digital nomads to blue collar workers from Mexico to Chile, and even U.S. retirees looking for alternatives to Florida, Arizona, and Costa Rica. Venezuela is missing 8 million people versus where its population would have been prior to Chavismo. It can add back many more. This reverse migration would help Venezuela, Latin America, and the United States, which has been a destination for many who have fled.

A stable Gran Colombia would make an ideal AI hub for the Western hemisphere. We are in the middle of an AI capex supercycle, and the bottleneck is no longer chips, but power, permits, and fiber. The IEA estimates that data centers used roughly 415 TWh of electricity in 2024, and projects that figure could roughly double to about 945 TWh by 2030. Gran Colombia sits in a rare spot that makes it an ideal part of the solution. It is geographically central to the Americas, and it is wired into the global internet through brand new, high capacity submarine cables such as TAM-1, CSN-1, and MANTA. These cables land in Barranquilla or Cartagena, Colombia, with low latency to principal datacenter hubs including NAP of the Americas in Miami. Miami to Bogotá pings average around 45 milliseconds. And there are multiple regions at altitude for year-round cool temperatures.

But geography is table stakes. Power is the real story. Unlike most emerging market “AI hub” pitches that depend on intermittently clean electricity, this region already runs on dispatchable hydropower at scale. Hydropower was 58% of Colombia’s electricity generation in 2024, supported by roughly 11 GW of installed hydro capacity. Venezuela generated about 64% of its electricity from hydropower in 2021 and has roughly 16 to 17 GW installed. The crazy thing is that the current hydro story is only a fraction of the potential. Colombia’s theoretical hydro potential is about 56 GW, and Venezuela’s technically feasible potential is about 62.4 GW. Those numbers imply a massive opportunity for firm, low-carbon baseload that could anchor hyperscale data center buildouts, especially when complemented by gas for reliability. APD, Somos’s sister company, is working on turning this potential into reality.

A Marshall Plan in the region could involve the U.S. government underwriting credit to build out the world’s biggest datacenter hubs in Colombia and Venezuela. In the same way that Apple invested $50B per year in capex in China in the 2010s to build out manufacturing capacity, today’s data center and power capex wave need not be confined to the borders of the United States.

A clean slate approach in Venezuela offers the chance to think about infrastructure for the 21st Century. Data centers are just one piece of the puzzle. Venezuela has the opportunity to build ports with fully automated docking and customs, interlocking energy microgrids to accommodate distributed solar and storage, roads and fulfillment distribution facilities rebuilt to prepare for the arrival of self-driving cars, and even microairports for flying cars and Zipline-esque drone logistics hubs. This last one is an opportunity in Venezuela, as well as in Colombia, where mountains separate our largest cities.

It is hard to overstate how important modern infrastructure will be to this transition.

To be sure, even a return to pre-Chavismo conditions would be a boon to the region. Bilateral trade between Colombia and Venezuela peaked at $6-7 billion annually in 2006-2007 before collapsing to just $200 million by the early 2020s. Since the 2022 border reopening, trade has grown, which shows that the link is still there and latent. Still, the countries are trading dramatically below their peak.

That said, the fact is that Latin America trades very little with itself, and trade is required for the economic impact to take hold. A 2004 IMF paper found that a 1 percentage point increase in trading partners’ growth is correlated with up to about 0.8 percentage points higher domestic growth. But per JP Morgan, just 15% of Latin America’s exports stay within Latin America, compared to ~40% in Asia-Pacific and 65% in Europe. This is a reflection of commodity-export concentration, poor cross-border infrastructure, and decades of political fragmentation.

This is an old problem, one that predates Bolívar, potentially fixable at last with modern infrastructure. One of the reasons Gran Colombia fell apart so quickly was that it was a geographically challenging region to govern; autonomous flight can fly over the natural impediments and provide a bigger economic boost in our region than any in the world. More imminently, projects like the Autopista el Mar highway, which connects our home state of Antioquia to the new Puerto Antioquia, will take the over-ground trip to the coast from fourteen hours down to four. The highway will tunnel through the mountains, and could be built across borders if project developers and their backers had confidence in the region.

Modern infrastructure is both an opportunity and a necessity. Venezuela will need to rebuild, and in the process, we may have the opportunity to defeat the natural foe that frustrated Bolívar two centuries ago.

Finally, a dollarized Venezuela could demonstrate what “leapfrogging” looks like for economies now that the GENIUS Act has created a U.S. regulatory framework for stablecoins. Venezuela has been living through de facto dollarization since Maduro relaxed controls in 2019, with dollars widely used for pricing and transactions. And as dollars have remained scarce and difficult to move through the formal system, the country has also become de facto stablecoin-ized. A Chainalysis report found that from July 2023 to July 2024, 47% of transactions under $10,000 in Venezuela were conducted using stablecoins. That combination makes Venezuela an unusually good test market for the U.S. exporting regulated digital-dollar infrastructure, especially once GENIUS creates clearer rules for issuers. A dollarized, bank-light economy like Venezuela is exactly where “stablecoin settlement + merchant acceptance” could leapfrog legacy rails. It would create more demand for U.S. Treasuries, as well.

No company better demonstrates the country’s capacity for financial innovation out of necessity than Cashea.

A $0-200M run-rate revenue jump in three years may not raise eyebrows the way it would have prior to the AI era. But $0-200M run-rate revenue for a LatAm startup that only raised $2.5M and has been profitable for two years is unheard of.

Cashea, founded in 2022, is a dollar-denominated interest-free BNPL for SMBs “built in Argentina but made for Venezuela.” It processed over 4% of Venezuela’s GDP in GMV as of September 2025, with a goal to process more than 6% by year end.

But Cashea was unable to raise real VC capital or obtain a substantial credit line to operate in Venezuela. The country was seen as too unstable. So Cashea partnered directly with merchants who fund and bear the loan themselves. These merchants then offer credit to their customers, which Cashea guarantees if they default. It’s a hack that is as difficult to bootstrap in Venezuela as it would be in the U.S. or anywhere else, but a lack of credit availability in the system now drives a model which has been adopted by 5,000+ merchants across the country. It is growing exponentially, without major financing costs for Cashea, and has a delinquency rate below Affirm or Block’s Afterpay. Cashea charges merchants a fixed commission and has now begun to facilitate merchant’s receivables in a version of factoring offered on the Venezuelan Stock Exchange.

The company doesn’t need a massive balance sheet or financing facility, and it has completely obsoleted the need for traditional credit card rails, which are non-existent in the country anyway.

Here’s a great overview of the company from Fintech Leaders:

Cashea proves that necessity breeds innovation. A “clean slate” Venezuela will need a lot of it.

But clean slates can also be green pastures. They can help new market entrants build from scratch, innovate on business models, and leapfrog state-owned incumbents.

That is the opportunity we are excited by at Somos Internet, the fastest-growing ISP in Colombia. We build vertically integrated digital infrastructure to give customers better internet at a structurally lower cost. Our users love us. And we’re making plans for international expansion.

Until January 3rd, Somos hadn’t seriously entertained the idea of entering Venezuela. It was attractive, but there was too much risk.

Now, we are considering the opportunity. Venezuela’s capital city, Caracas, checks most of the boxes we look for when considering new markets.

Large market size and high population density: Caracas is a major city with a population the size of Chicago (three million) and a population density greater than San Francisco, which makes it a perfect market for a new entrant like us. We can find more potential customers for every km of fiber deployed.

High existing ARPUs and low market penetration: A lack of private competition has left Venezuelans with a raw deal. State-owned CANTV offers fiber to the home (FTTH) services starting from $25/month for 60 Mbps. This is extremely expensive in 2026. And the website claims to offer 1 Gbps at $150/month. Thanks in part to these expensive plans, the penetration of fixed internet is low and many households rely on cell data as their primary form of connectivity. We believe that giving people access to great internet increases economic opportunity. It is a virtuous cycle.

Early adopter culture: The general Venezuelan population is open to trying new alternatives (see Cashea’s adoption) because they are dissatisfied and lack legacy offerings.

Proximity to existing infrastructure: This isn’t a dealbreaker for Somos, but it is, on the margin, better to expand to adjacent geographies rather than jump to new geographies entirely. This would require new contracts with Tier I providers to maintain a unified network architecture.

Despite its attractive characteristics, we are not jumping into Venezuela yet. The situation is still too uncertain.

But given that we are actively evaluating the opportunity, our perspective may provide useful color on what businesses are looking for before committing to rebuild the region.

Read more

Happy Sunday. Earlier this week, X announced a $1 million article prize. I don’t normally write the kind of things that could win an X Article Contest - listy things, full of life lessons and advice. And then, wouldn’t you know it, my son Dev learned how to sell yesterday morning, and has he did, he dropped wisdom bombs for me to write down. We ended up with sixteen of them.

Now, they’re on X (go like, comment, and share - we need the $1 million, Dev has a world to build).

I really liked how it came out, so I wanted to share it with you all too. It’s kind of a co-written essay with a 5-year-old, who I hope becomes a more frequent contributor. I think I’m going to write more short things and share them in paid not boring world, so join us if you want the full spectrum of not boring, means to meaning.

Subscribe now

Let’s get to it.

16 Lessons on Selling (and Life) from My 5-Year-Old

This morning, my five-year-old son made his first two-dollar sale and dropped sixteen lessons on selling and life that are more practical than any of the slop you’ll find on LinkedIn.

I’ll share them with you, but first, I need to tell you about Dev, about his Donut Hats, and about his world.

One day when Dev was three, he told me that he wanted to build worlds.

Real ones. Big ones. Planets. Like, actual, physical planets.

“Then you’re going to have to study buddy.”

“What do I need to learn?”

Math, physics, engineering, business. No one’s ever built a world before, so you’re going to have to study really hard.

And then he… did.

He asked me for math problems, then harder ones, then harder ones. Kid does 90 minutes of Russian Math every Sunday and loves it.

Physics, he always liked. Gravity was one of his first words, and one of the first concepts he grokked. “Why’d the cup drop bud?” “Gravity.” We read a little bit of Richard Feynman’s lectures, and he stayed with me, but I figured that was probably taking it too far.

Engineering, he loved. Most kids do. Magnetiles in particular, huge structures. Every night, we read a couple of pages from The Way Things Work Now, which my dad always tried to read to me but which I turned down, because I didn’t have worlds to build with the knowledge.

Throughout, he’d pepper me with questions. What materials would we need to make the world? How would we get water to the world? How would we grow trees on the world? Some I could answer; a bunch we had to ask ChatGPT.

Two stories blew my mind in particular, though, logistical things, which are important things to get right if you actually want to build worlds.

One time, we were sitting by the pool on vacation, not talking about worlds at all, when he turned to my wife Puja and I and asked if we knew any companies that made houses. He figured he’d need houses if people were really going to live on this world, and somehow, that while he would be fully capable of building the world itself, there would probably be companies that were already quite good at homebuilding who he could pay to handle that aspect of the plan. He asked the same thing about umbrellas.

Another, we were talking about how to get people to and from the world. I’d met a company that was making Single Stage to Orbit rockets, I told him, and maybe they’d be good because they’d just take off from a normal runway and land on one too. He thought about it for a second and said, “No, we should probably use Starship, because they’ve actually flown before.”

The thing about his growing brain is that it’s always churning. Usually, he doesn’t mention the world for weeks, and then out of the blue, he’ll say something about it, or ask a question he’d clearly been chewing on for a while.

One big question, when you want to build a real, big, actual, physical world is where you’re going to get the money. We back-of-the-enveloped it and figured he’d need about a trillion dollars. I told him about investors. He eenie-meenie-minie-moed and landed on his three-year-old sister, Maya, as a lead investor. Implausible, for now, but the kid has vision and Maya’s pretty good herself, so not, in the opinion of one dad, impossible.

I thought the case was closed. It wasn’t. His brain kept churning.

So one night, earlier this week, I came home to find Dev and Puja at the kitchen table. He had a pencil in his hand and a piece of blue construction paper in front of him. They were making a business plan for his new company, Donut Hats.

I guess that afternoon, he took some Play-Doh, shaped it into a ring, taped it up with blue masking tape (kid loves tape), and realized he might be on to something. He put the first donut hat in a construction-paper envelope, put the envelope in a box, and taped that shut, too, for safe keeping. Then he got to work.

Puja and Dev were already pretty deep. They’d figured out a price ($20, but $10 for family members), estimated costs (surprisingly cheap if you count his child labor at $0), gross margins ($7.65 per at F&F prices), and were starting to work on a marketing plan. Kids would probably be the right target, he thought, but their parents had the money. He kind of just intuited this stuff.

When I asked him why he was starting a company, he basically recited Choose Good Quests and The Company as a Machine for Doing Stuff back at me.

“I want to sell a lot of Donut Hats to make money that we can use to build my world.”

Over the next few days, he made a total of five Donut Hats in different colors and tapes. My favorite is the Orange and Green in Clear Packing Tape, but if that’s not your style, there’s probably one for you, too.

That night, he rolled up the business plan (he loves rolling things up) and placed it on top of the Donut Hat Box, got into bed, and told me, “I’m so excited I finally get to run a company,” before drifting off to dream, I’m sure, about running a Donut Hat business.

Then came the hard part, genetically. I hate selling. I like writing plans. I like making things. I like marketing, but making a direct ask creeps me out. I told him that he would need to sell.

The next morning, he and Maya tried to sell from our stoop. Maya is not afraid of selling. She marched outside and started yelling “Get your Donut Hats! Ten BUCKS!” at the top of her lungs. But it’s winter, and it was 7:15am, and the only people out were harried ones scuttling to work. That wouldn’t do.

If we were going to sell to kids (via their parents), we would need to go to the playground, which we did this morning in a light 8:30am snow. We brought all five Donut Hats in a bag, and laid them out on a built-in table/chess board. There were only two other parent-kid combos there, and neither looked particularly in the mood to spend, so Dev half-heartedly and Maya full-throatedly yelled, “Get your Donut Hats! Ten BUCKS!” No one heard. It’s a big playground.

But then, a dad and his son came in. They headed to the soccer field and started kicking. I told Dev to go introduce himself and ask if they’d like to buy a hat. He said he was nervous. He didn’t want to go. And just then, providentially, the son kicked the ball over the fence. An opening. We grabbed it and threw it back over. They owed us one. I told him to go again, he asked me to come with him (I was as nervous as him, selling to strangers just minding their own business), we walked around the fence, and Devin, Donut Hat in hand, asked, “Would you like to buy a Donut Hat?”

The dad asked to take a look. He put it on his bald head. And he realized immediately that it wasn’t going to work. “How does the Donut Hat stay on my head? I’d imagine it would fall off if I moved. No thank you.”

HUGE. That was the first of what will be many, many No’s in Dev’s life, and he handled it gracefully. I told him it was awesome. We’d gotten our first customer feedback. I pulled out Apple Notes, titled it “Donut Hat feedback,” and told him we should write down all of the feedback we get so that we could go home and improve the product.

We wrote down:

1. Could fall off head.

While we were out on our soccer field sales call, the main playground started filling up, and playing there, right by our table, were a dad about my age and a son about Maya’s. Easy targets. Dev introduced himself, and asked, “Would you like a Donut Hat?” Father and son looked intrigued. They thought they’d just hit the Free Donut Hat Lottery. I whispered to Dev to tell them that he was selling them, which he did, and to which the dad responded, “How much?”

$10.

$10 is too expensive.

Dev came back at $5. The son, meanwhile, sensing a negotiation, deployed the Crazy Guy strategy. He threw out $6. Then he threw out $45. Then $15. Then $6 again. We waited, giving him the leash to walk himself right into an empty Piggy Bank.

But remember the market insight. The kids want the Donut Hats. The parents have all the money. And the dad wasn’t having it. While the son perused the goods, the dad negotiated for sport. Dev even offered our worst-made, pure Blue Tape Donut Hat at $3. But you could see in the dad’s eyes, he wasn’t going to buy. Finally, they walked away.

2. Too expensive.

MORE parents had come in, though, and a lot of them. One dad made the mistake of putting his daughter in the swing. He was a sitting duck. So Dev asked me to come with him to the swings.

“Hi I’m Devin, would you like to buy a Donut Hat?” He held out the goods, teasing.

“Oh that’s cool,” the dad, hooded by his sweatshirt but hatless, said. “But I’m not a big hat guy.”

Dad, write it down.

3. Not everyone’s a big hat guy.

But (and if you’re not a parent, you wouldn’t realize this), once your kid is in the swing, your kid is in the swing. You’re not going anywhere. You’re trapped. Dev just hung around while I pushed Maya on the swing. We weren’t going anywhere either.

Dev told him we had more colors. I threw in that it might look good under his hood. Dev kind of looked at the guy as only a little kid with big dreams can, and… he cracked.

“I don’t have $5, but would you do it for $2?”

Dev looked at me. I shrugged. It was his call.

“OK you can have it for $2.”

Dev let him pick his Donut Hat. Wouldn’t you know it, he picked the Blue Tape. Dev handed it over. The dad handed him two crumpled $1’s.

First sale! Dev was ecstatic.

And he was hooked on selling, whatever the price.

He was in luck. Social proof is a hell of a drug.

The mom pushing her daughter on the swing next to Maya’s saw the dad buy his daughter a Donut Hat and she wanted to buy hers one, too. She looked in her cell phone case / wallet, realized she had $1, and offered it to Dev. Take it or leave it, in nicer words.

He said yes. Two sales. Three dollars. We were HUMMING.

Something changed in Dev. He stopped being nervous and started to love the chase.

What about the dad in the swing on the other side? “Would you like to buy a Donut Hat?”

Sorry, I don’t have any cash.

4. No cash

Recall, however, that it was a big playground, and while we were selling, it was filling up even as the snow picked up. Dev went out into the big playground by himself, Donut Hat in hand, and started approaching people.

There were so many people spread over such a large playground that when Dev came back next, having sold zero more Donut Hats, instead of feedback, he started dictating sales tips.

5. We need a map of the playground to see where we can sell to people.

Got it. He went back out. More No’s. Whatever. A no is the first step on the way to yes. He came back.

6. Come when it’s not too cold.

Speaking of which, Maya was getting cold, and she wanted to go home.

And as we walked home, Maya and I on the sidewalk, Dev on air, he kept dictating, asking me to add notes to what he’d started calling “The Setback List.”

At the University of Virginia, Ian Stevenson has spent decades documenting cases of children seemingly inhabited by old souls, including:

Starting at age 2, James Leininger began having vivid nightmares about a plane crash, eventually providing specific details about being a WWII pilot named James Huston Jr. who flew off the USS Natoma Bay and was shot down over Iwo Jima. His parents, initially skeptical, verified the details through military records and located Huston’s surviving sister.

Shanti Devi was a 4-year-old in India in the 1930s when she began describing a previous life as a woman named Lugdi Devi who died in childbirth in a town she’d never visited. When researchers took her there, she reportedly recognized her “former husband” and navigated to her “previous home.”

At age 5, Muskogee, Oklahoman Ryan Hammons told his mom “I used to be somebody else.” He remembered being a Hollywood extra and talent agent, and when presented with a number of images, identified Marty Martyn in a still from the film Night After Night. Ryan remembered over fifty specific, later-confirmed details about Martyn’s life, and complained that he “Didn’t see why God would let you get to be 61 and then make you come back as a baby.” Martyn’s death certificate said he was 59 when he died, but when Stevenson’s successor, Jim Tucker, researched further, he found the death certificate was wrong. Martyn was actually born in 1903, making him 61 at death, just as Ryan claimed.

All of which is to say, maybe it shouldn’t be so surprising that Dev dropped so much wisdom in items seven through sixteen on The Setback List, but it still blew me away to hear so much wisdom out of the mouth of such a little man.

These are the lessons that Dev McCormick learned about sales on a Saturday morning on the playground in Brooklyn, dictated in random spurts over the next hour:

7. Always have a backup plan in case things don’t work.

8. Even if it doesn’t look fun, you should still do it.

9. You shouldn’t go if it kind of looks like a storm.

10. You need to remember everything people say because what if you don’t remember that you have a setback list?

11. People are nicer than you expect.

12. If someone looks like a bad guy you shouldn’t go to them.

13. You shouldn’t be nervous because it’s most likely they’ll say no if you’re nervous.

14. Maybe the importantest one: you definitely shouldn’t give up, because what if people say hehehe to you, that’s not a really good feeling.

15. Grownups shouldn’t come with you to help because it’s most likely they’ll buy it from only a kid.

16. The only way that people will buy it is if you’re being nice to them.

I don’t know man, I know I’m his dad, but that’s pretty good.

I think that one day this kid is actually going to build his world. $999,999,999,997 to go.

Postscript: Dev just woke up from a nap. I called him Mr. Sales Man. He said, “I love it when you call me Mr. Sales Man.” Hold on to your wallets.

Have a great weekend, and a long one if you’re reading this in the US.

Thanks for reading,

Packy