Martin grew up working class in Brooklyn. His parents are Albanian immigrants. His dad was a janitor. He started working on Wall Street as a teenager and built a pharmaceutical company from $2 million to $500 million.

I respect him tremendously for being an unabashed champion for capitalism. He said,

“I didn’t care about the party or anything like that. What I cared about was that ever since I studied economics it is a company’s right to charge whatever price it wants for its products and services. [...] The idea that a company chooses its price is holy to me. [...].

That’s what makes capitalism/economics in America great. The day that you say that you’re going to look at my portfolio and decide what the prices are going to be, you might as well move to another country. It’s over.”1

He might be one of the most disobedient people in the world. The entire media, Congress, and public turned against him. He did not change his position.

Drug Creationism

Most people think drugs should be free. Martin illustrates this with Vertex, which developed a cure for cystic fibrosis.

“Miracle product. Literally a cure for cystic fibrosis. Cystic fibrosis is no longer a medical problem for 90% of people with it. This company should have parades every year. But instead they get shit for it. One of my friends in high school died of cystic fibrosis at 19. These guys are doing - I use this term very carefully - God’s work.”2

Bernie Sanders asked “why is it so expensive?” They just cured a disease that killed 30,000 people.

I think this is Economic Creationism. The belief that goods and services can be legislated into existence. That you can wave a magic wand and they appear. Applied to drugs, education, or any essential service.

No value is directly created by the state. Drugs included. Only individuals create value. Entrepreneurs and employees who take enormous risk to research a drug and build a company that distributes it.

If you are excited about what Vertex did for cystic fibrosis, look into Sana Biotechnology (my writeup here). They are engineering pancreas cells to be immune-evasive, so they can be transplanted into type 1 diabetes patients to produce insulin. A one-time cure. I think the company is severely undervalued.

Healthcare/pharma industry

Most people do not understand that increasing the price of a drug can actually increase its access. Insurances pay for most of the cost. Higher profits incentivize more production.

Martin on his own drug:

“I 20x’d the price of a drug and the number of people on the drug went up. If you think the price of drugs and access to drugs are related - I hate to say this - but they’re not.”3

Unlimited marginal healthcare demand

The demand for marginal healthcare is virtually infinite. You would sacrifice marginal clothing or marginal food. You would not sacrifice marginal healthcare. We have gone from 5% to 20% of GDP on healthcare and he thinks it will go to 50%.

Why? Because once our basic needs are met we ultimately just want to live forever. I have argued before that wanting to be immortal is rational. Healthcare is not like other goods. It is the one thing demand never saturates for.

AI cannot “discover drugs”

Martin is friends with Vivek Ramaswamy. They ran rival companies with the same model: find underappreciated drug assets, buy them, and commercialize them. They both searched for underappreciated drugs in PubMed, a database of 38 million biomedical papers. Martin had a team of 18 PhDs reading papers. Vivek had 80.

“If you sit there long enough and you have the passion you can become a billionaire. It’s just a matter of sitting there and sifting and sifting and sifting.”4

Finding an undervalued drug is a creative task. It requires generating new conjectures about what a molecule could do, which drugs are undervalued, what disease they could treat, and what market they could serve.

It is not just ‘pattern matching’ across existing literature. LLMs can summarize papers. They cannot deeply understand a disease and conjecture a novel use for an existing compound. This is what I argue in Why LLMs Will Never Be Creative.

The creator/co-founder of Tailwind (a popular CSS library) recently rejected a code contribution claiming that they need to focus on recovering from the damage that AI has done to their business. He claimed he had to lay off 75% of his employees (~3 out of 4 people) because AI had a “brutal impact” on his business.

The real reason is that he (with the help of many open-source contributors) created a valuable open-source library but failed to monetize it and hired people he couldn’t pay.

What is disturbing is how much hate the comments received that criticized the Tailwind founder’s decision or that suggested improvements to his monetization strategy. It appears that most people reading this thread thought that these were inappropriate or disrespectful. Many of the critical comments received hundreds of downvotes (8-15 times the number of upvotes).

Businesses die every day. Errors happen and are corrected. The creator could simply stop working on this CSS library and build the next successful open source company. MongoDB is an open-source company with a $33.2B market cap. That is an opportunity anyone has thanks to capitalism.

2. Spotify

I recently spent some time in a Discord server dedicated to an electronic music community called Claws. Several people complained that Spotify’s take rate is roughly twice that of Apple Music. They argued that this is unfair, and that Spotify exploits artists.

I proposed that according to that claim, it would be “unfair” to continue to stream music from Spotify. In response, that person insulted me and justified their own continued use of Spotify by claiming that it is a monopoly. They then banned me from the Discord server.

As of 2026, Spotify has ~32% of the music streaming market. There are several other music streaming platforms the listeners and artists can use. Even if Spotify were a monopoly in the colloquial sense, it would still vie with other services for users’ attention.

Spotify has such a large share of the market because they provide a better product. Artists upload their music there voluntarily and benefit from the algorithms that drive traffic to their music.

3. DoorDash

A few weeks ago, a Reddit post by an alleged backend engineer at a large food delivery platform described the practices of his employer in disgust. He accused the platform of exploiting the drivers and assigning “Desperation Scores” to them. He claimed that this system offered the less profitable jobs (with a lower order value) to the more desperate drivers.

The post and the people sharing it make it sound like the drivers for that platform have no other work opportunities and hate their job. But that framing ignores revealed preference. If these jobs were truly as unbearable as described, far fewer people would choose them, including migrants who incur real costs and risks to do this work. Analytically, this treats drivers as passive subjects rather than participants making economic trade-offs.

A server in a restaurant can’t decide when they want to work and which customers they want to serve. The massive demand for food delivery drivers only started a decade ago and will likely go away with the rise of autonomous cars and drones. But the universe and the desires of humans are infinite. It is up to participants in the job market to find high-demand problems to solve.

Morally, portraying food delivery drivers as desperate victims is also insulting to those drivers who are proud of their job and providing a quality service to their clients.

Conclusion

I think all these phenomena spring from a suppression of criticism, ignorance of basic economic mechanisms, and the moral narratives that follow from those errors.

I wrote about this dynamic on my blog before, arguing that many incorrect moral judgments, from claiming that billionaires and super babies are bad to claiming that copying ideas is wrong, stem from the same explanatory failures.

1. The readers of the Github discussion dismiss criticism of the founder’s business model in order to preserve the victim narrative.

2. The aspiring indie artist rejects criticism of their false theory of platform economics, according to which Spotify is obliged to pay them and their friends more.

3. The DoorDash backend engineer relies on a mistaken model of labor markets, portraying drivers as exploited and denying that they have agency.

Thanks to Logan Chipkin, Dirk, and Tom Hyde for reading drafts of this. If you enjoyed this you might like:

Discoverability in Design

Product market fit is not just about building a product that solves a real problem. It only happens if the user discovers the feature that enables the solution. Discoverability describes how easily users can conjecture what functionality your application offers by looking at the interface. In his book The Design of Everyday Things Don Norman wrote:

Two of the most important characteristics of good design are discoverability and understanding. Discoverability: Is it possible to even figure out what actions are possible and where and how to perform them? Understanding: What does it all mean? How is the product supposed to be used? What do all the different controls and settings mean?

However, I don’t think these are separate processes. You can only discover something if you understand its function.

Apple example

The audio message feature in iMessage is notoriously bad. But did you know you can press-tap the playback button to change the playback speed? I did not until a few months ago. Here is Apple’s feature and WhatsApp’s in comparison.

Did you know you can activate the iOS Safari reader mode via one press-tap? This feature is not very discoverable. However, if you enable it using the settings, the icon lights up which helps you conjecture that you can press-tap on the icon to activate it directly.

Apparently you can also move apps between pages on the home screen and images between apps by holding them and swiping/navigating with another finger.

Spotify

Another favorite example is the Blend feature in Spotify. Initially, you had to go to the global search and type “Blend,” which would show an option called “Create Blend” without explaining what the feature even is. If you tapped on that, you could invite a friend.

It is now prominently displayed in the Create option in the bottom left of the navbar.

Tutorials and Onboarding UI

Tutorials and product tours are a sign of bad discoverability. A product whose features are discoverable does not need these. Robust solutions to users learning about more advanced features are checklists and empty states.

These make it easy to discover core features that drive user retention. This example of the Evernote onboarding shows both: An almost empty note that the user can edit and a checklist to encourage them to create reminders, save a web article, and sync between their devices:

Is that a way of “promoting features”? Yes and no. We are making hard to vary conjectures about which features the user will get a lot of value from and make them retain and making them easily discoverable (through the default UI state). The user should not have to search for a valuable piece of functionality. Let’s assume we have someone using a ChatGPT for finance product for the first time. They do not know what the exact capabilities of the product are. They heard the product offers some sort of library of automation. Let’s assume this is a feature which once it is used increases retention significantly. After signing up, our users start looking around in the interface for “automation”, “workflows”, “agents” or any related icons. It happens so that you can also just ask the main chatbot to create an automation, but the user doesn’t know that, and it’s not yet a familiar pattern to them. Until this is a familiar pattern to the large majority of our users, they’ll fail to find the valuable feature.

Simplicity and Familiar Patterns

Some might argue that adding discoverability clutters the UI and point to extremely simple products like ChatGPT or Google. However, in ChatGPT all the major features are visible in the sidebar and the large “+”-button in the search input.

ChatGPT initially shipped without Search, GPTs, or Company Knowledge. The chat UX pattern itself was unfamiliar. Starter prompts made the core capabilities of the model discoverable.

Discoverability and User Research

Designers, PMs, and engineers often fail to discover the real or underlying problems and the importance of these for their customers. This leads to bad prioritization. Companies can fight misprioritization by ensuring frequent and direct contact with customers. This advice is encapsulated in the Paul Graham-mantra “Talk to your users”. Talking to them will help you conjecture their problems and how they can best discover a feature.

Do This

The simplest solution to help you and your team understand your customers’ problems and build solutions in a way that makes them discoverable within your application is to:

1. Observe and interview your customers about their problems and friction points.

2. Watch a new user onboard to your software (without any tutorials). You can give them pointers if they get stuck.

You’ll likely be surprised by how much more detailed your user model becomes and how hard it is for them to use your product.

I think everyone working within the company should do at least one of these sessions every month.

At 24, Grady became the youngest person Sequoia ever hired. His investments include OpenAI ($300bn), ServiceNow ($177bn), Snowflake ($86bn), and Zoom ($25bn), among others including HubSpot, Notion, Qualtrics, Okta, OpenEvidence, and Harvey.

Relatedly, I previously created a podcast titled The Almanack of Keith Rabois.

Desperation & Irrelevance

Grady grew up in Gillette, Wyoming—a coal mining town—desperate to find out what else was out there.

In high school, his greatest fear was boredom. Today, it’s irrelevance.

Different people have different answers to the question, what is the meaning of life? I think for some people, the greatest happiness comes from some feeling of productivity or some feeling that you’ve contributed something to the world around you. When I say irrelevance is my greatest fear, it would be the point at which I am not a good husband, not a good father, not a good partner, not contributing to the world around me.¹

This thinking pervades everything at Sequoia.

The default is if we don’t go out there and earn it, tomorrow we are irrelevant.¹

Doug Leone became senior steward in 2012. He discovered only he was safe—nobody could vote him off. He changed the legal documents so he could be fired like anyone else.

As soon as you’re safe, you start to get a little bit complacent. And as soon as you get complacent, you start to become mediocrity.²

Look for What’s Right

When Grady joined the partnership he systematically looked for flaws in investments and in people. Roelof Botha told him: the hard thing is coming up with the reason to invest, knowing there are so many things broken. The job is to figure out what is so overwhelmingly great or tantalizingly promising that it’s worth dealing with all the stuff that’s broken.³

When I meet a new person, I try to think, what’s special about this person? What can we learn from this person? I think it’s made me a better investor and hopefully a better and certainly a happier person.³

Understanding People

The biggest mistake in sales is telling people how awesome you are. They don’t care how awesome you are. They want to know how awesome they have a chance to become.²

The thing we really want to do is understand who are you, what do you want to become, and what is it you want to build? If we can understand those things, and we can feed that back to you to show you that we understand those things, and we are interested in those things, and we want to be a part of those things—the main thing is not our resources. The main thing is your vision and your dream.²

Doug Leone’s method: 30-40 minutes of rapid-fire questions. A perfect 2-minute summary capturing what the founder hopes to build, acknowledging every flaw in the business, gently suggesting how Sequoia could help, and reaffirming the vision. Ninety-five percent of the time, the founder would be desperate to work with them.⁴

The most powerful thing you can do is to make a founder feel like they’re seen. You understand them and try to validate their ambition through the way that you describe their business.²

On negotiation, Doug was unbelievably transparent. Grady recounts a typical negotiation: “You’re going to want to pay X, we’re going to want to pay one half X. Why don’t we just call it 0.75X and call it a day? The whole thing would take 30 seconds.”²

The simpler you can keep things, the more straightforward you can be, the more transparent you can be, the more people are going to trust you, and the more people trust you, the easier life becomes.²

Grady applies this to all communication. The objective isn’t to impress. It’s understanding.²

My objective is for them to understand. The thing you’re optimizing for is clarity and understanding, not trying to impress upon people how smart you are.²

When assessing an opportunity Grady thinks about two variables:

The market determines how big a company can get, the founder determines how big the company will get. Market in some ways establishes the ceiling on an opportunity. But you could have the biggest market in the world, if the founder is just not that good, you’re not going to get anywhere. I will take a market of modest size being attacked by a spectacular founder over a market of gigantic size being attacked by an okay founder.¹

Personal Growth

Grady creates scaffolding around himself to maintain pressure.

I have my personal long-term plans and then my personal annual OKRs, which cascade down into my quarterly OKRs, which cascade down into what I’m doing this week, this day, this minute.²

When asked to rate Sequoia on winning deals, Grady answered: “We don’t give 10 out of 10s.” Even for an area where Sequoia excels, he rated them a nine.¹

He joined at 24 as the youngest person they’d ever hired. He recounts:

Pretty much every single day on the drive home, I was berating myself for not having done enough that day. When we did our annual reviews, I’d probably spend 48 hours on my self-review. It was almost like a self-therapy session where I was just constantly tearing myself apart for not having done more and not having lived up to the name Sequoia better that year.²

Infinite work capacity kept him in good standing. A heart of gold; true dedication to doing the right thing.⁴

Michael Moritz told Max Rhodes, founder of FAIR, what separates legendary companies: relentless application of force. Grady agreed.²

1. Pat Grady, “Sequoia Partner on Investing Lessons from Doug Leone, Roelof Botha and Alfred Lin,” 20VC Podcast Episode 1174

2. Pat Grady, “Relentless Application of Force,” Invest Like the Best Podcast Episode 378

3. Pat Grady, Loom recording on lessons from Roelof Botha

4. Pat Grady, “Sequoia Capital’s Pat Grady on Building Enduring Companies, the Sequoia Values and AI’s Next Wave”

Most people hold fundamentally wrong ideas about morality. This includes thinking that copying business ideas or having CRISPR-edited super babies is immoral, that death is good, that the existence of billionaires is wrong, and that not helping others is immoral. They all spring from an avoidance of conflict: the conflict of ideas, the conflict of accepting not to exist, the conflict of envy, etc. This “morality” becomes destructive because it’s scared of the turning forces of evolution by evading conflict and adversity of any kind. It’s suffocating.

Morality is best explained by epistemology (the philosophy of knowledge), not religion or intuition. With this lens morality is the knowledge about what to want, and what to strive for1.

The most fundamental tenet of morality is to not remove the means of error-correction2. Since error-correction is the only way our knowledge can improve.

If society hinders a scientist from inventing and distributing a cure for cancer, that is deeply immoral. Many regulations that restrict the freedom of people and their ideas are immoral.

Copying someone's business

The Samwer brothers famously copied Airbnb and other companies, but these companies provided the solution to people in different geographies or demographics, improving access to the solution.

Their Airbnb clone, Wimdu, largely failed because it was only a surface-level copy that didn’t innovate on any aspect of the business. It incentivized Airbnb to innovate on better host support, internationalization, trust infrastructure, and regulatory compliance.

Opening a lemonade stand two blocks from an existing one incentivizes both lemonade stand operators to improve their lemonade. Competition leads to innovation.

Now, why do most people believe that copying someone’s business is immoral?

I think the main reason is that people think ideas can be “stolen”. That is wrong. Ideas are non-rivalrous3. And everyone should be incentivized to reproduce them and correct their errors. Problems are solvable4, and there is an infinity of problems (people always want more). IP law is another way to incentivize people to innovate. Unfortunately, large companies like Amazon (with hundreds of people in their internal legal department working on IP law) exploit this system to prevent competition.

But isn't competition for losers?

Peter Thiel famously proclaimed this in his book Zero to One. In a demand-constrained market, this is true. Building another AI headshot app wouldn’t be a great idea if the demand for AI headshots was shrinking rapidly. It is not.

Even Founders Fund (Thiel’s venture firm) invested in companies with strong competition:

• Ramp: Launched two years after Brex; both grew quickly as financial operations digitalized.

• Spotify: Entered a crowded market (iTunes, Pandora, Rhapsody) just as music streaming took off.

• Rippling: Entered HR/payroll after Gusto, ADP, and Paychex; succeeded by bundling HR, IT, and finance as businesses moved to the cloud.

• Postmates: Was started after Grubhub and Seamless, but grew fast as on-demand delivery became a habit.

• Icon.com: Was started after there were already countless AI ad generator platforms. As more people consume short-form video content and realistic AI-image and video generation becomes possible demand for this kind of software is exploding.

It would be stupid to claim these companies are immoral because they copied another business. Here are some other moral mistakes I have considered lately. They range from economics to biology to altruism.

Altruism

Helping can be immoral if it prevents people from learning to solve their own problems. It can also be immoral if the invested resources could have led to a greater error correction. Socialism, with its underlying philosophy of altruism, is the embodiment of this error. For example, the US government taxes the 343 million Americans ~$1.2 trillion per year (that’s ~$3,580 per person on avg.) to fund a public education system that is stuck in the 1800s. These people could have used the resources the state took from them to buy education services from private companies that have a clear profit incentive to improve their service. To quote my previous post, “They don’t care if the students hate school” nor if they end up in student debt.

Being immortal

Death is not the only reason that life is “precious”. There are better reasons.

The idea that the shortness of life is what gives it its “meaning” is wrong. Meaning is not zero-sum. I.e., your life doesn’t have a fixed amount of meaning that is reduced if you live longer. The opposite is true, a longer life leads to more meaningful experiences. Life is more like water when you’re thirsty than a rare diamond; its value comes from how much you can experience, not from being limited in quantity.

Life is not its own meaning. Meaning comes from the things you do and create. If your life is unbounded, so is the meaning you can create within it.

People wear seatbelts. Nobody says, “I want to get hit by a car, because my life will have more meaning if it ends sooner.” There’s nothing inherently meaningful about limitation or scarcity by itself. If that were true, a two-day life would be more meaningful than a 70-year life.

Some argue that meaningfulness of life is binary. But if knowing the exact day of your death feels terrifying, that shows that it’s not simply mortality that gives meaning, but the unpredictable and exciting adventure of living itself.

Even Christianity calls death an enemy from humanity’s fall, promising its end: “The last enemy to be destroyed is death” (1 Corinthians 15:26). In paradise, “death shall be no more” (Revelation 21:4). It sees death as a problem to be solved, not a source of meaning.

Ultimately I think arguments that claim death is moral originated to cope with the fear of death. The thought that our existence suddenly stops. All the laughter, parties, friends, interesting conversation, sex, beautiful landscapes and skylines are suddenly taken from us, for eternity.

Nick Bostrom in his essay The Fable of the Dragon‑Tyrant calls this "deathist" ideology—the passive acceptance of what should be humanity's greatest enemy. My friend Arjun criticized other common arguments against immortality in his blog post.

Involuntary death isn’t a cosmic necessity; it’s a biological error, possibly the greatest error of all. And it’s one we may soon be able to fix. Just imagine the inventions we never experienced and how many more lives would flourish today if people like Euclid, Newton, Tesla, Marie Curie, or Alan Turing would still be alive.

CRISPER-edited Super Babies

Some people don’t like the idea of us playing God. But that is what we are now. We are as gods and might as well get good at it5.

Others claim that gene-editing super babies (to be free from disease or just smarter) would be an innovation that only the rich could afford, thereby creating a segregated society. The 1997 movie Gattaca is a dramatization of this claim.

That logic is wrong because it impedes the spread of an innovation (and error correction). The assumption that non-gene-improved people will be relegated to menial jobs is also wrong because there are many wealthy people with fun jobs who are low in IQ or disabled.

History shows that innovation disperses. Initially only the rich could afford better nutrition or computers. Should we have waited before allowing people to buy a computer until they are cheap enough for eight billion people to have their own? No.

Opposing gene-editing is like opposing antibiotics because not everyone could afford penicillin in 1928.

The technology is here. Using embryo selection or gene editing to have a super baby is a choice about your body and procreation. Some people don't think you should have this choice. But it's not their choice to make. It's yours6.

Billionaires

Some people claim billionaires shouldn’t exist. That is wrong. I'd even go so far to say that not wanting to be a billionaire (or trillionaire) is wrong. Ambition is a consequence of optimism. In his Davos 2024 speech Javier Milei said:

Let no one tell you that your ambition is immoral

We are like billionaires to people living 5000 years ago. If some of these people did not desire immense wealth, we’d probably still live in mud huts now. We're like iron age peasants to the people that will live 2000 years from now.

Like every other destruction of optimism, whether in a whole civilization or in a single individual, these must have been unspeakable catastrophes for those who had dared to expect progress. But we should feel more than sympathy for those people. We should take it personally. For if any of those earlier experiments in optimism had succeeded, our species would be exploring the stars by now, and you and I would be immortal.7

I think the disgust of the existence of billionaires likely stems from envy. But that response is wrong. What matters is not the relative, but the absolute wealth and well-being of people. In the future, inequality will become ever greater. Some people will own entire galaxies, while others will only own a continent on a single planet. Everyone will be immensely more wealthy.

Thanks for reading this. I’ll now continue playing the infinite game of capitalism.

Thanks to my friends for feedback on this essay: Tom Hyde, Milan Scharfe, Ben Chugg, Vaden Masrani, Maxime Desalle, Saketh, Jesse Ray Nichols, David Taurat, Logan Chipkin, Brian Chau, Remus

[Note: I wrote this in Apr 2023 but never published it, so some numbers in here might be outdated]

The Thiel Fellowship succeeded in creating an extremely high status credential/club membership. And it succeeded in helping create about $220 billion in value.

While it did create an alternative to college for a select few, it didn’t create a sustainable alternative to a broken education system.

But you could make it an incredibly profitable model. By simply taking 2% or so of each fellow’s (future) companies.

Multiples

In total, the fellowship granted $22.8 million in $100k fellowships. It helped create about 220 billion in venture returns. Let’s assume another 400k in total every year in operating costs of the fellowship. Since its inception in 2010, that’s $5.2 million. So the total cost might be around $28 million.

A 2% stake in the $220 billion outcome would be worth $4.4 billion. This represents a return multiple of approximately 157x from the initial $28 million investment.

In comparison, top venture capital firms like Sequoia have an average multiple of 4-8x. The Thiel Fellowship for equity model could generate returns that are approximately 19.6 to 39.2 times greater than those generated by the best VC funds.

Even if we assume a higher price (equity payment) discouraged many from applying to the fellowship, and it returned only $110 billion, that's still a return of $2.2 billion, i.e., a return multiple of approximately 78.6x from the initial $28 million investment.

However, the fellowship also served as a deal flow pipeline for Thiel and Founders Fund, thus allowing them to capture a respectable part of the mentioned $4.4B return.

Status signal dilution

Now you might think: But this will reduce the value of the signal? If fellows paid equity for a Thiel credential it becomes less valuable. That is true, but it’s not the true limitation of the “Thiel Fellowship for equity”-model. Stanford and Harvard students pay exorbitant amounts (avg of $200-350k depending on financial aid) for every semester and the status signal is high regardless. (The credential probably retains a high status value despite the high cost exactly because the price is opaque for others).

The true limitation is scaling up the batch sizes. This is what happened to YC. I couldn’t find precise “reputation metrics” on this but a lot of founders and VCs will confirm that YC isn’t as attractive anymore because of the ubiquity of the signal.

The best way to scale up the batch size is to increase the perceived value of the deal. Stanford’s high status is not only due to its exclusivity but also due to the program quality. Stanford students learn from the best of the best and live in an environment that resembles a resort.

Potential Value proposition

Since we’re now talking about a venture capital/school model we need to re-evaluate our value proposition. The service needs to be extremely attractive which allows us to scale up the batches. We want this to be a model that can revolutionize the education system.

We also need to make sure joining the program has a high success rate in making its users into geniuses that build massive companies. Otherwise, this won't be an attractive offering. How do we do that? By offering the world's best education. The company that’s running this program will be an institution that is more incentivized to do so than any other existing educational institution.

It’s trivial to deliver something significantly better than the current education system. Currently, students spend a big portion of their lives following a mandated curriculum cramming facts for tests which they then forget about.

The best education is 1:1 tutoring combined with project based, curiosity driven learning. Most of the people we call geniuses received 1:1 (aristocratic) tutoring from an early age. Famous examples include Darwin, John Stuart Mill, Hannah Arendt, Einstein, Jon van Neumann, Bill Gates, and Bertrand Russel1. Bloom’s 2-sigma problem is often quoted as proof of this effect.

The earlier this program starts, the better for our program’s success rate. However, that also increases the operating cost and our return cycles significantly.

Other objections

Incentives Ethics

Some might still argue that this model is “profiting from young entrepreneurs ideas and efforts” (describing it as something negative). Yes. That is the definition of an efficient education system. In the current system the tutors and school administrators have no incentive to innovate and improve the education of the students. They don’t care if the students hate school.

This is the same reason non-profit’s or NGO’s incentive structures are inefficient. The user of the service doesn’t pay for the product directly. A third party pays for it. This is also why I would not trust news sources that are donation or government funded, as they do not have an incentive to grow their customer base to the widest possible audience and deliver the best product.

Even private schools are misaligned in some ways. They get paid by their customers for a credential that demonstrates that they were able to pass some test, and personality traits such as conscientiousness and conformity. The incentive of a conventional college is to persuade employers that you are a conscientious and conformity seeking worker.

The Thiel Fellowship for equity model skips this step completely. Its only incentive for its students is to generate massive amounts of value for society.

Talent Pool limitation

Another possible objection is that the talent pool for this kind of program is too limited. This assumes that building an incredibly impressive company is something magical that you can’t learn or that requires an incredibly high IQ. That is not true. Building a company is not significantly more difficult than working as senior ranking executive or programmer.

Indentured servitude

The association of a program like this with indentured servitude appears to be the biggest obstacle to making the long-term strategy version of this a mainstream educational model.

Indentured servitude is a historical labor system where workers (mostly immigrants) were bound to work for a specific period of time in exchange for a passage, room, or some other service.

There are a number of important differences with this historical phenomenon: There is no upfront cost in the form of physical labor. Our version is voluntary, fairly compensated, and the value exchange happens asynchronously, in the form of a work artifact (a valuable company) that is created by the student after we provide our educational and incubation service.

The association is likely less strong in Europe where a version of this model is already practiced as apprenticeships. In Switzerland, for example, 70% of teens are apprentices, giving the country a steady stream of trained, experienced workers (including areas such as accounting, healthcare, aerospace, and software engineering).

Others who wrote about this

In modern times, the first notable mention of the concept of Income Share Agreements (ISAs) was by Nobel-prize winning economist Milton Friedman in his 1955 essay The Role of Government in Education. Last year The New Yorker published a piece “Is Selling Shares in Yourself the Way of the Future?” about two brothers that sold shares in their personal holding company2. The Residency provides, housing, food, and cash for 3 months in exchange for 2% equity in resident's startups3. Sam Altman also tweeted about this last year4.

Sam Lessin (co-founder and GP of Slow Ventures) also talked about the concept in an interview with Brian Requarth.

Questions I still have

• How important is it to have an initial status signal / association with some high status individual such as Peter Thiel or Elon Musk?

• How can you effectively sell the program to 10 or 16-year-olds? The earlier you start with the program the more you have to persuade the parents.

• What legal edge cases could arise? If a student disputes the contract later, as contracts with minors are often voidable. However, contracts deemed beneficial, such as those involving apprenticeships or services, may still be enforceable if their terms are fair and reasonable.

• How profitable was the fellowship for Thiel and founders fund? If anyone knows how much Thiel put into Luminar, Figma, DoNotPay, Scale AI, Oyo Rooms, etc. that would be interesting.

1. The Beginning of Infinity: Explanations That Transform the World (David Deutsch)
   My favorite book by far. It’s also Naval’s and Sam Altman's favorite book. Most of my blog posts touch on some of Deutsch’s ideas. Some favorites include Disobedience is Fundamental to Intelligence and Critical Rationalism for Founders and Managers.

2. Zero to One: Notes on Startups, or How to Build the Future (Peter Thiel)
   I reference Thiel's work on my blog in my post about startup ideas, LLMs not being creative, Transaction Cost, and others.

3. Atlas Shrugged (Ayn Rand)
   I found out about this book because it was one of the most recommended books in Tools of Titans by Tim Ferriss (where he interviews successful people and asks them the same questions, including what books influenced them the most). I was more of a centrist or even left-leaning before reading this. Taught me a lot about ambition and obsession.

4. Meditations (Marcus Aurelius)

   Read this about 5 times. Notes (German).

5. Philosophy and the Real World: An Introduction to Karl Popper (Brian Magee)
   Great overview of Karl Popper’s work. To be honest, I found it hard to get through some of Popper’s books because of his academic writing style, but Magee’s book was a deeply enjoyable read.

6. Peak Human: What We Can Learn From History’s Greatest Civilizations (Johan Norberg)

   16.5 mins into the audiobook Norberg quotes David Deutsch (on cultural relativism). Great read to understand why civilizations thrive and decline. Explains: Ancient Athens, The Roman Empire, The Abbasid Caliphate, Song Dynasty China, Renaissance Italy, The Dutch Republic, and The Anglosphere.

7. No Room for Small Dreams: Courage, Imagination and the Making of Modern Israel (Shimon Peres)
   Great portrait of Israeli values and perseverance. Extremely moving. Read this after reading:

8. Startup Nation: The Story of Israel's Economic Miracle (Dan Senor and Saul Singer)
   An incredible success story of capitalism and Western values. Israel, at the time of this writing, is only 77 years old and is still the 25th largest economy in the world by nominal GDP. Notes.

9. How Innovation Works: Serendipity, Energy and the Saving of Time (Matt Ridley)
   Made me a lot more excited about private sector innovation/research. Also, check out Naval’s interviews with Ridley: Part 1, Part 2.

10. The Founders: The Story of Paypal and the Entrepreneurs Who Shaped Silicon Valley (Jimmy Soni)
    I’m currently re-reading this after having read:

11. The PayPal Wars: Battles with Ebay, the Media, the Mafia, and the Rest of Planet Earth (Eric M. Jackson)
    Very detailed account of stories from an early product team member that reported to David Sachs.

12. The Fabric of Reality: The Science of Parallel Universes--and Its Implications (David Deutsch)
    Made me realize that human understanding is unbounded (and not primarily based on facts). Notes.

13. Several Short Sentences About Writing (Verlyn Klinkenborg)
    Best book on good writing. Aligns well with my intuition about the value of simplicity.

14. Elon Musk (Walter Isaacson)
    Great refresher from reading the Ashlee Vance bio about 10 years ago. Elon’s focus on simplifying processes is incredible.

15. The Power Law: Venture Capital and the Making of the New Future (Sebastian Mallaby)
    Best book on the history of Venture Capital. I also highly recommend reading Tom Perkins’ biography, Valley Boy, Michael Moritz’s portrait of Don Valentine (Founder of Sequoia), and Eboys.

16. The Upstarts: How Uber, Airbnb, and the Killer Companies of the New Silicon Valley Are Changing the World (Brad Stone)
    Made me appreciate network effects (and hacking them) a lot more.

17. The Nvidia Way: Jensen Huang and the Making of a Tech Giant (Tae Kim)
    Companies constantly need to reinvent themselves.

18. Rainbows End (Vernor Vinge)
    My favorite hard sci-fi book. Notes. If you like hard sci-fi too, I wrote a blog titled The 7 Best Optimistic Hard Sci-Fi Short Stories.

19. An Elegant Puzzle: Systems of Engineering Management (Will Larson)
    Best book on eng management. Great read for anyone leading a team of (software) engineers.

20. Scientific Freedom (Donald W. Braben)
    Helped me realize the power of curiosity and open-ended science. Quoted this in my blog post Open-endedness in Society.

21. Winners Dream: A Journey from Corner Store to Corner Office (Bill McDermott)

    McDermott is probably in the top 10 storytellers on this planet. His journey from a corner store in Long Island to CEO of ServiceNow is exceptional. If you’re interested in learning about ServiceNow’s legendary investor, check out my post about Pat Grady.

22. Basic Economics: A Citizen's Guide to the Economy (Thomas Sowell)
    Must-read for anyone who wants to deeply understand how capitalism works. Notes.

23. How to Heal Your Metabolism: Stop blaming aging for your slowing metabolism (Kate Deering)
    Best book about nutrition and health. I wrote a corresponding blog post titled The Bioenergetic Explanation of Health.

24. Why We Die: The New Science of Aging and the Quest for Immortality Hardcover (Venki Ramakrishnan)
    I read about four books about the evolution of death and this was my favorite one neatly explaining the most prominent explanations. It also explains why death is the price we pay for sex.

25. Bigger Leaner Stronger: The Simple Science of Building the Ultimate Male Body (Michael Matthews)
    This was one of the first 2 books that showed up when I searched strength training on Audible. It includes several great workout plans and the theory behind them. Dispels the popular myth that you can't lose fat and gain muscle at the same time.

26. The Information: A History, a Theory, a Flood (James Gleick)
    Great introduction to computing and information theory.

27. The Parasitic Mind (Gad Saad)

    Read this in Feb 2021. First time reading about virtue signaling. Elon Musk is a big Gad Saad fan, too. I also wrote about virtue signaling in Criticism and The Ascent of Man.

28. The Myth of Artificial Intelligence: Why Computers Can’t Think the Way We Do (Erik J. Larson)
    I interviewed Erik here. This is the only book I’ve ever seen that Peter Thiel endorsed with a blurb.

29. Linked: How Everything Is Connected to Everything Else and What It Means for Business, Science, and Everyday Life (Albert-László Barabási)
    Helped me develop an appreciation for graph theory and network science. Made me realize networks of knowledge are modular, scale-free networks. Notes.

30. Influence: Science and Practice (Robert Cialdini)

31. Becoming Steve Jobs (Brent Schlender and Rick Tetzeli)

    I prefer this version over the Isaacson one.

32. Science Fictions: Exposing Fraud, Bias, Negligence and Hype in Science (Stuart Ritchie)
    Explained how flawed most of science is.

33. Why Beautiful People Have More Daughters: From Dating, Shopping, and Praying to Going to War and Becoming a Billionaire — Two Evolutionary Psychologists Explain Why We Do What We Do (Alan S. Miller, Satoshi Kanazawa)
    Fascinating evolutionary biology and psychology theories.

34. Scale: The Universal Laws of Growth, Innovation, Sustainability, and the Pace of Life in Organisms, Cities, Economies, and Companies (Geoffrey Brian West)
    Interesting analogies between types of organisms and organizations. Notes.

35. Living With Complexity (Donald A. Norman)
    Phenomenal book on simplicity and design.

36. Accelerando (Charles Stross)
    Actually stopped reading this about 70% into the book as it felt too unrealistic. But the first part is very enthralling.

37. Stories of Your Life and Others (Ted Chiang)
    Very thought-provoking. I wrote about two of my favorite stories here.

38. Complexity: A Guided Tour (Melanie Mitchell)

    Another great book on complexity theory. Explains how to measure complexity. Notes.

39. The Art of Thinking Clearly (Rolf Dobelli)
    First self-development, non-fiction book I read when I was ~15. Taught me, among others, that reading the news is a waste of time.

40. Actionable Gamification: Beyond Points, Badges, and Leaderboards (Yu-kai Chou)
    Best book on gamification, a very underappreciated area of ux design.

41. The Science of Science (Albert-László Barabási, Dashun Wang)

    Great book for understanding science as an industry. Notes.

42. Being Logical: A Guide to Good Thinking (D. Q. McInerny)
    Logic is underrated.

If you've read this far, I would like to again ask you: Is there any book that should be part of this list/I should prioritize? Please comment the name of the book below.

Leave a comment

Also, let’s connect on goodreads: https://www.goodreads.com/moritzw

Books I had on this list but removed:

• Eboys: The First Inside Account of Venture Capitalists at Work (Randall E. Stross).
  Legendary stories of Benchmark.

• Who: The A Method for Hiring (Geoff Smart)
  Best book on hiring. Notes. I also recommend this guide by the Gem.com founder. Moved this out of the list as it is not super relevant if you’re not recruiting a ton of people.

• The Rational Optimist: How Prosperity Evolves (Matt Ridley)
  I would have put this in the other list, but it didn’t fit.

• Competing in the Age of AI: Strategy and Leadership When Algorithms and Networks Run the World (Marco Iansiti)

  Good explanation about data vs. network effects. Written pre-GPT3 but still very relevant..

• Reinventing Discovery: The New Era of Networked Science (Michael Nielsen)

• Sapiens: A Brief History of Humankind (Yuval Noah Harari)
  Good book on the history of humanity. Great storytelling. However, Harari is a pessimist.

• The Meme Machine (Susan Blackmore)
  Good, but Deutsch’s writing on this is a lot better.

• AI 2041: Ten Visions for Our Future (Kai-Fu Lee)

  Collections of short stories by the author of “AI Superpowers: China, Silicon Valley, and the New World Order”. Love this type of sci-fi. I mentioned some of these in my post here.

If you like reading, you might also like this post:

He explains how conjecture and criticism are analogous to variation and selection in evolution. This makes the growth of knowledge an evolutionary process. Creativity is the most powerful piece of software in the entire universe, as it allows us to create knowledge, solve problems, and transform entire solar systems1.

Try asking AGI-soon-proponents: Would you have to prompt it? LLMs, lacking agency, disobedience, and creativity cannot rival or threaten humans. Thus fearing them is irrational. While LLMs can be misused to persuade, defraud, hack, or design bioweapons, they also enable protection. For instance, an LLM might suggest a DDoS attack, but human creativity, augmented by LLMs, can devise novel defenses like adaptive firewalls.

Claiming that LLMs will somehow gain creativity is prophecy. A prediction draws conclusions about future events that follow from good explanations. A prophecy is something that purports to know what is not yet knowable, such as the invention of a creative program. Moore’s law is a prediction based on an explanation of engineering density scaling principles (photolithography advancements, material science innovations, and power-efficiency trends). The claim that “more AI researchers working on improving LLMs will lead to creativity” is not a good explanation. AI scaling laws predict performance (e.g., loss) from larger models, more data, and compute, not creativity, while test-time compute scaling enhances reasoning during inference but still lacks creative conjecture.

AlphaEvolve, a recent Google advancement, optimizes algorithms like matrix multiplication within a defined search space2 using clear evaluation metrics, unable to independently conjecture new problems or solutions as humans do.

Expecting to create a program that is creative without first understanding in detail how creativity works is like expecting skyscrapers to learn to fly if we build them tall enough. The inventors of the steam engine Savery and Newcomen, while relying on experimentation and mechanical intuition, had rudimentary and intuitive knowledge of thermodynamics. Similarly, the Wright brothers understood aerodynamics before inventing airplanes.

Machine learning fundamentally works by applying inductive probability to the occurrence of tokens in its training corpus3. But induction can’t create new knowledge4. For centuries Europeans thought all swans were white. In 1697, Dutch explorers discovered black swans in Australia. Inducing that all swans are white didn’t create new knowledge. Similarly, recently LLMs failed to generate a picture of a wine glass that is full to the brim. The LLM fundamentally lacks an understanding of these concepts and can’t conjecture how a full-to-the-brim wine glass would look like without having seen hundreds of labelled examples. In the below illustration the full-to-the-brim wine glass and reverse-traffic lights are black swans in the LLM’s training data.

This video explains a great watches prompt example:

robinsonerhardt

A post shared by @robinsonerhardt

The prompt is:

Draw a group of watches, showing 3 minutes after 12.

I also came up with this “Double Mirrored hallway”-example. The prompt is:

A man stands in a mirrored hallway holding a soccer ball. Multiple reflections create an infinity mirror effect, showing him in different poses with the ball stretching endlessly. The room has bright artificial lighting. Camera angle is from above looking down.

“LLM-intelligence” is different from human intelligence. As Peter Thiel said

Computers are tools, not rivals. [...] As computers become more and more powerful, they won’t be substitutes for humans: they’ll be complements.

Augmentation is thus more effective than a complete LLM-replacement for tasks that require creativity. This is the reason Cursor (AI IDE) is overall a more successful product than Devin (AI coding agent). If you have ever used Cursor’s agent mode for difficult (debugging) problems without intervening or guiding it, you will notice that it tends to get stuck. This is because it cannot creatively conjecture solutions5.

Thanks to Tom Walczak, James Bergust, Vaden Masrani, Logan Chipkin, Simon Berens, Laser Nite, and others for giving feedback on drafts of this.

This post is partly about what I learned about co-founders and partly about my own search.

My search

I realized some of the best people to ask as part of my search are the subscribers to this blog and their friends. I enjoy working with and spending time with people who are intellectually curious. They should have the ambition to build a massive, fast-growing, and venture-backed company solving an important fast-growing problem. Ideally, they're even more technical than I am.

While I can get excited about many ideas that have the potential to grow into a generational company, I like problems that involve the creation or dissemination of knowledge. Some areas I have thought about recently are ai executive coaching, compliance ai for aerospace and aviation, ai roll-ups (LPOs and SEO agencies), and eval labeling/expert marketplaces.

If someone comes to mind that is also exploring or planning to start something, please forward them this blog post or blurb:

I (Moritz Wallawitsch) previously co-founded BenchFlow.ai (backed by Jeff Dean) and RemNote.com (GC-backed, ~2M ARR). Exploring several ideas including BPO ai roll-ups and ai exec coaching, but open to others. Interested in connecting with other technical founders in similar situations. (moritzw.com)

Share

What I realized and learned

Early-stage VCs are right to prioritize the co-founders and their relationship over everything else because a great co-founder team is able to pivot around until they find the right problem and reach PMF. 70% of successful startups pivoted at least once before finding the right business model. I wrote about this ideation process on my blog before. The Cursor/Anysphere team for example, started in text to CAD, moved to an ai email client idea before they started working on Cursor.

A great co-founder is someone you would have met anyway. I.e., you share existing friends with them. You should also enjoy spending time with their friends and partner. You can test this by going to the same event or hosting a dinner or poker night.

Another interesting criterion I re-discovered is the airplane seat check. I.e., would you enjoy sitting next to them on a long plane ride. If in a meeting, do you naturally go over the allotted time?

In my experience, it is a lot more fun, rewarding, and productive to come up with the idea together (or independently come up with the same idea) that you work on. Analogously, parents feel more invested in children that they helped conceive compared to those from a previous relationship.

The majority of the companies I admire had multiple co-founders. Examples include: Palantir, Anduril, early Apple and PayPal, Flexport, Faire, Scale AI, Cursor, Mercor, Ramp, Stripe, and Contextual AI.

I think a match in skills is important but underappreciated is a match in thinking styles. The minds of the co-founders should complement and augment each other. Brian Armstrong said:

I was searching for someone who would be genuinely excited about my idea, but also complementary to my skills—someone who could push me to think bigger and offer perspectives I hadn't considered. I wanted to leave meetings feeling energized and having learned something new.

Resources

• How to Find a Co-Founder - Erik Torenberg

• The Co-Founder Playbook with Julian Weisser, Co-Founder & CEO at ODF

• Brian Armstrong on co-founders

• Co-Founders: Finding the Right Partner to Build Your Business

• How to find a co-founder for your startup (Carta)

• How To Find a Co-Founder: Navigating Your Startup's Biggest Decision (HubSpot)

• The Founder Dating Playbook – Here’s the Process I Used to Find My Co-Founder (First Round Review)

• Finding a Cofounder for Your Startup (SVB)

• How to Get Lucky by Max Gunther (1986, to read)

• Guide: How to find, pick, and succeed with a co-founder (Mercury)

• My 5-step blueprint for finding the ideal cofounder (Indie Hackers)

• A Systematic Guide to Finding a Co-founder (F2 Venture Capital)

• Cofounder Courtship: How to Find the Right Mate—for Your Startup (HBS)

I recently looked up in the sky and asked myself: where are the flying cars (electric Takeoff and Landing Vehicles)? VTOL jet aircraft were invented in the 1960s. 65 years ago. The first successful manned eVTOL test flight was by Marcus Leng, founder of Opener Oero (now Pivotal Aero) in 2008.1 17 years ago.

Who is depriving us of the joy of cutting our commute from 60 to 6 minutes? The answer, as some of you might expect already, is regulation. But I wanted to understand what kind of regulation and compliance exactly is the problem and how I could solve this. I want to see our skies full of eVTOLs sooner than in 10 years.

Studies show that each extra minute of commuting lowers life satisfaction, with one finding a 20-minute longer commute reduces job satisfaction as much as a 19% pay cut.

Secondly, eVTOLs have the potential to drastically increase global GDP because they accelerate trade. History shows that where you've accelerated the rate of transportation within a network (reducing the economic distance) it increases trade and thus economic growth. As Walter Isard wrote:

Distance is the great separator; transportation, the great integrator.

On my quest to understand this missed opportunity multiple people have pointed me to the book “Where Is My Flying Car?: A Memoir of Future Past” which is severely outdated as it was published in April 2018. (It was republished by Stripe Press in 2021 without updates). For perspective, Archer Aviation (ACHR:NYSE, the second most valuable eVTOL startup with $3.7B) was founded in October 2018. The book also doesn’t explain the regulatory problems faced by VTOL companies. This is my attempt at doing so.

Regulation and Compliance

Aircraft Certification

History

In 2019 the FAA planned to certify winged eVTOLs under airplane rules. However, after a long debate the FAA decided that winged eVTOLs can’t be certified as part of existing regulatory categories2.

In May 2022, the FAA designated eVTOLs as "powered-lift" vehicles, requiring them to undergo a special certification process under 14 CFR Section 21.17(b). This section in the Code of Federal Regulations empowers the FAA to establish custom airworthiness requirements for aircraft that don't fit existing categories. It states:

For special classes of aircraft, including the engines and propellers installed thereon (e.g., gliders, airships, and other nonconventional aircraft), for which airworthiness standards have not been issued…, the applicable requirements will be the portions of those other airworthiness requirements contained in Parts 23, 25, 27, 29, 31, 33, and 35 found by the FAA to be appropriate for the aircraft and applicable to a specific type design, or such airworthiness criteria as the FAA may find provide an equivalent level of safety to those parts.

In practice, this means the FAA determines certification requirements for each eVTOL on a case-by-case basis by selectively applying standards from existing aircraft categories (such as helicopters, small aircraft, and transport planes). The FAA doesn't establish a comprehensive set of fixed standards for all eVTOLs upfront; instead, they review each design individually and specify which existing requirements apply and where new criteria are needed. This ad-hoc approach introduces significant uncertainty into the certification process.

Performance-based specifications

Both agencies use performance-based specifications, instead of rigid design standards (narrow metrics and materials) that an aircraft needs to check off.

To meet performance-based standards manufacturers can propose their own Means of Compliance (MOCs). The FAA evaluates each proposal individually causing uncertainty and delays. This is a costly and bureaucratic process.

Joby Aviation reportedly spent $100 million on type certification alone, with its airworthiness criteria spanning 130 pages (covering structural, mechanical, electrical systems, cybersecurity, human factors, and noise compliance), while broader certification efforts require thousands of pages of technical documentation covering safety assessments, system designs, and operational compliance. Other companies face similar obstacles: Lilium has incurred over $1 billion in losses as it pursues FAA and EASA certification, Archer Aviation has raised $1.5 billion to fund its Midnight aircraft program, and Vertical Aerospace, while projecting a lower cost of $350 million, has delayed its timeline twice, now targeting 2026. Industry-wide estimates suggest that the average timeline for eVTOL type certification has stretched from an initially projected 3–4 years to as much as 7–8 years due to the complexity of meeting both existing and novel regulatory requirements.

Certification Process

The certification of a new aircraft requires three approvals:

1. Type Certification (design approval):

   • A certification of the type (think SKU) of aircraft. Manufacturers need this to manufacture a specific type of aircraft3.

   • Phase 1: The first phase is about agreeing on a set of requirements with the FAA which results in what is called a G1 issue paper. Next the applicant and the FAA agree on detailed design requirements to meet these high level safety and performance benchmarks. Finally, they plan tests for these requirements.

   • Phase 2: The company must prove that it is able to produce the parts for the aircraft it wants to certify and once that is done perform the planned tests.

2. Production Certification (manufacturing approval):

   • Manufacturers need to apply for a production certificate to ensure its production facilities, quality system, and organization comply with the TC.

3. Operational Authorities:

   • Beyond the aircraft, operators need approval to fly commercially, tying into pilot training and airspace rules (discussed below).

Certification Privatization

In his piece Why aviation innovation matters Eli Durado argues that the certification process should be privatized completely. He points out that in the car industry the NHTSA (National Highway Traffic Safety Administration) allow sophisticated car companies to self-certify that they meet regulatory requirements. As a result cars have improved drastically in the last 50 years compared to aircraft. This stagnation is evident in our aging fleets - as of April 2025, the average age of United Airlines' fleet is 15.8 years.

Environmental Protection

Additionally, the Committee on Aviation Environmental Protection (CAEP) under ICAO sets environmental standards for aviation, which eVTOLs must meet regarding noise and emissions. This adds another layer to the certification process, ensuring eVTOLs align with international environmental goals.

Pilot Certification

The initial 2016 plan to regulate winged eVTOLs under airplane rules would have allowed pilots to leverage existing airplane certification frameworks. However, the 2022 pivot to power-lift lacks established pilot certification standards.

In 2023, the FAA proposed Special Federal Aviation Regulation (SFAR) for powered-lift pilot certification. Under this regulation pilots would need to earn powered-lift ratings specific to each eVTOL type. Manufacturer-employed pilots would need to serve as instructors. The extensive training requirements include ground school, simulation time, and actual flight training in dual-control eVTOL with certified instructors.

Dual control means that the aircraft has duplicated controls; one for the student pilot and instructor. Manufacturers would have to build and certify entirely different variant of their aircraft solely for training. This proposal also causes a Chicken-and-Egg problem. You need certified pilots to operate, but you need a certified training aircraft to create pilots.

Finally, the proposal ignores advancements in flight simulator technology which can realistically replicate eVTOL flight characteristics and emergency scenarios reducing the need for excessive flight training.

A more practical solution would be to add a “powered-lift” type rating to existing aircraft and helicopter licenses.

In August 2024, Beta Technologies became one of the first companies to receive FAA authorization to begin eVTOL pilot training. However, this is just for manufacturer and FAA personnel, not commercial operations.

Operational & Infrastructure Requirements

Vertiport Regulation: FAA’s draft vertiport design standards, issued in September 2022, outline safety and charging requirements, but final rules are not expected until late 2025. Historical precedent—like the 4-year gap from heliport standards draft to adoption in the 1990s—suggests vertiport approvals could delay urban construction into 2026 or 2027. Companies like Joby Aviation, targeting a 2026 commercial launch, may initially rely on existing heliports, with broader networks forming by 2028-2030.

Air Traffic/Space Regulation: The powered-lift classification still lacks an urban airspace framework. The FAA released an “Airspace Blueprint for Air Taxis” in May 2023 and aims for initial integration by 2026. However, low altitude traffic management systems historically take 5–7 years post-initial rules. For example, Drone Unmanned Aircraft System Traffic Management (UTM) started in 2016 and matured in 2022, implying eVTOL airspace readiness by 2028-2030.

Operational Procedures: The certification allowing air carriers to operate commercial flights is called Part 135 certification. Archer is progressing toward an 2026 approval. Pilot training rules, proposed in 2023’s SFAR, need 2–3 years to finalize—similar to the 2-year timeline for new pilot ratings in the 2000s. Beta’s 2024 training start hints at small-scale ops by 2027, but widespread use awaits full standards.

Other regulation: The FAA and EASA are drafting noise regulations, aiming to finalize them by 2025-2026, which could push back eVTOL launches as companies ramp up testing to comply. The International Civil Aviation Organization’s Committee on Aviation Environmental Protection (ICAO’s CAEP) drives environmental sustainability, enforcing noise and emissions targets for eVTOLs, but unclear battery lifecycle rules might limit scalability beyond 2026. Collision avoidance and maintenance regulations set by the FAA are also still unclear.

Loopholes and Workarounds

I came across three interesting ways to work around the certification regulation.

1. Public Aircraft Operations (PAO): This framework lets public services like first responder (FR) missions operate with fewer regulatory restrictions. For example, allowing rescue helicopters to land almost anywhere in an emergency. The eVTOL company Jump Aero leverages this framework to sidestep regulation and bring a first responder eVTOL to market that could save millions of lives. Pivotal is also exploring POA applications including border control.

2. FAA Part 103 Compliance: The FAA currently allows anyone to fly a single-seated aircraft if it’s lighter than 88 kg (194 lbs) and slower than 63 mph (100 km/h). They also can’t have a fuel capacity higher than 5 gallons (18 L), and are only allowed to be flown during daylight and over non-congested areas (they are not allowed to fly along a highway but over it), and for recreational/non-commercial purposes. Pivotal.aero designed their tilt-aircraft with simplified controls specifically to meet this existing standard. The aircraft of the Italian company Jetson (the Jetson One) would meet these requirements too. Here is an epic video of the Pivotal aircraft flying around the Golden Gate Bridge:

3. MOSAIC Rule: The FAA's proposed Modernization of Special Airworthiness Certification (MOSAIC) rule could enable small eVTOLs to be certified as light sport aircraft, provided they meet specific performance requirements. This could expedite regulation for smaller, personal, non-commercial models. However, it has not yet been finalized.

4. Military Contracts: More of a GTM strategy than a workaround but nevertheless interesting. Companies like Orb.aero (led by Thiel Fellow Alex Taylor) have secured defense contracts with the Navy. Their hybrid VTOL will be used as a decentralized infrastructure i.e., as a flying cell tower, a charging station for drones, and more:

Conclusion

It is likely that we’ll see the skies full of eVTOLs only in the beginning of the 2030s. The FAA is slow. A TC takes about 2–3 years, and it takes another year for a production certificate. The timelines of the vertiport development are a lot less clear.

I’m optimistic though that the day will come. I’d love to help make this vision a reality. If you know someone that is working on eVTOLs I’d love to talk to them and learn about their problems. I’m eager to make the vision of skies full of eVTOLs a reality. As Henry Ford said in 1940:

Mark my word: A combination airplane and motor car is coming. You may smile. But it will come.

Appendix

Batteries vs generators

From 2016 to 2018 DARPA ran a program resulting in Aurora XV-24 LightningStrike which looks strikingly similar to the famous Lilium jet. It relied on a hybrid electrical propulsion system. It had a generator (a gas-powered Rolls-Royce AE1107C turboshaft engine) which powered 24 ducted fans. Here is a simple energy comparison:

However, the generator likely emits significant noise making it less perfect for urban environments.

Global Airline Passengers growth

Humans want faster travel. The growth of global airlines passengers per year is exponential. The number of passengers in 2040 is projected to reach 10 billion. We can expect a similar growth for other forms of air travel such as eVTOLs once the regulations allow for it.

VTOL jets

VTOL (vertical takeoff and landing) aircraft were invented in the 1960s. One of the most majestic ones built in 1970 was the German Dornier Do 31. It was the only VTOL transport jet plane ever built. It could transport 39 people from SF to Palo Alto in ca 10 minutes (466 mph or 750 km/h top speed).

Historic Air Travel Deregulation

De-Regulation of air travel could cause incredibly cost savings. Historic events demonstrate this. In the 1970s, airfares in the US were regulated by the Civil Aeronautics Board (CAB). It regulated flights since 1938, controlling commercial aviation schedules and prices.

But recession-inflation in the 1970s raised concerns for the government, fearing that high prices for customers in a weak economy could lead to a collapse in the airline industry.

In 1978 the Government passed a bill through US Congress that gradually eliminated the regulation of airlines. By 1983, all airlines were free of domestic CAB regulation, which meant they could set prices and effectively compete with other carriers.

A cross-country flight between Los Angeles and Boston cost $4,439 in 1941, $915.82 in 1978, $408.89 in 2015, and a mere $119.67 in 2024. And that is despite airport fees and taxes significantly increasing in recent years. For example, the typical airport fee for a ticket has risen from $1.65 in 1979 to $29.70 in 2024.

Open questions

Some questions I haven’t researched that came up after I wrote this (if you have more please add them in the comments of this article and I’ll try to answer them):

• Why exactly did Lilium fail?

• How will I have to fly before I can fly vertically?

• What is the state of unmanned traffic management (UTM)?

Links

• Uber Elevate Paper (2016)

  • Relevant quote: The introduction of a new type of aircraft, however, requires a new certification basis, developed in parallel with the type certificate, and this could extend the end-to-end certification process to 4 to 8 years (and in the case of the AW609 as long as 20 years

• Why aviation innovation matters by Eli Dourado

• Interview of Geoff Bower (Chief Engineer at Archer) in Vertical Space Podcast (discussion about certification starts at about 33 minutes)

• Wikipedia: https://en.wikipedia.org/wiki/EVTOL

To Read

• https://readplaintext.com/a-big-dig-in-the-skies-bb9371726b1e

• https://www.wired.com/2013/04/then-internet-now-airspace-dont-stifle-innovation-on-the-next-great-platform/

What is vLLM?

vLLM is an open-source LLM inference and serving engine that utilizes a novel memory allocation algorithm called PagedAttention. It can run your models with up to 24x higher throughput than HuggingFace Transformers (HF) and up to 3.5x higher throughput than HuggingFace Text Generation Inference (TGI).

How does vLLM achieve this throughput improvement? Existing systems waste 60%-80% of the KV-Cache (LLM memory), whereas vLLM achieves near-optimal memory usage with a mere waste of under 4%. Because of this improved memory usage, it requires fewer GPUs to achieve the same output improving the throughput significantly.

Thousands of companies use vLLM to reduce inference costs. For example, LMSYS, the organization behind Chatbot Arena and Vicuna, cut the number of GPUs used to serve their growing traffic of ~45k daily requests by 50% while serving 2-3x more requests per second. vLLM leverages PagedAttention to reduce memory waste and achieve these efficiency improvements. 

What is PagedAttention, and how does it work?

In early 2023, the authors behind vLLM noticed that existing inference engines only used 20%-40% of the available GPU memory. When exploring solutions to this problem, they remembered the concept of Memory Paging from operating systems. They applied it to better utilize KV-Cache, where transformers store attention calculations during LLM inference. Let's start with understanding Memory Paging.

Inspiration—Memory Paging

In their paper, “Efficient Memory Management for Large Language Model Serving with PagedAttention,” the authors introduce PagedAttention, a novel algorithm inspired by virtual memory paging. Memory paging is used in operating systems, including Windows and Unix, to manage and allocate sections of a computer's memory, called page frames. It allows for more efficient memory usage by swapping out parts of memory that are not actively used. PagedAttention adapts this approach to optimize how memory is used in LLM serving, enabling more efficient memory allocation and reducing waste.

Understanding KV Cache

Have you noticed that the first few tokens returned from an LLM take longer than the last few? This effect is especially pronounced for longer prompts due to how LLMs calculate attention, a process heavily reliant on the efficiency of the Key-Value (KV) Cache.

The KV Cache is pivotal in how transformers process text. Transformers use an attention mechanism to determine which parts of the text are most relevant to the current tokens generated. This mechanism compares a newly generated query vector (Q) against all previously stored key vectors (K). Each key vector corresponds to earlier context the model considers, while each value vector (V) represents how important these contexts are in the current computation.

Instead of re-calculating these KV pairs for each new token generation, we append them to the KV cache so they can be reused in subsequent calculations. 

The Cache is initially empty, and the transformer needs to precompute it for the entire sequence to start generating new tokens. This explains the slow speed of the first few tokens. However, traditional implementations of KV cache waste a lot of space.

How PagedAttention Optimizes KV Cache Utilization

PagedAttention addresses these inefficiencies through a more dynamic memory allocation using the previously unused space. Let's understand how existing systems waste memory space. 

The parameters of a 13B LLM like llama-13b take up about 26GB. For an A100 40GB GPU, this already takes up 65% of its memory. We now only have about 30% of the memory for managing the KV Cache.

Existing systems store KV cache pairs in continuous memory spaces. In other words, they allocate a fixed, unbroken block of space for every request. However, requests to an LLM can vary in size widely, which leads to a significant waste of memory blocks.

The researchers behind vLLM break this down further into three types of memory waste:

1. Internal fragmentation: Slots of memory for a sequence output are allocated but never used, because the system does not know how many tokens the model will generate. The input prompt's complexity and the inherent design of autoregressive models like transformers influence the output sequence length. These models generate one token at a time, adjusting their output based on the tokens generated so far, which makes the exact sequence length variable.

2. Reservation: To ensure sequences are processed without interruption, the system reserves the entire block of memory for the duration of the request. Even if only a part of the reserved memory is used, the rest remains blocked off and cannot be used by other requests. In other words, the system prioritizes the stability and continuity of the current processing task over memory efficiency.

3. External fragmentation: External fragmentation occurs when fixed-size memory blocks do not match the variable lengths of sequences, leaving unused gaps between them that cannot be allocated to other processes, unlike internal fragmentation, which involves excess space within allocated blocks.

As a result, only 20-40% of the KV cache being used to store token states.

This continuous or fixed memory space allocation can suffice if the input and output lengths are static and known but is inefficient if the sequence lengths are dynamic, as is the case with most LLM inference use cases. 

PagedAttention applies the concept of Memory Paging to more efficiently manage the KV Cache. It allocates blocks dynamically, enabling the system to store continuous blocks in arbitrary parts of the memory space. They are allocated on demand. The mapping between the logical KV cache and the physical memory space is stored in the Block table data structure.

With PagedAttention, vLLM eliminates external fragmentation—where gaps between fixed memory blocks go unused—and minimizes internal fragmentation, where allocated memory exceeds the actual requirement of the sequence. This efficiency means nearly all allocated memory is effectively used, improving overall system performance.

While previous systems waste 60%-80% of the KV cache memory, vLLM achieves near-optimal memory usage with a mere waste of under 4%. 

The enhanced memory efficiency achieved through PagedAttention allows for larger batch sizes during model inference. This allows for more requests to be processed simultaneously and means that GPU resources are used more completely and efficiently, reducing idle times and increasing throughput.

Memory Sharing with PagedAttention

Another advantage of PagedAttention's dynamic block mapping is memory sharing between separate requests and their corresponding sequences.

Let's look at two LLM-inference decoding algorithms where this memory sharing helps us:

1. Parallel sampling/decoding: This method generates multiple output samples from the same prompt. It's particularly useful for applications where selecting the optimal response is crucial. For instance, in a content generation application, users can compare and choose the best variation. Similarly, in model testing scenarios, you can use it to compare several outputs. vLLM enables the sharing of the computation and memory for the prompt.

2. Beam Search: Beam search is a strategy used in tasks like machine translation to find the most likely sequence of words. It expands the most promising options at each step, known as beams. As more of the sequence is generated, these options are narrowed down by continuing the paths with the highest probabilities. vLLM efficiently manages the memory of shared beam branches.

3. Shared Prefix: Shared Prefix allows multiple users to start their requests with a common set of instructions or examples, known as a system prompt. This common starting point is pre-stored by vLLM to save computation time, enhancing efficiency and accuracy for subsequent tasks.

4. Mixed Decoding Methods: allow vLLM to efficiently manage requests that utilize various decoding strategies, such as greedy decoding, beam search, and top-k sampling. By optimizing memory sharing, vLLM can handle these diverse preferences concurrently within its system architecture, significantly enhancing throughput and system flexibility without the need for simultaneous decoding within a single request.

Other Performance Optimizations Enabled by vLLM—Quantization, Prefix Cashing & Speculative Decoding

vLLM implements several other techniques to increase performance:

Quantization

8-bit quantization reduces the precision of the model's numerical data from the standard 32 bits to just 8 bits. Applying this technique to a large model like Llama 13B, for example, which initially requires 52 GB in a 32-bit format (13b*32 bits = 13b*4 bytes = 52GB), would now only need 13 GB post-quantization. 

This reduction in size allows models to fit better in memory, enabling a model like Llama 13B to run on a single 24 GB VRAM GPU. The process usually comes at the expense of some performance loss, but the tradeoff is worth the performance bump for many LLM inference use cases.

Automatic Prefix Caching

Improves efficiency by caching repeated computation across different inference requests, particularly beneficial in scenarios with repeated prompts​, including multi-round conversations.

Speculative Decoding

Speculative decoding is when a smaller model is used to predict the output of a larger model, thus inferring responses more quickly and potentially doubling decoding speeds​​.

Multi-LoRA Support

Facilitates using Low-Rank Adaptation (LoRA) to manage multiple model adaptations efficiently. This allows a single base model to be used with various adapters, optimizing memory usage and computational efficiency across different tasks. An adapter, in this context, is a small, lower-rank matrix that fine-tunes the model on specific tasks without extensive retraining, enhancing versatility and resource efficiency​.

How does vLLM Compare With Other Inference Engines?

vLLM vs. TGI and Hugging Face Transformers

The Berkeley researchers behind vLLM compared the throughput of LLaMA-7B on an NVIDIA A10G GPU and LLaMA-13B on an NVIDIA A100 GPU (40 GB). vLLM achieves 24x higher throughput compared with HuggingFace Transformers (HF) and a 3.5x higher throughput compared with HuggingFace Text Generation Inference (TGI).

vLLM vs. TensorRT-LLM

A direct performance comparison with TensorRT-LLM is difficult as the effectiveness of each can vary based on the model configuration, quantization precision, and specific hardware used. For instance, while vLLM demonstrates robust performance with AWQ quantization, TensorRT-LLM can leverage fp8 precision, optimized for NVIDIA's latest H100 GPUs, to enhance performance.

That said, we can compare the setup time and complexity of these two inference engines:

• vLLM: streamlined setup – simply install the Python package using pip and define the model you want to use. vLLM is also OpenAI-compatible, making it easier to integrate into your existing infrastructure. While vLLM simplifies the setup process with its straightforward installation and OpenAI-compatible API, its user-friendly approach extends beyond just installation. vLLM also excels in versatility, supporting many models to cater to various application needs.

pip install vllm

llm = LLM(model="meta-llama/Meta-Llama-3-70B") # Create an LLM.
outputs = llm.generate(prompts) # Generate texts from the prompts.

• TensorRT-LLM: Setting up TensorRT-LLM involves a multistep process requiring familiarity with NVIDIA's software stack, Docker, and detailed configuration settings. This setup includes compiling models into TensorRT engines and configuring the NVIDIA Triton Inference Server for deployment, which may pose a steep learning curve for inexperienced developers.

What models does vLLM support?

vLLM supports the most popular open-source LLMs on HuggingFace, including:

• Classic Transformer LLMs (e.g., Llama, GPT-2, GPT-J)

• Mixture-of-Expert LLMs (e.g., Mixtral, Qwen2MoE)

• Multi-modal LLMs (e.g., LLaVA, GPT-NeoX, StableLM)

You can find a complete list here.

vLLM's Development & Popularity

Since July 2023, vLLM has rapidly gained over 20,000 GitHub stars and has a thriving community of 350+ open-source contributors.

Today, thousands of developers use vLLM to optimize their inference, and notable companies like AMD, AWS, Databricks, Dropbox, NVIDIA, and Roblox support it. The University of California, Berkeley, and the University of California, San Diego also directly support the project.

What do these seminal management ideas have in common? They all contribute to the elimination of risk, transaction costs, or both, by employing a modular organisation.

Transaction Cost Theory

In his Nobel Prize-winning college essay, The Nature of the Firm economist Ronald Coase argued that firms exist to eliminate transaction costs. 

Imagine a scenario in which a company outsources every single one of its tasks to an army of contractors. The company uses the free market to select the best and cheapest services. All these contractors compete with each other for these jobs. 

However, this would result in a lot of contracts and back and forth. Moreover, collaborations often create additional value not captured by the initial contract. This is a case of the hold-up problem, which arises when unforeseeable changes in the collaboration's value or terms increase one party's bargaining power, leading to renegotiation or conflict. 

As the number of contracts increases, so does the level of detail within each, often reducing the flexibility available to contractors. This specificity can constrain their ability to address unforeseen challenges creatively or adapt their approach as needed, leading to potential inefficiencies and suboptimal or unintended outcomes.

Companies hire full-time employees to eliminate these transaction costs.

Carl Dahlman later described three of the major types of transaction cost: Search cost, bargaining cost, and enforcement cost. In his essay about marketplaces, Dan Hockenmaier added a fourth: Distribution cost. 

But transactions within a company also have a cost. These are called organizational costs. They include:

• Coordination Costs

  • Communication Costs

  • Planning and Scheduling Costs

• Management Costs

  • Supervision and Monitoring Costs

  • Decision-Making Costs

• Administrative Costs

  • Record-Keeping and Reporting Costs

  • Compliance Costs

• Human Resource Costs

  • Recruitment and Training Costs

  • Employee Benefits and Compensation Costs

• Operational Costs

  • Inventory Management Costs

  • Maintenance and Repair Costs

Notably, as organizational networks expand, transaction costs can increase quadratically rather than linearly, a kind of reverse network effect where costs escalate disproportionately with network growth.

I’m fascinated by what product or platform would need to exist so that more contract work becomes possible and valuable. What transaction cost can we eliminate now that we couldn’t before? If you feel similarly excited about these ideas or contract work platforms in general and would like to discuss this further feel free to reply to this email or dm me on X. 

Around 2002, Jeff Bezos issued his famous API mandate, which later partly leaked. In it, he described how teams within Amazon should interface with each other to reduce transaction costs. (Bezos’ famous Two-Pizza Teams rule has a similar objective).

The mandate enforces a modular organization and architecture design. Each module, or an internally developed service, is now less “internal” (company-internal instead of team-internal). The entire organization can now utilize and criticize the module’s effectiveness.

Four years after the API mandate was shared, Amazon released S3 and EC2. The success of these and other AWS services demonstrates the effectiveness of his approach. AWS not only improved Amazon’s operations but also eliminated the transaction cost of Amazon’s customers developing and maintaining their own infrastructure.

Amazon shows that Coase’s insights apply not only to companies but also to teams within companies. The modular, API-driven approach also eliminated significant risks for his company, including lack of standards, documentation, excessive or unnecessary communication, and an inability to capture the value of initial internal services sold to the free market.

Paul Graham wrote:

[A] large organization could only avoid slowing down if they avoided tree structure. And since human nature limits the size of group that can work together, the only way I can imagine for larger groups to avoid tree structure would be to have no structure: to have each group actually be independent, and to work together the way components of a market economy do.

Subscribe now

Leave a comment

Risk Elimination

Most startup founders just hire to fill functional vacancies when uncompleted tasks pile up. Most managers or CEOs give a handful of tasks with differing priorities to their reports and hope that all get completed. Khosla and Thiel suggest that this is wrong.

In his article, Gene Pool Engineering For Entrepreneurs, Vinod Khosla recommends hiring people that are able to solve the problems that reflect the company's biggest current or future risks. The steps in this process are:

1. Identify the five largest risks facing the company.

2. Define the skill sets necessary to address those risks.

3. Locate centers of excellence for each risk (institutions to hire from).

4. List top experts at each center of excellence.

5. Recruit candidates from this “gene pool” to build a diverse team.

Counterintuitively, many pre-PMF (Product Market Fit) startups obsess about only hiring engineers. However, in this situation, they’d be better served by hiring an experienced PM or engineer with a product background that can help eliminate the risk of not finding PMF.

Khosla’s recruiting philosophy follows the process of derisking a startup idea. (Leo Polovet from Susa/Humba Ventures wrote a great post about startup de-risking). In some cases, Gene Pool Engineering can be part of the de-risking process if your startup inception depends on bringing together the right experts.

In his book Zero to One Peter Thiel wrote:

The best thing I did as a manager at PayPal was to make every person in the company responsible for doing just one thing. Every employee’s one thing was unique, and everyone knew I would evaluate him only on that one thing. I had started doing this just to simplify the task of managing people. But then I noticed a deeper result: defining roles reduced conflict. Most fights inside a company happen when colleagues compete for the same responsibilities. [...]. Eliminating competition makes it easier for everyone to build the kinds of long-term relationships that transcend mere professionalism.

Keith Rabois (ex. PayPal’s executive and Founders Fund partner) recalls that Thiel “would refuse to discuss virtually anything else with you except what was currently assigned as your #1 initiative.” 

A big benefit of this strategy comes from the reduced cognitive load that context-switching between projects or tasks requires. There is always an opportunity cost to starting a new task vs. finishing the existing one. This also applies to personal projects. This switching cost can be seen as a type of transaction cost. 

In another account Rabois phrased this strategy as:

[This] philosophy of building a startup is to take on the most challenging problems first, not the easiest. A lot of people focus on the things that they know how to solve, and they do them, and they defer solving the really intricate, difficult technology, unit economics, or customer acquisition problems until later.

If you don't innovate, and you don't come up with interesting breakthroughs, your company isn't going to be one of the companies that people remember forever.

As an illustration, he mentioned Stripe deferring to innovate on the financial risk function, even though this issue almost killed PayPal in 2001. Because Stripe’s founders knew that other companies had already proven it could be done, it could be safely deferred until much later.

This “One Thing” management strategy not only focuses on eliminating the most pressing risks the company faces but also reduces organisation costs of competition between employees for tasks.

In organizational design, this strategy is called Single-threaded leadership (STL). A computer program that is single-threaded processes one process at a time. 

Jeff Wilke (Ex-CEO of Amazon's worldwide consumer business) wrote: 

Separable means almost as separable organizationally as APIs are for software. Single-threaded means they don’t work on anything else.

This organisation around key initiatives (or risks) enables the organizations to run more initiatives in parallel. Another tenant of STLs is to minimize dependencies to other parts of the organisation (or companies). This allows teams to complete tasks without blocking, waiting for, or aligning decisions with others.

Conclusion

One shared theme of these strategies is that they make the organisation more modular. The design advantage of (hierarchical) modular systems is that they permit a quasi-independent evolution of their parts. Here are some examples:

• Automotive Industry: Car manufacturers can upgrade individual components like transmissions without overhauling the entire engine, reducing the risk of cascading failures and associated costs.

• Engineering Design: Chip designers enhance processor speeds independently of peripheral components like keyboards or USB ports, mitigating risks related to full-system upgrades.

• Software Engineering: Modular and uncoupled code facilitates easier refactoring and maintenance, decreasing the risk of introducing bugs across unrelated modules and reducing maintenance costs.

• Corporate Architecture: Companies with modular architectures can more readily outsource tasks, allowing them to expand product features or incorporate vendor-supplied services more efficiently, thus reducing both risk and transaction costs.

• Biological Sciences: In genetics, the localized impact of mutations shows the advantages of modular biological systems in managing risks associated with genetic variations.

• Hiring: Companies mitigate significant risks and decrease search and training costs by strategically hiring employees with specific skills to address these risks.

• Marketplaces: Platforms like Zillow, eBay, Airbnb, and Amazon tailor their services to efficiently eliminate specific risks and transaction costs—search, bargaining, enforcement, and distribution—faced by the parties, as reflected in their varying commission structures. 

In summary, organizations tackle key risks and reduce various transaction cost via a modular organisation. 

Thanks to Varun Shenoy, Brian Chau, Logan Chipkin, Martin Thaulow, Bart Vanderhaegen, and others for giving feedback on drafts of this.

This article draws inspiration from the excellent lecture "Stanford CS25: V3 I Retrieval Augmented Language Models" by Douwe Kiela, who, along with Patrick Lewis, Ethan Perez, et al., invented RAG in May 2020.

The idea of enabling computers to extract information from a knowledge base to assist in language tasks goes back decades, with early question-answering systems from the 1960s and IBM's Watson Jeopardy system having similar conceptual underpinnings. To understand the origins of the first RAG-like system in 2017 and its invention in 2021, we must understand the underlying retrieval technology.

Subscribe now

Retrieval

Sparse vs. Dense Retrieval

Sparse vectors are called sparse because they are sparsely populated with information (a lot of values are zero because most words don’t appear frequently). They require fewer computational resources to process and are often used to find information about a specific brand or object (ex. “Apple Inc.” not “Pear”) but can’t handle semantic meaning. Popular embedding examples are BM25 and TF-IDF (term frequency-inverse document frequency). This technique was used in one of the first instances of a RAG-like (LSTM as generator/reader) system for Q&A in 2017 (DrQA Chen et al.).

Dense retrieval enabled searching for semantic similarity. Unlike sparse vectors, the numbers in a dense vector represent learned relationships between words, compactly capturing their meaning. This means semantically similar words (like "doctor" and "physician") will have similar embeddings.

ORQA: Latent Retrieval for Weakly Supervised Open Domain Question Answering (Lee et al. 2019)

One of the first Q&A systems built on dense embeddings is ORQA. It is trained end-to-end to jointly learn evidence retrieval and answering, using only question-answer pairs. It treats retrieval as an unsupervised, latent variable initialized via pre-training on an Inverse Cloze Task (predicting a sentence's surrounding context).

Vector DBs & Sparse-Dense Hybrids 

Maximum Inner-Product Search (MIPS) involves finding the vector in a given set that maximizes the inner product with a query vector. In simpler terms, given a set of vectors and a query vector, MIPS aims to identify the vector in the set most similar to the query vector based on the vector's inner product.

Faiss: A library for efficient similarity search (Johnson et al. 2019):

A powerful library for similarity search and clustering of dense vectors that implements approximate nearest neighbor search (ANN) to solve MIPS search problems is FAISS, which laid the foundation for many of today’s popular vector DBs. 

ColBERT: Efficient and Effective Passage Search via Contextualized Late Interaction over BERT (Khattab et al. 2020):

ColBert is a state-of-the-art neural search model. It enables efficient semantic search by independently encoding queries and documents before comparing their fine-grained similarity via late interaction (delaying interaction until after separate encodings are created). It finds maximum similarity (MaxSim) matches between each query token and document tokens, aggregating these to efficiently estimate overall relevance (up to 170x faster compared with prior BERT-based retrieval models).

SPLADE: Sparse Lexical and Expansion Model for First Stage Ranking (Formal et al. 2021):

An interesting hybrid between Sparse and Dense Retrievers is SPLADE. It is a sparse retriever that uses query expansion, identifying synonyms and related terms for the query, enhancing its ability to capture semantic meaning even when not contained in the query. 

DRAGON: Diverse Augmentation Towards Generalizable Dense Retrieval (Lin et al. 2023):

DRAGON, a generalized dense retriever, undergoes training with progressive data augmentation. This method gradually introduces more challenging supervisions and diverse relevance labels over multiple training iterations, enabling the model to learn complex relevance patterns effectively. By exposing the model to varied supervisions and sampling difficult negatives, DRAGON can generate high-quality representations, improving retrieval effectiveness across queries and documents. This strategy enhances DRAGON's language representation, leading to superior performance on unseen queries and documents.

SANTA: Structure-Aware Language Model Pretraining Improves Dense Retrieval on Structured Data (Li et al. 2023)

Structure Aware DeNse ReTrievAl (SANTA) addresses the challenge of aligning queries with structured external documents, especially when addressing the incongruity between structured data (such as code or product specifications) and unstructured data (such as text descriptions). It enhances the retriever's sensitivity to structured information through two pre-training strategies:

1. leveraging the intrinsic alignment between structured and unstructured data to inform contrastive learning in a structured-aware pre-training scheme, and

2. implementing Masked Entity Prediction (utilizing entity-centric masking strategy that encourages LMs to predict and fill in the masked entities, fostering a deeper understanding of structured data).

🧊 Frozen vs Dynamic RAG 🔥

The industry has mostly viewed the components of the RAG architecture as separate components that work in isolation. We can call this “Frozen RAG”. In contrast, some research has focused on iteratively improving the individual components (we can call this “Advanced RAG”).

Ideally, in a “Fully Dynamic” model, the gradients from the loss function would flow back into the entire system (end-to-end training): retriever, generator, and document encoder. However, this is computationally challenging and has not been done successfully.

🔥 Dynamic Retriever but Fixed Generator 🧊

In-Context Retrieval-Augmented Language Models (Ram et al. 2023):

The authors of this paper introduce a re-ranker, which ranks the retrieved results (using a simple, sparse BM25) before passing them into the LLMs context. This component is dynamic, i.e., the training signal of the entire model is backpropagated into the re-ranker. They show that this optimization can result in performance gains, allowing a 345M parameter GPT-2 model to exceed the performance of a 1.5B GPT-2 model. 

REPLUG: Retrieval-Augmented Black-Box Language Models (Shi et al. 2023):

In this framework, the language model is treated as a frozen black box but is augmented with a tunable retriever model. The name stems from the idea that you can plug any LM into the system. The retrieved documents/elements are presented to the LM separately, and we compute the perplexity of the model, given the query and the retrieved item (LM likelihood). This information is used to train the retriever to select: 1. the highest Retriever Likelihood (using a similarity score) and 2. the lowest perplexity documents. This framework does not work for any model that doesn’t provide a perplexity score.

DREditor: A Time-efficient Approach for Building a Domain-specific Dense Retrieval Model (Huang et al. 2024)

The authors propose DREditor, a time-efficient approach to customizing off-the-shelf dense retrieval models to new domains by directly editing their semantic matching rules (i.e., how the model compares vectors in the embedding). Motivated by needs in enterprise search for scalable and fast search engine specialization across corpora, DREditor calibrates model output embeddings using an efficient closed-form linear mapping (calculating the adjustment) instead of the usual long adaptation fine-tuning (similar to what REPLUG is doing). Experiments on domain transfer and zero-shot setups show 100-300x faster run times than fine-tuning while maintaining or improving accuracy.

🧊 Fixed Retriever but Dynamic Generator 🔥 

FiD: Fusion-in-Decoder (lzacard & Trave 2020):

This paper addresses a core limitation of many RAG systems: we have to cram all the documents into the LM context. This is limited to the model’s context size. In this framework, we combine (concatenate) the query vector and the retrieved passages vectors before decoding them together into an answer.

KG-FiD: Infusing Knowledge Graph in Fusion-in-Decoder for Open-Domain Question Answering (Yu et al. 2021)

This paper adds another Graph Neural Net (GNN) re-ranking/filtering step to the FiD pipeline explained above. The authors correctly point out that FiD and other RAG frameworks wrongly assume that the contents of the retrieved passages are independent of each other. However, the entities referenced in the retrieved passages are likely related to each other, and their relationship can be modeled. The steps of the framework can be summarized as:

1. Retrieve relevant passages & embeddings via Dense Passage Retrieval (DPR).

2. Construct a knowledge graph from WikiData and neighboring context passages.

3. Utilize Graph Neural Network (GNN) for iterative re-ranking of passages based on semantic relationships.

4. Update passage embeddings to eliminate less relevant passages and enhance top-ranked selections for answer generation.

(For more details, you can find the main author’s presentation of the paper here).

SURGE: Knowledge Graph-Augmented Language Models for Knowledge-Grounded Dialogue Generation (Kang et al. 2023)

SUbgraph Retrieval-augmented GEneration (SURGE) addresses the problem with prior graph-based retrieval techniques where the LM can get confused by irrelevant content. Their framework aims to retrieve only a context-relevant subgraph, which is end-to-end trainable along with a generative model.

The GNN-based context-relevant subgraph retriever extracts relevant pieces of knowledge from a Knowledge Graph (no vector DB) and extracts candidate triplets (3 nodes). We generate a Retrieval Distribution for each triplet by calculating the inner product between the Context Embedding (based on the Dialog History) and our candidate triplet. This process involves exponentiating the inner product of the triplet embedding and the context embedding, resulting in a score that determines the relevance of the triplet to the dialogue history.

The authors further leverage contrastive learning to train the model to distinguish between knowledge-grounded responses (using the retrieved subgraph) and irrelevant alternatives, mitigating exposure bias that arises from only showing input and a single "correct" output during training.

KNN-LM: Generalization through Memorization: Nearest Neighbor Language Models (Khandelwal et al. 2019):

This is another interesting paper in which the authors attempt to make the LM outputs more grounded. This is done by comparing the vector distance of the model’s initial prediction to similar/neighboring passages from a data store. In this case, the database is a collection of key-value pairs comprised of a token and its proceeding tokens (context). Finally, the normalized KNN model outputs, ranked by their distances, and the LM output distribution are merged (interpolated) to converge on the final output. 

RAG: Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (Lewis et al. 2020):

This paper, mentioned in the introduction of this blog post, is the origin of the idea of a dynamic, end-to-end trained RAG system backpropagating into both the retriever and the generator. However, the document encoder step in this and the next paper is still static.

RETRO: Improving language models by retrieving from trillions of tokens (Borgead et al. 2022):

This paper showed that RAG using an LM pre-trained from scratch can outperform a 25x bigger model in terms of perplexity. I won't go into too much detail about the exact architecture because it is unclear if this paper is reproducible. (DeepMind never published the paper’s code, and I heard that other tier-1 AI research companies failed to reproduce it).

Essentially, in RETRO, the retrieved chunks are selected similarly to the Generalization through the Memorization/KNN process above, then added to the query, and processed by the transformer encoder (using chunked cross-attention). In contrast, in RAG and related architectures, the retrieved passages are used as additional context for the transformer decoder.

Fully Dynamic RAG

REALM: Retrieval-Augmented Language Model Pre-Training (Guu et al. 2020):

This paper represents the first fully dynamic RAG model (in which both the retriever, the generator, and the document encoder are updated). Its main limitation or downside is that it is not truly generative, just BERT-based, limiting its ability to produce completely novel/free-form text. Updating the document encoder is costly. REALM introduces asynchronous updates, where the knowledge base is re-embedded in batches.

Other Retrieval Research

Query/retrieval system related

RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval (Sarthi et al. 2024)

Sometimes, reasoning about a text requires a more abstract or holistic understanding of it. This is where simple chunk-based retrieval fails.

In the RAPTOR technique, similar text chunks are clustered and then summarized. The summaries, in turn, are then clustered and summarized. These leaves and summaries are organized into a tree structure.

RAPTOR supports two retrieval strategies: Tree Traversal, which retrieves nodes layer by layer, and Collapsed Tree, which flattens the tree for a breadth-first search. Collapsed Tree is more effective in most cases because it retrieves cluster summaries and leaf chunks.

Benchmarks show RAPTOR's effectiveness in improving RAG's accuracy by up to 20%, especially for queries requiring comprehensive contextual understanding.

FLARE: Forward-looking active retrieval augmentation (Jiang, Xu, Gao, Sun et al. 2023):

A limitation of some of the above-explained techniques is that they sequentially retrieve and then generate. This paper proposes a system that iteratively predicts the next sentence to retrieve relevant context if it contains low-confidence tokens.

HyDE: Hypothetical Document Embeddings (Gao, Ma et al. 2022):

A core problem of retrieval is that the user's query might not capture their actual intent. I.e., there is a difference between what someone thinks they want to know about vs. what they actually want to know about. The paper aims to address that through an intermediate query-rephrasing step. This leads to a process where the main weakness of vanilla LLMs is dampened by their main weakness: Hallucination against hallucination.

MuGI: Enhancing Information Retrieval through Multi-Text Generation Integration with Large Language Models (Zhang et al. 2024)

This paper builds on the above Query Re-writing idea. It introduces a framework named Multi-Text Generation Integration (MuGI).

The framework involves prompting an LLM to generate multiple pseudo-references, which are then dynamically integrated with the original query for retrieval. The model is used both for re-ranking and retrieval.

Query Rewriting for Retrieval-Augmented Large Language Models (Ma et al. 2023)

This paper introduces a trainable rewrite-retrieve-read framework (reversing the traditional retrieval and reading order, focusing on query rewriting) that utilizes the LLM performance as a reinforcement learning incentive for a rewriting module.

Context related

Lost in the Middle: How Language Models Use Long Contexts (Liu et al. 2023):

This paper points out a core problem with passing a list of context items into the model, sorted by their relevance, as it will attend more to documents at the beginning and the end. 

SILO Language Models: Isolating Legal Risk In a Nonparametric Datastore (Min, Gururangan et al. 2023)

This paper suggests a solution to recent copyright infringement lawsuits where companies like the New York Times are suing model training companies for training on their paywalled data. The authors suggest only using public domain data during training but augmenting the model with “higher-risk data” during test/inference time.

Augmentation/interactivity

CRAG: Corrective Retrieval Augmented Generation (Yan et al. 2024)

This paper proposes a method to improve the robustness of retrieval-augmented language models when retrieval fails to return relevant documents. CRAG tackles this by: 1. Assessing retrieved document quality with a confidence score, 2. Launching web searches for inaccurate retrievals, and 3. Refining knowledge with a decompose-then-recompose algorithm (segmenting the document into fine-grained strips, scoring each for relevance, filtering out irrelevant strips, and concatenating the relevant ones). CRAG improves RAG performance on short- and long-form generation tasks across diverse datasets, showcasing its generalizability and robustness.

WebGPT: Browser-assisted question-answering with human feedback (Nakano et al 2021):

The system presented here could be termed Web Search Augmented Generation. The Information Retrieval Model receives the query and can output browser commands (like clicks and scrolling) to extract relevant paragraphs from web pages that it determines as informative. It's trained on human demonstrations with Imitation Learning (Behavior cloning). In the second step, a separate Text Synthesis Model synthesizes the answers. Finally, a reward model predicts the system output score. The entire system is then fine-tuned using human feedback, i.e., the reward model (RLHF).

Toolformer: Language Models Can Teach Themselves to Use Tools (Shick et al 2021):

This paper generalizes the idea of augmented generation. It presents a solution that allows LLMs to use external tools via simple APIs. Tool usage shown in the paper includes a calculator, a Q&A system, search engines, a translation system, and a calendar. The steps can be summarized as follows: 1. The authors annotate a large text dataset and sample potential locations in the text where tool API calls could be useful; 2. At each location, they generate possible API calls to different tools; 3. They execute these API calls and insert the call+response back into the original text (like "[QA(Who founded Apple?) -> Steve Jobs]"), 4. They check if adding the app call reduced the perplexity loss of the LM for predicting the following token and keep the API call if it did, and 5. The resulting training data is used to fine-tune the original LM.

This system has many limitations, such as the inability to use tools in a chain, its ability to use tools interactively, or its ability to take into account the cost of the use of a tool.

Gorilla: Large Language Model Connected with Massive APIs (Patil et al. 2023)

One limitation of the previous paper is that tool use is limited to a small set of tools. In contrast, the authors of this paper develop a retrieval-based finetuning strategy to train an LLM, called Gorilla, to use over 1,600 different deep learning model APIs (e.g., from HuggingFace or TensorFlow Hub) for problem-solving. First, it downloads the API documentation of various tools. It then uses this data to create a question-answer pair dataset (using self-instruct). Finally, the 7B model is finetuned over this dataset in a retrieval-aware manner.

Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection (Asai et al. 2023):

The authors point out the problem with most RAG systems: they retrieve passages indiscriminately regardless of whether the factual grounding is helpful. The Self-RAG algorithm uses a special type of token called a "reflection token" to communicate between the different parts of the algorithm: Retrieve, IsRel (relevance), IsSup (fully or not supporting), and IsUse (useful response).

GRAG: Graph Retrieval-Augmented Generation (Hu et al. 2024):

Unlike RAG approaches that focus solely on text-based entity retrieval, GRAG maintains an acute awareness of the graph topology, which helps generate contextually and factually coherent responses. The researchers claim that GRAG significantly outperforms current SOTA RAG methods while effectively mitigating hallucinations.

CodeGRAG: Extracting Composed Syntax Graphs for Retrieval Augmented Cross-Lingual Code Generation (Du et al. 2024)

The core idea of CodeGRAG is to extract and summarize the control flow and data flow information from code blocks, creating a composed syntax graph. This graph serves as a bridge between natural language and various programming languages, capturing both semantic and logical information. The approach involves three main stages:

1. Syntax Knowledge Base Preparation: Code blocks are analyzed to extract graphical views representing their syntax and control information.

2. Knowledge Querying: Given a target problem, an informative query is generated and used to retrieve relevant knowledge from the prepared knowledge base.

3. Knowledge Augmented Generation: The retrieved graphical syntax knowledge is used to inform the LLM for enhanced code generation.

The paper emphasizes the importance of the composed syntax graph, which combines Abstract Syntax Tree (AST), Data Flow Graph (DFG), and Control Flow Graph (CFG) information. This representation allows for better preservation of code structure and logic compared to raw code text.

GRIT: Generative Representational Instruction Tuning (Muennighoff et al. 2024):

GRIT addresses a similar problem to the above-mentioned paper while being very performant. The authors train a single LLM to perform both text generation and embedding tasks via Generative Instruction Tuning. In other words, the model architecture of GRITLM allows it to process input text, create embeddings, and generate output text.

Beyond the conditional tool use capability, performance is further enhanced by re-using: 1. The query’s vector representations for retrieval and generation and 2. Reusing the document key-value store (basically the raw retrieved vector db data) for generation.

GRITLM sets a new benchmark by outperforming all generative models up to its size of 7 billion parameters, excelling in both generative and embedding tasks as demonstrated on the Massive Text Embedding Benchmark (MTEB) and various other evaluation benchmarks.

Conclusion

All the products that some of us use daily, like Intercom.com’s AI chatbot, Perplexity.ai, You.com, phind.com, Komo.ai, or ChatGPT Browse with Bing, implement a frozen or less dynamic form of RAG.

Much of the research summarized above lies dormant and has seen little application. It will be exciting to see companies commercializing secure and performant applications of this technology, enabling faster and more reliable access to knowledge.

Further Reading

Other summaries/literature reviews

• Retrieval-Augmented Generation (RAG): Paradigms, Technologies, and Trends (Presentation: Haofen Wang Tongji University)

• Retrieval-Augmented Generation for Large Language Models: A Survey (Paper: Gao et al. 2023)

• Knowledge-Augmented Large Language Models with Personalized Knowledge Representation, Retrieval, Injection, and Verification (Research Statement: Baek)

• Awesome papers about unifying LLMs and KGs (GitHub Readme)

• Top Information Retrieval Papers of the Week (Substack Newsletter)

• A Practitioners Guide to Retrieval Augmented Generation (Blog post)

More Papers

I plan to summarize more of these here when I find the time. If you think I’m missing a paper, feel free to leave a comment or DM me about it (⭐ = paper I’ll add in the next days/weeks):

Filtering and ranking

• Retrieving Texts based on Abstract Descriptions (Ravfogel et al. 2023)

• Retrieve to Explain: Evidence-driven Predictions with Language Models (Patel et al. 2024) ⭐

• RAG-Fusion: a New Take on Retrieval-Augmented Generation (Rackauchas 2024)

• Similarity is Not All You Need: Endowing Retrieval-Augmented Generation with Multi–layered Thoughts (Gan et al. 2024) ⭐

Transformer memory

• Memorizing Transformers (Wu et al. 2022)

• TRIME: Training Language Models with Memory Augmentation (Zhong et al. 2022) ⭐

• Transformer Memory as a Differentiable Search Index (Tay et al. 2022)

Multi-modality

• Cross-Modal Retrieval Augmentation for Multi-Modal Classification (Gur et al. 2021) ⭐

• Retrieval-Augmented Multimodal Language Modeling (Yasunaga et al. 2022)

• Towards Language Models That Can See: Computer Vision Through the LENS of Natural Language (Berrios et al. 2023)

Knowledge Graphs & Reasoning

• Knowledge Guided Text Retrieval and Reading for Open Domain Question Answering (Min et al. 2020)

• UniK-QA: Unified Representations of Structured and Unstructured Knowledge for Open-Domain Question Answering (Oguz et al. 2020)

• Select and Augment: Enhanced Dense Retrieval Knowledge Graph Augmentation (Alfaifi and Alfaifi 2022)

• Neuro-Symbolic Language Modeling with Automaton-augmented Retrieval (Alon et al. 2022)

• Enhanced Story Comprehension for Large Language Models through Dynamic Document-Based Knowledge Graphs (Andrus et al. 2022) ⭐

• An efficient memory-augmented transformer for knowledge-intensive NLP tasks (Wu et al. 2022) ⭐

• Improving Wikipedia verifiability with AI (Petroni et al. 2023)

• KnowledGPT: Enhancing Large Language Models with Retrieval and Storage Access on Knowledge Bases (Wang et al. 2023) ⭐

• Knowledge Graph-Augmented Language Models for Knowledge-Grounded Dialogue Generation (Kang et al. 2023)

• G-Retriever: Retrieval-Augmented Generation for Textual Graph Understanding and Question Answering (He et al. 2024)

• CodeGRAG: Extracting Composed Syntax Graphs for Retrieval Augmented Cross-Lingual Code Generation (Du et al. 2024) ⭐

Other Reasoning Techniques

• StrategyQA: Question Answering Benchmark with Implicit Reasoning Strategies (Geva et al. 2021)

• Answering Complex Open-Domain Questions with Multi-Hop Dense Retrieval (Xiong et al. 2021)

• Demonstrate-Search-Predict: Composing retrieval and language models for knowledge-intensive NLP (KhattaB et al. 2022)

• ASQA: Factoid Questions Meet Long-Form Answers (Stelmakh et al. 2023)

• SPRING: Studying the Paper and Reasoning to Play Games (Wu et al. 2023) ⭐

• Do Large Language Models Latently Perform Multi-Hop Reasoning? (Yang et al. 2024)

Instruction & Memory

• Recitation-Augmented Language Models (Sun et al. 2022)

• InstructRetro: Instruction Tuning post Retrieval-Augmented Pretraining (Wang et al. 2023)

• RA-DIT: Retrieval-Augmented Dual Instruction Tuning (Lin, Chen, Chen et al. 2023)

• MemGPT: Towards LLMs as Operating Systems (Packer et al. 2023) ⭐

  • video

• Lift Yourself Up: Retrieval-augmented Text Generation with Self Memory (Cheng et al. 2023)

• Chain-of-Note: Enhancing Robustness in Retrieval-Augmented Language Models (Yu et al. 2023)

My other writing on LLMs/AI

A similar process is employed when deciding on a dietary approach. One such explanation that has gained traction is the “bioenergetic view/explanation” of health, which holds that impeded energy production (low metabolic function) is the sole cause of disease and degeneration. Thus, optimizing metabolism and minimizing foods that degrade it (for ex. polyunsaturated fatty acids) forms a hard-to-vary explanation of health.

A hard-to-vary explanation1 is one that is very precise. In the context of nutrition, this means describing the exact impact of a compound on an organism. Not just a small part of it. The explanation doesn’t use vague words or unexplained analogies. All of its entities play a functional role in the explanation and correspond closely to reality.

Contrarianism

You may have encountered the notion that seed oils are bad, while sugar is not the enemy, that milk is healthy, and a high pulse is good. Maybe you saw the seed oils and ‘soy boy’ memes (both soy and soy boys are very high in Polyunsaturated fatty acids).

Such claims challenge authoritative health/alphabet soup bodies like the Health and Human Services (HHS) and the Food and Drug Administration (FDA).

A proponent of these ideas is biologist Dr. Ray Peat, whose thinking starkly contrasts mainstream dietary advice, and who has not profited from his research (unlike other “help experts”).

Critical Rationalism

Peat’s style of thinking is highly aligned with the philosophy called Critical Rationalism, which I wrote about here on my blog before. This philosophy explains why Justificationism (the belief that knowledge comes from authority) is wrong. Knowledge doesn’t come from authority, it comes from conjectures. Peat writes2 about the same phenomenon leading to learned helplessness:

Professors of medicine see themselves as models of the authority that their students will need to apply in dealing with patients, and the physicians trained in the authoritarian style are likely to see their patients as recipients of their medical knowledge, rather than as occasions for listening and learning something new.

…

When our own steering system is commandeered by the authorities, our patterns of knowledge will be compartmented, and arranged in a fixed pattern. This kind of knowledge either deteriorates, or it seeks more of its own kind.

While self-regulation and the generation of knowledge are pleasurable, having knowledge imposed isn't.

…

For Pavlov, the study of psychology or physiology without consciousness was simply crazy. Pavlov said that he studied nutrition to understand consciousness and the nervous system, because eating is our closest interaction with the world. Our brain is part of our digestive system. But eating has become highly institutionalized and influenced by our cultural beliefs. If people begin to think about the meanings of eating, they are beginning a process of cultural and philosophical criticism.

Detailed Explanation

The bioenergetic explanation of health, as advocated by Peat, hypothesizes that the organization and flow of energy gives rise to biological complexity. The decline of this energy flow is the basis of diseases and aging.

This explanation seeks to increase the processes of life, contrary to dietary approaches like veganism and the carnivore diet that aim to slow them down3. Increasing the metabolic rate (i.e., cellular respiration) requires eliminating foods that dampen metabolic function and increase inflammation.

Dietary Implications 

Avoid

Polyunsaturated fatty acids (PUFAs) are molecularly fragile and easily oxidize spontaneously, preventing proper thyroid function4. Examples of PUFA-rich foods are vegetable oils like sunflower oil, corn oil, soybean oil, and fish rich in omega-3 fatty acids like salmon and mackerel.

Consume

Saturated fats (SFAs), are beneficial for their protective, thyroid-promoting, and pro-metabolic properties. Examples of SFAs are coconut oil, and animal fats like butter and ghee.

Certain simple carbohydrates, such as monosaccharides (glucose, fructose) and disaccharides (sucrose, lactose), have lower glycemic indices due to their processing pathways.

Examples of monosaccharides include glucose found in fruits and honey, and fructose found in fruits, honey, and root vegetables. Disaccharides examples are sucrose (table sugar) and lactose found in milk and dairy products.

Notably, fructose is partly processed by the liver, leading to a slower and more steady rise in blood sugar, thus providing a sustainable source of energy.

This contrasts with polysaccharides (complex carbs) that can lead to rapid spikes in blood sugar levels due to their more direct metabolic pathway.

Milk, packed with balanced macronutrients, calcium, and vitamins A, D, and K, aids metabolism and has been historically used for its therapeutic, anti-stress properties.

Ensure a balanced protein intake from muscle meats and gelatin-rich foods to maintain an optimal amino acid profile and reduce inflammation. Gelatin-rich foods include bone broth, crab meat, shrimp, beef liver and kidney.

Measurement

Gauge metabolism using resting pulse and body temperature (heat is a by-product of energy production) measurements to evaluate the efficiency of your health practices. Any decrease in body temperature after eating may suggest a metabolism that relies on stress hormones.

Summary

If the Lord delights in us, he will bring us into this land and give it to us, a land that flows with milk and honey (Numbers 14:8)

The bioenergetic explanation of health offers a potentially superior explanation of health. However, all theories, including those of health, remain fallible. Beyond that, I’d argue, that the quest for ideal health (protocols) is part of the quest for truth.

If this piqued your interest, you should check out the following:

Resources

Introduction

• Ray Peat: Where to Start by Yago (article)

• Bioenergetics I - Life, Energy & PUFAs by Yago (article)

• How to Heal Your Metabolism by Kate Deering (book)

• Ray Peat Inspired Nutrition (playlist)

• Raypeat.com (website)

People to follow

1. Nate Lawrence (Substack, YouTube)

2. Danny Roddy (Substack, YouTube

3. Yago (Substack, Twitter)

On Milk

• The Cosmic Diet: Eating in a way in which to support the flow of cosmic energy through an organism

• Milk-Wars on

• ARTICLE Milk in context: allergies, ecology, and some myths

• Real Milk Cures Many Diseases

• Milk Manifesto by @anabology

On Obesity

• We Used To Eat A Lot More Without Becoming Obese by Sven

• Fats and degeneration by Ray Peat

Thanks to Noah Glanz, Yago, and Nate Lawrence, for their feedback on drafts of this.

But what makes a good startup idea, secret, or technology inflection point? And how do you find them? In this article, I summarize 50+ blog posts and podcasts on the topic of startup ideas by the world's most successful founders and VCs (including Andy Rachleff, Peter Thiel, Marc Andreessen, Vinod Khosla, Paul Graham, Sam Altman, Naval, Elad Gil, Geoff Lewis, and Keith Rabois).

This piece aims to be as practical as possible. I included examples in every section, with a total of 54 company examples. I’ll also touch on my personal experience co-founding RemNote.com (backed by General Catalyst).

Content:

1. Problems

2. Secrets and Trends

   1. Secret and Trend Types

      1. Secrets in Nature (disruptive v. sustaining tech)

      2. Secrets in Behavior

   2. Narrative Violations & Mirages

3. Where to look

   1. Proxies and Competition

4. Exploration

Problems

Problem-solving is how humanity makes progress. In theory, any problem can be solved except if its solution is prohibited by the laws of physics.

A solution to your own problem or one you notice is called an organic startup idea2. (See the questions that help you find organic startup in the last section of this post). These startups can be easy to build because you save time understanding your customers and what they want. YC claims that 70%3 of the ideas of the successful startups they surveyed were organic. One reason that these are more successful is that they are likely based on a real problem. Michael Seibel (managing director at YC) said4:

I would start thinking about problems. [A] lot of people keep little idea books with start-up ideas. I think they'd have a lot more success if they kept little problem books and wrote down notes on the problems that they encounter.

The most important aspect of a problem is its severity: How much time or money is the problem causing your customer or end user, and how frequently and urgently do they experience the problem? How much do they want a solution?

A solution in search of a problem (SISP) is often addressing a made-up and less severe problem. Though, as we’ll see in the next section, Andy Rachleff has a more optimistic view of SISPs.

In the last section of this post, I’ll also talk about the process of finding and discovering interesting problems.

Secrets and Trends

In economics, disbelief in secrets leads to faith in efficient markets. But the existence of financial bubbles shows that markets can have extraordinary inefficiencies—Peter Thiel (Zero to One)

In other words, a secret is a market inefficiency.

Similarly, Chamath Palihapitiya said5:

There is always a scarce resource somewhere. So whether it was all of us deciding to organize around safety, inventing language, creating money, building a machine, creating a smartphone, somewhere along the way, someone was asking the question, what is the next great problem that needs to be solved? Where I stand today, what is the scarce resource that exists? What limits progress?

Contrary to popular belief, the number of startup ideas and secrets is not shrinking. It is growing6. Humanity gains new desires and market inefficiencies with every new human on earth. Humanity’s future is infinite, as is the number of world-changing companies.

Similar to secrets, broader trends are often overlooked, though less secret. Trends are also referred to as the "why now" in pitch decks7. To identify secrets and trends, ask:

• How is the world changing in fundamental ways?8

• What valuable company is nobody building?9

• What company would I like to own in 10 years?10

A way to assess the current market size is to ask11: 1. how many potential customers are there, and 2. how much will each of them be worth to you if everything goes right? This is summarized by the formula:

Market Size = (# of Customers) × (Revenue/Customer)

Predicting the future market size or latent demand is tricky. Many underestimated the potential market sizes of Airbnb, Lyft, and Uber because they didn’t understand the potential or future market size, demand elasticity12, and disruption potential of taxi and travel accommodations.

Andy Rachleff (co-founder of Benchmark) describes the importance of a deep understanding of what products a technology enables and how a product can address a desperate market as the quality of the insights13:

Almost everything you read about entrepreneurship would lead you to believe that entrepreneurs evaluate markets, look for problems, and come up with Solutions. That will lead you to the right and consensus quadrant14. That's really easy to do. Anybody can do that. What great technology entrepreneurs do is look for an inflection point in technology that enables a new product, and then they try to find a market that wants that product…

It’s all about the quality of the insight and if you’re able to find a market that is desperate for that insight. [Great founders] live in the future. They can see the implications of that inflection point, and what that can solve, way before anyone else can. 

This framing closely mirrors the definition of an invention that Brian Arthur gave in his book The Nature Of Technology: What It Is And How It Evolves15. In his book, he explains that we invent by linking the following two aspects (conceptually & physically):

1. A demand,

2. An exploitable effect (or set of effects), i.e., principle

Examples: 

• Jeff Hawkins built PalmPilot’s PDAs for a price point of $300 by removing unnecessary functionality while everyone else was trying to build them for $2000. Hawkins understood the technology inflection points that cheap PDAs with a high margin were possible and saw the desperate demand for them.

• Anne Wojcicki's 23andMe anticipated the continuously declining costs of genetic data, the rise of Web 2.0, and social networks. She transformed users into active research participants, pioneering a novel crowdsourced genetic study model, thereby disrupting incumbent research institutions like Stanford and Pfizer16.

• Commonwealth Fusion Systems (CFS) became viable once rare-earth barium copper oxide (REBCO) superconducting tapes were commercially available. These enabled the high-field, high-temperature magnets essential for CFS's compact tokamak designs, allowing for smaller, more cost-effective fusion reactors capable of net energy production.

• Andy Rachleff explained that the idea of Wealthfront came when he learned about new APIs in the brokerage world that would enable building an automated investment service, replicating the sophisticated models of endowment funds, thereby making the models available to the masses17.

• Founded in 2008, Twilio's founder, Jeff Lawson, realized that creating an easy-to-use and cost-effective communication API service had just become feasible, and he accurately anticipated the latent/growing demand for such services.

Peter Thiel differentiates secrets in nature and secrets in behavior. Similarly, we can differentiate between Technology trends and market trends:

Secret and Trend Types

Like surfers who aspire to greatness, the founders sought out their wave, often seeing it much sooner than almost anyone else. They had the foresight to see the future first, the insight to build the right kind of surfboard, the courage to paddle very far out into the ocean amongst the sharks (and naysayers) long before the wave hit, and ultimately the steely determination to ride atop the wave, even as it rose to scary heights, and broke in ways they couldn’t foresee.

—Greg McAdoo (ex. partner at Sequoia)18

Secrets in Nature 

These are technological possibilities that are widely underappreciated. To find technology trends, we can ask: 

• Which technology or platform adoption curves are starting to hit scale over the next few years? 

• What’s the science fiction version of product X?

• What did just recently become possible? (​​Any time you can think of something that is possible this year and wasn’t possible last year, you should pay attention).

• What did just recently become necessary? (A good example here is Checkr, which does background checks for delivery startup gig workers).

• What costs have plummeted or will plummet? (Many successful companies, including YouTube and Dropbox, were enabled by the plummeting database hosting costs).

A useful framework to evaluate the technologies behind a startup idea is that of disruptive/strong and sustaining/weak technologies. Keith Rabois (partner at Founders Fund) says that the most valuable companies are often built on the former, those that challenge the status quo and shift the paradigm19. Whilst strong technologies may initially seem expensive, risky, unusual, or toy-like20, they capture the imagination of enthusiasts. They tend to define new eras, forcing the world to evolve and mold to their requirements. In his post, Chris Dixon (partner at a16z) gives a few examples of pairs of strong and weak technologies:

Unlike disruptive technologies, incumbents are often quick to adopt sustaining or weak technologies. Generative AI might fall more into the weak technologies bucket because companies of all sizes are rapidly adopting it. However, the adoption itself represents a behavioral trend (see below).

Disruptive/strong technologies often develop according to the Perez21/Gartner/Amara22 hype cycle. That is, after an innovation trigger and peak of inflated expectations, there is a trough of disillusionment in which the technology loses in hype. During that time, a weak technology might appear nearer to mainstream adoption.

Similarly, the Founders Fund23 recommends commercializing the technologies that have languished:

[C]omputers and communication technologies advanced enormously (even if Windows 2000 is a far cry from Hal 9000) and the Internet has evolved into something far more powerful and pervasive than its architects had ever hoped for. But a lot of what seemed futuristic then remains futuristic now, in part because these technologies never received the sustained funding lavished on the electronics industries. Commercializing the technologies that have languished seems as good a place as any to start looking for ideas.

Examples:

• In 1995, Nintendo released a virtual reality console called the “Virtual Boy” that sold dismally. In 2012, Oculus started working on its first headset, Oculus Rift.

• In 2012, Databricks was founded by a team from U.C. Berkeley who noticed Silicon Valley tech companies successfully used a different technology, now known as Apache Spark, for AI and machine learning. Despite initial resistance due to its perceived academic nature, the team started Databricks to promote this technology, proving its enterprise readiness and potential to revolutionize the industry24.

• In 2017, Vectordash (YC W19) tried to build a GPU cloud service focused on Machine Learning. The company pivoted into cloud gaming because of a lack of demand and eventually shut down. In 2022, many crypto-focused GPU server centers were underutilized, while LLM technology saw fast adoption. This enabled the success of Machine Learning focused GPU marketplaces like RunPod.io.

Secrets in Behavior 

These are things that people don’t know about themselves or things they hide because they don’t want others to know. “Market trends” is a combined term for consumer or business behavior changes, regulation changes, etc. Marc Andreessen wrote25:

In a great market – a market with lots of real potential customers – the market pulls product out of the startup. The market needs to be fulfilled, and the market will be fulfilled, by the first viable product that comes along. The product doesn’t need to be great; it just has to basically work. And, the market doesn’t care how good the team is, as long as the team can produce that viable product.

To find secrets in behavior, you might ask: 

• What hacky, bizarre behavior could become mainstream26?

• What is a trend (a long-term shift in behaviors or preferences often characterized by rapid growth), and what is a fad?27

• Which behavior or usage sounds dumb but has traction? (A new mainstream behavior is often viewed as dumb in the beginning. Andrew Chen (at a16z) described this as the dumb idea paradox28.

• What market is relatively small today, where something is changing to make it a big market tomorrow?29

• What area has a growing customer base but they’re being dramatically underserved30 or overserved (enables low-end disruption)?31

Examples:

• Mariam Naficy leveraged crowdsourcing and an untapped talent pool of bloggers to create Minted, a successful online design marketplace for high-end paper products, despite initial skepticism around the viability of print media32.

• Robinhood capitalized on the interest of short-term, commission-free stock trading on mobile phones. Before Robinhood, wealthy individuals traded stocks mostly through financial institutions.

• Affirm capitalized on the trend of millennials avoiding credit card debt by offering point-of-sale loans for e-commerce purchases.

Another strategy is to ask: which products, platforms, or behaviors are being bundled or unbundled?33

One approach is to take adjacent categories and determine which 20% of features have an 80% influence. Look at overlapping categories, and then start documenting the important features.

Examples:

• Rippling created a suite of bundled but independently useful and compatible HR-related SaaS products. This creates a large surface for upselling their other products.34

• Intercom bundled support chat, outbound messaging, onboarding, and analytics.

• Pendo came about from consolidating analytics and in-app personalization.

• RemNote (my first startup) capitalized on adoption of spaced repetition and knowledge management software. We noticed that many people were copying text from their note-taking apps into flashcard apps and solved this problem by creating a unified workspace for learning and thinking.

Narrative Violations & Mirages

Another type of Secrets is uncovered by violating narrative mirages, i.e., misleading or incorrect ideas about the world. They are fueled by popular narratives that create an illusion of a market or technology, leading to a rush of investments and potential bubbles.

This phenomenon, termed Narrative Mirages by Geoff Lewis (founder/VC at Bedrock Capital)35, can often mislead investors and entrepreneurs into believing in the promise of a nascent startup or a new market category.

A narrative violation occurs if an idea, investment thesis, or knowledge about a company, market, or technology contradicts the popular narrative. These often represent great opportunities. In his letter, Lewis goes on to list some examples:

Howard Marks, founder of Oaktree Capital wrote36:

Most great investments begin in discomfort. The things most people feel good about – investments where the underlying premise is widely accepted, the recent performance has been positive, and the outlook is rosy – are unlikely to be available at bargain prices. Rather, bargains are usually found among things that are controversial, that people are pessimistic about, and that have been performing badly of late.

[…]

Unconventional behavior is the only road to superior investment results, but it isn’t for everyone. In addition to superior skill, successful investing requires the ability to look wrong for a while and survive some mistakes. 

Narrowing down the search space (Where to Look)

In my own exploration, I found focusing on a narrow problem space extremely useful. This focus could mean focusing on a particular industry, customer, technology, or regulatory space. This focus is helpful because 1. your learnings from customer discovery calls within an industry will compound, and 2. it helps you to ensure the ideas you focus on have a good founder market fit.

Founder Market Fit

To increase your founder market fit, searching for an idea in fields that leverage your unique background, skills, and resources can be productive. 

It might make you aware of changes or problems others can’t see. To find organic startup ideas for each life experience, internship, job, etc., think carefully about the following:

• What problems did you come across?

• What did you learn that other people don’t know?

• What problems have you been in a special position to see?

For B2B ideas, you can ask:

• What internal products have you built again and again at past companies?

• What’s a tool you wish you had (and would have paid a lot of money for) at your previous companies?

Schlep blindness spaces

A place to look for secrets is where no one else is looking. Frequently overlooked ideas are those occluded by schlep blindness, as Paul Graham describes. For example, founders tend to shrink from uncomfortable idea spaces such as payments or payroll software, as they require complex state licenses. Boring ideas get left on the table.

Prone industries and markets are another good place to look for ideas. As Paul Graham wrote37:

Since startups often garbage-collect broken or stagnant companies and industries, it can be a good trick to look for those that are dying or deserve to, and try to imagine what kind of company would profit from their demise. For example, journalism is in free fall at the moment. But there may still be money to be made from something like journalism. What sort of company might cause people in the future to say "this replaced journalism" on some axis?

Many companies in these spaces are slowed down by bureaucracy and navel-gazing38.

Marketplace Criteria

Criteria for marketplace ideas are39:

• Currently massive or growing into a massive market size

• Manual existing processes

• Inefficient capacity utilization

  • Flexport rationalized international logistics to make optimal use of cargo space.

  • SpaceX’s reusable rockets reduce the cost per launch.

  • Opendoor’s home sales process reduces unoccupied homes.

  • Airbnb took an underutilized resource, broke it up into smaller modular units (day-based sublets), and created a low-friction marketplace to sell those units.

• High fragmentation (on the demand and supply side)

  • I.e., there are many small market participants, leading to inefficiencies due to a lack of standardization and complexity of the purchase process.

• Low NPS (net promoter score)

It is also helpful to think of any marketplace idea as a solution that reduces certain transaction costs (Dan Hockenmaier has a great post about this40). I also wrote about transaction cost and risk elimitation on this blog.

Proxies and Competition

A proxy is a large and successful company that does something similar to your potential startup idea but is not a competitor. However, Thiel pointed out that ideas based on an analogy to a different business are more likely to lack a true secret because others can also come up with the analogy. Similarly, bringing an existing idea to a different platform or applying it in a different industry can be fruitful.

Examples:

• Rappi, the most successful food delivery startup in Latin America, used the success of DoorDash as a proxy41. 

• Before Airbnb, there was Vrbo (founded in 1995) and couch surfing42. Neither facilitated payments nor provided professional photographers to their landlord customers.

• The founders of Zip got advice from their YC Partner to look for a large market with entrenched companies that weren’t great, which led them to the procurement space (Incumbents here include Coupa and Bill.com)43.

• Instagram brought the existing idea of photo sharing from popular desktop/web-based platforms like Flickr to mobile. 

• Right after Amazon showcased its first smart, checkout-free retail store, Standard Cognition (or Standard AI) realized that every grocery store would want the same technology.

• Palantir took big data analytics concepts (primarily used in consumer-facing industries for targeted advertising and recommendations) and applied them in the defense, security, and government sectors.

Counterintuitively, competition can be a sign of a good startup idea44. However, your startup idea typically needs to leverage a new insight (that makes your solution 10x better) to dominate your competition. Paul Graham wrote45: 

Worrying that you're late is one of the signs of a good idea. [...] You don't need to worry about entering a "crowded market" so long as you have a thesis about what everyone else in it is overlooking. In fact, that's a very promising starting point. Google was that type of idea. Your thesis has to be more precise… 

Examples: 

• Despite many other file-hosting competitors, Dropbox succeeded due to a significantly better UX, allowing users to connect the app to their operating system and sync files automatically. 

• Google’s secret was its superior page rank algorithm. 

• Facebook had fewer bugs and churn than Friendster and felt safer46. 

Exploration

Many of the best secrets are earned secrets. They are discovered while exploring a problem space.

Curiosity was designed by evolution to uncover problems and gaps in understanding, which is why it is a powerful drive to find areas for startup ideas. Vanta’s founder, Christina Cacioppo, said47:

There was this curiosity of thinking, these are smart, motivated people. How did they mess [their security/complience] up? That puzzle was just interesting to me, and it informed what Vanta ended up becoming.

Do you have some ideas already? Chris Dixon recommends creating a spreadsheet with your ideas and talking through each of them with the smartest people you know48. 

Nat Turner (who sold Flatiron Health for $1.9 billion) has an excellent approach to finding startup ideas, which might be described as “pitch to get feedback and iterate”. He recommends49:

1. Begin with a general space of interest and look for any seed of an idea

   1. Ask: What are the interesting problems in this space, and what is the root cause of these problems?

2. Network like crazy in the industry and take introductions extremely seriously

3. Create a deck and start pitching to anyone smart and relevant you can find on the specific idea. Take meticulous notes and follow up.

4. Tweak the deck between every meeting, hack together a demo, and start pitching (pre-sell software that doesn’t exist yet but will)

5. Find a trusted group of advisors to discuss key learnings from the field and get feedback.

6. Continue to iterate as hard and fast as possible on the feedback you’ve received.

Example: While still in college, Turner was interning at VideoEgg. After an executive during a sales meeting said, “We want to advertise on VideoEgg, but we don’t have a big production company to build a video ad”, he and his friend Zack Weinberg sat down and began to think of solutions. But instead of just researching the idea or talking casually with others in the online advertising space, they immediately began to pitch. Turner recounts the early days of Invite Media: 

We pitched everyone. We pitched potential advisors, investors, potential customers, friends, everyone. I can remember hundreds of meetings we had.

This research process aims to attain a sophisticated understanding of the problem space. A good startup idea is based on 100+ interviews rather than just 10.

From talking to several succesfull founders this process usually takes between 2 and 12 months.

Another approach is what I’d call the Ideal Customer Profile (ICP) or Title focused approach. Brian Long who co-founded the SMS marketing company Attentive outlined this interview template in his book "Problem Hunting”50:

1. What is your official title, and how long have you been in this role? How did you get this role?

2. What are your primary responsibilities?

3. We aim to build products that solve problems for you and your peers. What are the top three problems in your job today?

4. For each of the top three problems (repeat this section for problems #1, #2, and #3):

   • Can you explain more about this problem?

     • Why is it a problem?

     • What metrics track this issue?

     • What have you done to try and fix it?

   • On a scale from 1 to 10, how significant is this problem today?

     • 9-10 == burning problem

   • Why do you feel that way?

5. What are the top three problems for your company today?

   • repeat 4.

6. Is there anything you'd like to emphasize or that we might have missed?

7. If you were starting a company in this industry today, what would you focus on?

8. Do you have any other questions or feedback?

9. Would you be open to a follow-up conversation? What's the best way to reach you?

As mentioned in the introduction, the success of a startup idea is inherently unpredictable, which is why assuming predictable growth in a trend is error-prone. Existing secrets might become irrelevant, and trends might die down, but new ones can emerge. 

Exploring the idea maze before and during the company building process is an open-ended process. Paths from an initial niche idea into a giant company are hard to see. Good startup ideas can morph into great ones.

Examples: 

• Airbnb’s Brian Chesky, for example, read about major inefficiencies in the hotel industry after trying to rent out his air mattress.

• Zillow’s Rich Barton found the problem that price discovery is broken. When trying his original idea of running an auction on every house, he discovered the Zestimate (a house price estimation service). 

• Ramp: After Eric Glyman’s startup Paribus was acquired by Capital One, he talked to many customers in their credit card division. He asked, “Do you want points, cash back, or something different for rewards?” He found that they didn’t want any of that and just wanted to be better off. This led to the idea of a B2B credit card that helps you save money51. This is also an example of identifying a behavioral trend/secret.

• Coinbase’s founder Brian Armstrong applied to YC with a p2p bitcoin transfer application. He pivoted to focus on the problem of buying bitcoin52.

• Pinwheel's co-founders initially focused on health savings accounts (HSA), but found themselves struggling integrating with various payroll systems. Talking to people about their income and employment data problems, they noticed particular excitement about them offering direct-deposit switching (the process of changing the users linked bank account), which became their wedge into the market48.

• PillPack’s co-founders identified an innovative medication packaging solution and saw multiple potential routes for market entry (or “paths to value”), either complementing or challenging traditional pharmacies like CVS and Walgreens53.

• Twitch’s original idea was to create an online reality TV show.

• Lattice’s co-founder, Jack Altman, first built a tool for OKR tracking. It turned out that software wasn’t a good solution for managing OKRs. However, by building relationships with and talking to potential customers, they learned about their problems by conducting 360 performance reviews54.

• Scale AI’s original idea was to make it easier to book doctor's appointments. After discovering that interfacing with hospitals’ outdated information systems would be difficult, they revised their idea to put people on the back end to interface with hospitals. Users would input their insurance information, availability, location, etc., on Scale’s website, and they would take this information and call the hospitals and book an appointment for them. Next, they discovered that hiring the contractors for this work was not easy, leading to the idea of an API for human labor. Finally, they realized that the machine learning data labeling use case was in strong demand and decided to focus on that completely55.

• PayPal went through five business models before it found something that worked (including security software for Palm Pilots, sending money between Palm Pilots, and later sending money between emails)56.

Often, an innovation is combining two existing ideas. If you develop a better idea that fits into your original one, you can try to merge them.

Example: when coming up with the idea of Scribd, Trip Adler combined the idea of 1. “A way for academics, creative writers, or anyone to publish online more easily”, and 2. “The YouTube for all kinds of writing”57.   

This process is about deeply understanding the idea space and articulating the insight. Andy Rachleff explains58:

People who are really good at what they do, who have a tremendous command of what they do, are able to explain it in very, very simple terms, and they are the ones who are more likely to be able to pull off the non-consensus bet.

[…]

When you succeed you often stumble upon the market that's desperate.

Thanks to Robert Wachen, Charlie Curnin, Andrew Duca, Vincent Weisser, Miles McCain, Sven Schnieders, Jason Zhao, and Bani Singh for feedback on drafts of this.

34

luttig's learnings

Rippling and the return of ambition

In the 1970s, the tech industry was several orders of magnitude smaller than it is today. But each company had far larger product ambition. And each engineer had an outsized scope. Apple released its iconic Apple II computer in 1977 with just 25 employees…

Read more

4 years ago · 78 likes · 9 comments · John Luttig

Anton on Twitter: ⁠https://twitter.com/atroyn⁠

Moritz on Twitter: ⁠https://twitter.com/MoritzW42⁠

Mentioned Resources

• Unlimiformer: Long-Range Transformers with Unlimited Length Input: ⁠arxiv.org/abs/2305.01625⁠

• Memorizing Transformers: ⁠arxiv.org/abs/2203.08913⁠

• Chroma: ⁠https://www.trychroma.com⁠

• ⁠You and your Research⁠: ⁠https://www.cs.virginia.edu/~robins/YouAndYourResearch.html⁠

• Against Method⁠: ⁠https://en.wikipedia.org/wiki/Against_Method⁠

Contents

(00:00) Preview

(01:14) Introduction

(06:13) Edech

(12:12) Airchips cost structure

(24:06) Techno Optimism

(25:09) Convenience vs Enabling Technology

(31:13) Chroma and LLM Landscape

(35:37) Minecraft Bot

(37:05) Knowledge Retrieval

(39:38) New Knowledge Retrieval Research and Techniques: (Direct Retrieval Augmentation of Transformers, Conditional Retrieval)

(42:35) AI Native Platforms

(45:43) Theorem Proofing and Generation

(49:40) Do LLMs Reason

(52:18) Computational Complexity and LessWrong

(53:58) Frameworks and Taking Your Work Seriously

(55:53) Job Displacement and Jobs Supply vs Demand Side

(58:29) The Printing Press

Links

Twitter: ⁠Moritz (@moritzW42)⁠

Blog: ⁠https://scalingknowledge.substack.com⁠

About Erik: ⁠https://en.wikipedia.org/wiki/Erik_J._Larson⁠

Erik's book: ⁠https://www.amazon.com/Myth-Artificial-Intelligence-Computers-Think/dp/0674983513⁠

Transcript: ⁠https://scalingknowledge.substack.com/p/the-myth-of-ai-with-erik-j-larson#details⁠

Content

(00:00) Preview

(01:33) Intro

(06:55) Unpredictability of Invention

(10:09) Limits of Deep learning progress

(12:59) Abduction

(16:45) Creativity and Serendipity

(19:16) Neumann on Self Improving Machines

(24:30) His upcoming book

(26:40) Progress & Innovation decentralisation

(28:21) Neurosymbolic AI

(30:29) AI research progress

(33:16) Automation, Displament, and Alienation

(34:58) Outro

Transcript (Raw/Approximation)

[00:00:00] Erik: In a highly technocratic society, you have a lot of people telling you that they understand what's gonna happen next. So I don't buy the extrapolation. Now, you could certainly say, okay, we have GPT four, we're gonna have GPT 4.5. Although Sam Altman, who as I'm sure you know, is the head of open ai, he said we can't keep doing this.

We're gonna run out of text on the web and we're gonna route run out of computers to train it. It's actually a quote he said. His quote was, we're at the end of an era. The media has been saying that since Ford Motor, like since you know, Ford Motor Company, you know, and so it's just a perennial, it's a kind of Marxian idea that, you know, and advanced technology will displace and alienate workers.

John von Neuman, the famous mathematician. And he came out and said, you can't have a self-improving machine unless it's purely random. Because if it's actually planned, then you have to contain the blueprint for the improved machine in the machine and the lower machine. You know, if you, in other words, you, you have to pre, you have to put in the knowledge.

In other words, it's not going to think of it because it's a machine. It has to actually have access to. The components, the blueprint, right

[00:01:29] Moritz: today. I have the pleasure of speaking with Eric J. Larson, who is the author of the book, the Myth of Artificial Intelligence. So yeah, I'm really curious to hear more about your background.

[00:01:38] Erik: Okay, sure. So, I actually started in the field. We didn't really call it AI back then. It was sort of like, we called it machine learning. And now everything is ai. It's just kind of this meaningless title, you know. But I started working on machine learning you know, applications for, for language processing, for text processing.

And my first job was actually at a famous AI company in Austin, Texas called Site Corp. They're actually known for building this really large knowledge base that was supposed to. Exhibit common sense reasoning, but I started on January 3rd, 2000. So I can pretty much say that I've been working on AI for the entire century so far.

Right. So we're 23 years into the new century and I've pretty much been, yeah, so that's when I started. I came out of a background in mathematics, not computer science and philosophy. So I studied, I studied mathematics and philosophy as a double major. I. And I got interested in computer science, to be honest with you, when I, I needed to make more money, it's, there's, there's not a lot of money in philosophy or mathematics.

And so if you program you can, you can do well, especially in a town like Austin. It's very tech-savvy and, and. Very much in need of smart people to code. So that's what I started doing. I, I was a Java developer for a while at a big company, e d s, and then I transitioned back into doing more smaller artificial in artificial intelligence.

I was, Funded by Lockheed Martin's Advanced Technology Laboratory. For a while, I built a system to predict when a troublesome event was going to occur in a large corpus. I can't really get into it cuz some of it is actually classified, but, so I did, I did stuff like that. And then around 2006, I submitted a proposal for DARPA funding and probably, you know, about DARPA, and they, they gave, they gave me that funding and then they, they gave me the second year funding.

So I, I had a company and basically, it's kind of funny because people don't do blogs anymore, but we would basically say, In any language in the world, we would say, here's the blog, and here's what it's about. And, and it's funny because like nobody even thinks about blogs, but back in 2006, blogs were like really, really saturated information sources on the web.

There were blogs everywhere. It was a really big problem. How do you sort 'em, how do you search 'em? So we had a, a way of doing that. Around 2016, I started another company where I computed influence of. People on the web and said, I, you know, how do I, so my obsession was like, how do I assign this person a score?

How influential is this person holistically, like, taking everything into account? So how much information can I get about he or she and how do I actually assign a score that, you know, this person is 37, this person is 57. So I did that. It sold to an educational company, the best schools. Hmm. Yeah. And I made a, I made a, shouldn't talk about money, but that was a fairly lucrative venture.

Hmm.

[00:05:02] Moritz: And you also got DARPA funding for that one, right?

[00:05:05] Erik: Yeah, yeah. It was partially commercially funded, and it was al, it was partially funded through the d o D through DARPA. Yeah.

[00:05:13] Moritz: When you said commercially funded, what? I'm interested, like back in, or 10 years ago, what were the funding rounds for something like this?

How, how much did you guys, it

[00:05:20] Erik: was called strategic Investment. So it came not from vc, but actually came from the educational, the, the, the actual, they were potential acquire, they were owned by General Motors if you can believe that. Huh? Yeah. They, this education company was owned by General Motors, but yeah, so they, they put.

A couple of million dollars into it as what's called strategic investment. It's not vc. And the idea was that I can't, you know, I have a non-complete compete clause, so I can't go to another company. And basically, it was, you know, I. It was really great funding, but you're locked into that deal. And I took it.

And it, it worked out fairly well. I mean, it could've worked out better, but who knows? Yeah. And then, yeah, I started thinking I want to, I want to sort of, you know, set the record straight on ai. I've been working on, at that point, I'd been in, in the field for 15 years, around 2016, you know, And it's just like people are saying crazy stuff like they don't know what they're talking about.

They literally don't know what they're talking about. Yeah. And so that's, that was sort of the impetus for the book, you know. I finally got enough confidence with, you know, starting companies where it's like, okay, I'm gonna, I'm gonna say something here. You know, so, yeah. That, that's sort of my story.

Yeah. Let's

[00:06:32] Moritz: jump into epistemology a bit. So I am, I really like this quote you put in your book I think by Karl Popper quoted in this other book by US Asir McKen Tree After Virtue. Oh, okay. About basically prediction of innovation necessitates its invention.

[00:06:51] Erik: In a highly depiction of management society

[00:06:55] Moritz: with an example of the wheel, understand what's gonna, I'd be curious to hear sort of your, your summary of that point and also curious to learn more about this book after which, if you would recommend it.

[00:07:07] Erik: Oh, sure. Yeah. It's, it's one of the great books if you're into philosophical literature. It was sort of a WA watershed event. I think it came out in the eighties and still. Widely read and discussed. But his, his point was you ha he had, he had a whole constellation of points, but one of his points was, in a highly technocratic society, you have a lot of people telling you that they understand what's gonna happen next.

And he, he was trying to say, that's just Illusion is actually not true. And so how do you tell a highly technocratic society that they're that, that they don't know what's coming around the corner because it's dangerous to think that you know something that you don't? And so he pointed out inventions, and he used the example of the wheel, like the classic invention. Right? Right. Somebody, it was probably ac crude knowledge over time, but at some point somebody said, pulling this sled. Is way too much energy. And if we, you know, if we, if we had a, if we had a wheel, we could reduce the number of people pulling it. They could go do something else, division of labor, right?

Like, it, it just would, you know, and so somebody at some point had that idea. But if you say that you can predict the future, let's imagine you're a year before the invention of the wheel, just stipulate, right? How would you predict the wheel a year before? You either know what it is, in which case it's now, right?

It's not a year from now. It's now. In other words, you can't sort of fuzzily talk about the wheel. You either know what the invention is or you don't. So every invention sort of happens when it happens and you can't really predict it. And I use that as a way of saying like, you know, We either have artificial general intelligence and we know how it works, or we don't, you know, there's no point in saying it's coming in 20 years, it's coming in 30 years.

That's just dumb. It's just dumb. Yeah. I mean, we either know what it is or we don't, and right now we don't. So we don't know if it's a, if, if it's 30 years or 300 years, we just don't know, or if it's never going to happen. But claiming that we sort of know. Just, I don't mean to keep casting asperges, but it is a little bit stupid.

If you think, if you think about it for a few minutes, you'll get how it's dumb. It's like you don't know if you knew you would say so. Yeah, that was the point of me including that in the book.

[00:09:38] Moritz: When I ring up this point with friends or people, I'm debating on this day at this fuzzy thing in their head, that you can just like combine existing technologies and stitch 'em together into something that then works.

But that's essentially a, a new invention. So, It's something you, you cannot predict. But people ex like to ex extrapolate sort of the research funding that's going into these domains and yeah, especially one example is I think like reinforcement learning, combining that there's like a good amount of advance with like the Minecraft bot.

I'm not sure if you saw that. Like there's a bot that can play Minecraft and then save deer. It's a new sort of discovery to the, to the knowledge base basically. And you could combine that with like some of the GT four G T GPT technology. Yeah. Do you think there's some way we can predict these kinds of inventions or to what extent are we limited to do that?

[00:10:34] Erik: No, I don't. I, so when it comes to a real advance in ai, I think that we have to actually do, you know, we have to actually get the, the, the i, the blueprint first. So I don't buy the extrapolation. Now, you could certainly say, okay, we have GPT four, we're gonna have GPT 4.5. Although Sam Altman, who as I'm sure you know, is the head of open ai.

He said we can't keep doing this. We're gonna run out of text on the web and we're gonna run out of computers to train it. It's actually a quote he said, his quote was, we're at the end of an era. So he's already the guy who started all this. This craze about large language models is actually the one who's trying to shut it down now, you know, so we could extrapolate and say, we'll have GPT 4.5 and so on.

But what we can't, I don't think it's possible to extrapolate and say, We're gonna have a smart machine. That's generally intelligent, right? Right, because we just don't know what the blueprint is. And if you don't know the blueprint, you just really can't put a number on it. If you want to put a number on it, you're basically dealing with myth and religion.

You're not dealing with science. We just don't know. And it, it would be nice if the field sort of took that more seriously, but people just love talking about existential threats and so, you know, here we go. Yeah.

[00:11:53] Moritz: And then another point I liked in your book about abduction, and in my head, it sort of sounds like the same as conjecture or hypothesis generation and validation.

I'd love to hear your sort of explanation of abduction and maybe also how this relates to monotonic and defeasible inference.

[00:12:15] Erik: Yeah, so that's a lot of, there's a lot going on in that. So if you look at the, the, the classic case of induction is you have a, a sample of, you know, prior observations. So you can say, okay, I see a white swan, I see a hundred white swans.

I see a thousand white swans. And then my inference is going to be inductively that all swan swans are white. So you're actually looking at the hole. And then you're, you're trying to find a covering rule for it. What abduction does is say, I'm looking at this one event, right? Not, not, you know, what's happened in the past, I don't care.

It's typically a unique event. So it's something surprising that you see and, and it says, what's a plausible cause to account for this effect, right? Like, I make an observation. An individual's particular observation, well, what brought this about? Right? And so it. It's fundamentally different from what machine learning does, which is to say, I need more data.

The data is always from the past. How can it be in the future? Right? So I need more and more data, and then I'm gonna find a rule. The problem with all com, all AI research, including deep neural networks, everything right now is you can't handle anomalies or novelty, right? So generally speaking, this is the right text if you have a generative model, right?

Generally speaking, this is the right next word to put, right? But if something was really, if you had something that was really anomalous or off the system basically is just working off of what it is. What it has been, you know, exposed to in the past. So we don't do that. Like we use induction all the time, but we also have the ability to see something uniquely and reason about that.

I've never seen this before, ever, and I can still think about it. How is that possible? Right. How, how is it that I can see something for the first time and still wonder about what caused this? How, what explains this? How does this come about? So we clearly have a background, like a very large knowledge network that we're using in addition to machine learning.

We use something like neural networks. Obviously, we do, right? When we recognize images, it's pretty obvious that we're using them. Something like that works too, it's, it's not like at the level of neurons, it's not the same story, but like we're using something like, I've seen this before. I know what this is, but when we come across something that surprises us, we can't rely on machine learning.

Because machine learning is just painting a big umbrella over something. It's not saying this specifically is different, and human minds are way better than machines right now because we can do both, and we can do, you mentioned deduction, Monet. Monet just means if you're in a deductive system, once you conclude something, it never goes away.

And so Defeasible Logic says, Oh, I concluded this, but I got new information, so I'm gonna delete that premise. And Defeasible Logic was a big enthusiasm in AI in the 1990s, but it's very computationally expensive and it never got past like sort of toy textbook examples. Like nobody ever figured out how to sort of use it in the wild as it were.

So, The jury's still out on defeasible reasoning, but that, that's what that non-monotonic, by the way, defeasible, there's another term it's called non-monotonic, which means, hmm. Once you add something, you can take it back. Right. So I can say, oh, I think it was the butler. And then I can say, no, no, no. I think now it's the nurse, you know?

But in a purely deductive system, once you conclude, it's the butler. The butler just sits there and you can't get rid of it. That's a conclusion. Yeah.

[00:16:03] Moritz: Yeah. I am also interested in creativity. It seems like something that we haven't built, but a lot of people claim we already have with GPT. I'm curious, have you come across ways to formalize creativity or I think you also mentioned?

In your book, diversion Thinking?

[00:16:19] Erik: No, I mean, I don't have a recipe. I'm writing another book now. Mm-hmm. And so I usually sort of do something that doesn't make sense and so I, I try to find a way to corral or capture serendipity. So I'll, you know, for instance, I'm writing a book, another book, mm-hmm. And I spent about four or five months reading about the French Revolution.

It has nothing to do with my book, but Right. Like so I kind of like, I think like how we're better than machines is. You can, you can look for serendipity. You can say like, this doesn't, this isn't you know, this isn't connecting the dots. This is an A, B, C, D, you know, this is something different. And then you get inspiration and you think you, you have one thought that comes into your, in, into your brain that you wouldn't have had if you would've just, if I would've just read computer science literature, right?

I would've been boring, right? But I'm reading about the French Revolution and who knows what the connection is, but my writing got better. And so I think kind of like when I think about creativity, I think about finding new ways, find paths to serendipity where you're looking for something and you found something else.

Right? And that's, that's about all the theory I have on creativity. I'm not a psychologist, but that's what I do when I write, like I, I, I look for opportunities for serendipity. Yeah.

[00:17:46] Moritz: That's great. It really reminds me of the work of Kenneth O Stanley and his work on the Myth of the Objective. Have you seen it?

[00:17:56] Erik: No, I haven't heard that one actually. The myth of the objective sounds

[00:17:59] Moritz: interesting. Yeah. It can send you the link later. It's talks about open-endedness. Having an open-ended system that. Explores. And then move on. Moving on to super intelligence. I'd love to hear from you, why you think the concept of self-improving intelligence is flawed

[00:18:21] Erik: John von Neuman, the famous, mm-hmm. A mathematician who was largely responsible for the world's first, not the ENaC, but the vac, the next one, which had stored memory. That was John von Neuman's architecture. So he was also involved in the Manhattan Project, and he came out and said, you can't have.

A self-improving machine unless it's purely random. Because if it's actually planned, then you have to contain the blueprint for the improved machine in the machine and the lower machine. You know, if you, in other words, you, you have to pre, you have to put in the knowledge. In other words, it's not going to think of it because it's a machine.

It has to actually have access to. The components, the blueprint, right? And so he said it's basically impossible to ask a machine to involve like an organic thing. He made that argument in 1950. What's what's interesting though, is. We never let go of it. Like, like we, we still have this idea.

And I, and I'm not sure, I'm trying to figure out a way to say this, where I don't sound like I'm just, you know, pouring cold water. Mm-hmm. But I would turn it around and say, okay, we don't have any, we've never seen a machine. Improve itself and then that machine improves itself and so on. So we, we don't have that observation yet.

So what makes you think that the machine. Two years from now's gonna have it. In other words, there's this mystical idea that the machine is gonna become intelligent somewhere in the future. And I think that's actually just like a religion, like I, you know? Mm-hmm. The idea is that the machine is gonna come alive, and then once it's really smart, then it'll make us smarter.

Copy of itself. But we, we already have like really powerful computers now, and they don't do that. And so what are we missing? Well, they're just gonna get smarter. It's like, what do you mean by smarter? Like more memory. What is exactly you mean by smarter? Right. So I don't think it's a particularly well-founded research agenda to say that you can have a machine, a mechanical, let's be honest, a mechanical device somehow cognate.

A better copy of itself somehow out of the blue, it just comes in. You know, I don't think that's a particularly well-founded idea, actually. It's a very persistent idea. And I sort of, with, I'm not trying to be cynical, but I kind of look at that as like, there, it's a, it's, it's a, it's a mythological you know power within people to, so just something real, like all the way back to Frankenstein.

Remember Mary Shelley, you know, this thing came alive. We made it with science, we made this thing that's alive. And I think like, That just won't go away in the human psyche, but it has nothing to do with science or computer science. Like, I could be wrong.

That's, that's my take. Yeah.

[00:21:21] Moritz: Yeah. No, I think it's a similar point to what you, what we discussed earlier that you, it requires some amount of understanding of a thing to be able to. Predict its improvement. And so the machine itself would need to have a really deep understanding of intelligence itself to be able to like, yeah, improve its own code and like how do we get this?

[00:21:43] Erik: Understand it's the, but it's the dumber thing. How does it have it, right? Like there's a kind of contradiction almost. I tried to treat that in the book, kind of, sort of, you know, tongue in cheek with the. The woman who said, I'm gonna build this, you know, and it's like, well, you know, you either have this idea for how that it's, it's almost, it gets back to the, the point about the wheel, right?

You either have the blueprint or you don't. And if you don't have it, you're not gonna make something more intelligent because you don't know what even means to make. It's like, so I'm gonna ask you, Maurice. Yeah. I'm gonna ask you, Maurice, can you make a more intelligent copy of yourself? You're pretty smart.

How would you even go about doing that? Because if you had that, you would just become more smart. You would say, okay, here I am. You know, and so it, it, there's something very, very slippery about that entire discussion. It's not quite hitting the ground. Although I

[00:22:33] Moritz: think if I would need to steelman the other side one way could be sort of gene engineering.

Once we understand the Yeah, like iq relevant genes sufficiently, maybe we can like modify that and then we are a bit smarter, but. To assume that that goes like up to the right is also a bit farfetched, I guess.

[00:22:53] Erik: Yeah, yeah. I mean, we'll, you know, there are surprises in ai. I think. I think G P T was a, you know, chat, G P T was a surprise.

We didn't see that coming. So you, you can't say never, you know, never say never. Right. But you can see theoretical reasons why some of this discussion is just, There, there's, you know, the feet aren't on the ground, you know, the sun, you know, it's not, it is very difficult to see how you have a self-improving machine.

It would revolutionize our entire idea of what a machine was if we had that. And so, you know, who knows? Never say never, but I wouldn't Hold your breath. Yeah.

[00:23:31] Moritz: I'd love to go back to your previous point about your new book. Can you tell more on what you plan to write about, or is it still in the exploration phase?

[00:23:40] Erik: Yeah. I, I can't say too much because we're still pre-contract with the publishers, but right? I'm, I'm basically arguing that I need to, I need to work on my elevator pitch because I haven't thought about, I haven't, I haven't thought about how I want to sell this book. That the 21st century is a wash with futuristic ideas.

And there's this general kind of cheerleading that we're, you know, on a rocket ship to progress, but it's actually a very un-innovative century. We're basically just making new versions of cell phones, you know, and it's like we actually, the 1880s were vastly more innovative, than the two this century so far.

So I think like, and I don't say that because I want progress to slow down. I actually want to accelerate progress, right? So if you have bad ideas and misinformation, you're not gonna make progress and you, it's not enough just to say that we're, you know, exponentially progressing. You actually have to exponentially progress.

We don't have a lot of innovations, I mean, CRISPR is a four or five-generation downstream result of the discovery of DNA in the 1950s. By comparison, the 1950s were like a superhuman decade, right? We discovered DNA in the 1950s and this century we discovered a way to, you know, manipulate genes. It's a very downstream innovation, and we don't have, we no fundamental innovations are happening.

So I'm

[00:25:20] Moritz: What do you think is the, yeah, what do you think is the, the main like reasons for that? Is it regulation or self-delusion?

[00:25:29] Erik: I think we're just back in big business. You know, we're back in, in big business. We consolidated a lot of you know, the web started out, it was this big, you know, Power to the people decentralized platform.

Everybody's gonna be a citizen blogger. Everybody was gonna release their creativity. And by about 2009 we had just big, giant corporations like, you know, Ford Motor Company, Google, you know, and they just shut everything down and they're, you know, it's the same story. American business goes through these cycles.

I think those companies are just locking in their own profits and they have no interest in someone coming along with a legitimate. Innovation because that might actually, you know, bankrupt them, right? Mm-hmm. So basically, you know, we're just under the grips of big business, you know, the, you know and honestly, the AI we have is big, big iron, big tech, big business, you know, hyper funded, you know, so the whole culture has just moved into this really, really draconian sort of, you know, top-down phase.

And yet everybody, we're all still on Twitter saying about how we're so innovative. And to me, it's just like, you guys are just nuts. You don't even know what's going on. Like, it's just not innovative at all. And you're basically being controlled by these large corporations. It's time to get real, you know?

So yeah, I, I want to just, I want to get a different discussion going, you know, that that's my, that's my goal. Makes

[00:26:56] Moritz: sense. Yeah. And then with this current technology as mentioned sort of. There, I think are potential interesting ways to combine these induction based, like deep mo deep learning models with some of the symbolic approaches.

Have you come across any interesting work there recently or in the last few years? Yeah. There's

[00:27:16] Erik: a guy Pedro Domingo, who I think he's still at the University of Washington, but he's trying to use. Machine learning techniques to basically bootstrap like a more persistent knowledge graph.

So I think like, I don't know if that's gonna be a breakthrough or not, but we, we have two really well-established areas of research and ai. Knowledge. We, we used to call it knowledge, representation, and reasoning. Now they call it knowledge graphs, right? But the idea is that you have a persistent structure that you can reason over, but that structure has to sort of be informed by data.

So you take machine learning, you take a persistent structure, and then you get so something more powerful than a neural network. I think there's, there's room for that. And I think the, the field will start grabbing. Different stuff because, you know, you know, like, like Altman himself said like, we can't make g p T 100.

You know, at some point the game is over. It's just, you know, it was fun. It was a night neat trick, but. It can't continue. We have to find a different path forward. So I'm all for hybrid stuff. I'm all for combining stuff, but I don't see, to answer your question more directly, I don't see anybody that's really cracked that nut yet.

Yeah, right. We would know about it, right? It would be all, it was splashed all over the New York Times if they did and, but you know, I think that's the right idea is to sample from different sources and try to, you know, it's kind of the similar as this serendipity point I was saying earlier, right?

Mm-hmm. Go outside the box and just see what fits and what doesn't fit. And you know, there's an Einstein, not me, someone that's gonna see, Hey, wait a minute, you know? And there we go. There's a new idea in human culture. So, yeah. You know, experiment. Very interesting. Yeah. Yeah.

[00:29:10] Moritz: No. So just taking the current technologies we have without sort of taking into account like potential future breakthroughs, what sort of applications, what problem do you think are interesting to solve that ha we haven't solved?

[00:29:23] Erik: Yeah, so I think we're kind of back watered on robotics and Autonomous navigation. I think not. I, I really wanna see people put more effort into that. I think conversational AI took a huge step forward with large language models. And so I think we're gonna see really realistic dialogue. It's really almost, it's almost like solving the Turing test, right?

I think we're gonna see really realistic dialogue between a human and a machine. I think that that, that's really gonna push forward with large language models. I think that made that possible. We need more people thinking fundamentally, like what makes this system work and why does, why is it limited?

And like, sort of why do, why don't self-driving cars work? I, I would, I want to ask everyone right now, all your right priorities. Why don't self-driving cars work? Level five, fully autonomous. Why doesn't that work if AI is so smart? And it's not, it's not about, you know, this is an Irish expression, taking a piss.

You know, it's not about saying, it's not about saying like, there's something wrong with ai and I want, you know, it's, it's about saying like, we can't make progress if we don't identify what's wrong. So why don't, why doesn't it work? Do you have an idea why self-driving cars don't work level five?

[00:30:44] Moritz: I, I think a big aspect here is like the long tail problem of the weird situations that, that wasn't in, that weren't in the data.

Yeah. And there's some attempts to like synthesize the data and like generate a bunch of scenarios, but there will always be limited by the creativity to come up with these scenarios.

[00:31:01] Erik: So so, so is that a, is that an existential proof against. Agi, the fact that there will always be a long tail of, you know, anomalies on the open road.

Like, is that, is that actually an existential proof against agi? I, I think it might actually be, to be honest with

[00:31:18] Moritz: you, I think it is a proof against deep learning based ai.

[00:31:23] Erik: Yeah. Very smart. Yeah, very good. Yeah. It's a, it's a, it's a, it is a, it's a proof against one aspect, one approach to ai. I, I would, I'll go there with you.

Yeah, definitely. Cool. I'm also

[00:31:34] Moritz: interested in sort of the, your, your stance on the automation of jobs. I'm not sure if you've seen my article, but I think there's a distinction you can make between automation and augmentation. But yeah. I'm curious, do, would you agree with these people that say like 80% of all the knowledge worker jobs will be automated?

And do you think it's true to some extent

[00:32:00] Erik: the media has been saying that since Ford Motor, like since you know, Ford Motor Company. Right. You know, and so it's just a perennial, it's a kind of Marxian idea. That, you know advanced technology will displace and alienate workers to some extent.

Workers are alienated. I mean, to some extent if you're in a modern car factory, there's not a hell of a lot you can do. You know, you plug into a system, you're pretty much like flipping hamburgers, you know, and it used to be that you were actually building the car. So that kind of. Art artisan, like that sort of, you know, craftsmanship is kind of gone.

So I think that does alienate the workforce. And you see that everywhere in healthcare. Doctors constantly complain that they have to key in all this crap and they can't look at the patient, excuse me, like you just talk, you know, in my ear and all key it in. They used to actually, you know, make a connection.

So I think in all these industries, you, you can make a distinction between displacement and alienation. Hm. And I think we do have alienation from automation, but in terms of displacement, we tend to create more jobs. Automation actually creates jobs, right? So you don't see, I don't think that this trope of.

We're gonna get more and more computation and then less and less human in, in the workforce. I don't think that's really the point. Awesome.

[00:33:34] Moritz: Yeah. Is there anything else, Eric, you wanna

[00:33:35] Erik: share? Not really. I just, wish you the best of luck on the podcast and stay in touch.

[00:33:42] Moritz: Likewise. Yeah. Excited for your new book.

This is definitely one of my favorite books. The Myth of Artificial Intelligence. So, Yeah, keep us up to date and just by chatting.

[00:33:53] Erik: It was wonderful. Thank you very much for the opportunity. I really appreciate it, Mors. Have a good one. You bet. Thank you.

Created with ⁠wondercraft.ai⁠ (get 50% off with SCALINGKNOWLEDGE50)

Keith Rabois

• Twitter: ⁠twitter.com/rabois⁠

• Wikipedia: ⁠wikipedia.org/wiki/Keith_Rabois⁠

Mo Golshan

• ⁠OG written content version⁠

• Mo's twitter: ⁠twitter.com/mogolshan⁠

Mentioned links:

• Interview with John Sculley on Apple and Steve Jobs: ⁠John Sculley On Steve Jobs, The Full Interview Transcript | Cult of Mac⁠

• Read The Cognitive Style of PowerPoint by Edward Tufte: ⁠http://www.edwardtufte.com/tufte/powerpoint⁠

• Achieving Business Success One Person at a Time* by Peter Engel: ⁠The Exceptional Individual: Achieving Business Success One Person at a Time by Engel, Peter⁠

• Advisory board: ⁠Should Your Startup Have an Advisory Board? | by Mark Suster | Both Sides of the Table⁠

• Essay on hiring great PMs: ⁠How to Hire a Product Manager: the Classic Essay by Ken Norton⁠

• What is more important for a start-up: quality of team or size of market? ⁠pmarca archive.posterous.com/the pmarca guide to startups part 4⁠

• How to Raise money before launch? ⁠How to raise money before launch. by Delian Asparouhov⁠

• Should a student finish college or go work for a startup if given the chance to work for a YC, TechStars alumni? ⁠A Student's Guide to Startups⁠

Content:

(00:00): Intro

(01:00): Navigating the Ambitious Problem-Solving Journey: From Start to Success

(28:34): Navigating Startup Operations: From Team Building to Executive Hiring

(28:58): "Lessons on Assembling Your Founding Team: Complimentary Co-founders, Cult Creation, and Problem Solving"

(01:41:31): "Exploring Venture: Lessons on Assembling Founders, Succeeding in Early Stage Investing, and Venture Capital"

(01:41:53): "Assessing Extraordinary Founders: Traits, Strategies, and Unconventional Qualities"

(02:34:31): "Productivity and Life Advice: Strategies for Success in Venture and Beyond"

(02:34:50): Contrarian Thinking and Managing Time: Lessons for Success

(03:06:20): Outro

⁠

Brian’s Twitter: (@psychosort)⁠

Brian’s blog: From the New World⁠

Content:

(00:00) Brian's background

(03:20) On writing

(07:05) AI doomers: steel man and critique

(14:05) Job displacement

(19:20) Signaling jobs

(28:20) Future economics of art

(31:10) Qualitative v. quantitative moats (Google doc leak)

(37:00) Midwits and conformity selection

(40:00) How to fix Midwits phenomenon (Civil service reforms)

(42:00) Why the free market selects against the free market

(46:00) Problems with regulators (FDA and co.)

(52:00) The root of Altruism

(55:00) Origin of envy

(55:25) Conspiracy theories and egalitarianism

(1:00:00) Egalitarian murders

(1:06:45) Polarization and obstructionism

(1:09:05) Obstructionism and regulation

(1:10:25) Issue polarization, affective polarization, and partisanship

(1:12:00) Education system and the Freddie deBoer case

Arguments by analogy are fallacies. Almost any analogy between any two things contains some grain of truth, but one cannot tell what that is until one has an independent explanation for what is analogous to what, and why.

Most phenomena are not the same. History doesn't repeat itself. Assuming an analogous relationship requires a good explanation as to why the phenomena are analogous.

Mental Models

This is why many “mental models” are error-prone. One claimed benefit of mental models is that they are easy to remember and “simplify the complex into understandable and organizable chunks”. 

Because they are not based on hard-to-vary explanations they often achieve the opposite. Some theories are better than others. Better in which ways you might ask? They can be superseding theories, simplifications, or unification of existing theories. Michael Faraday’s and James Clerk Maxwell’s combined theory of electromagnetism is easier to remember than their separate inferior theories of electricity and magnetism. Arabic Numerals are easier to learn than Roman numerals.

Let’s look at an example from Shane Perish, who helped popularize the term of mental models, writing several books about the topic1: 

Inertia: An object in motion with a certain vector wants to continue moving in that direction unless acted upon. This is a fundamental physical principle of motion; however, individuals, systems, and organizations display the same effect. It allows them to minimize the use of energy, but can cause them to be destroyed or eroded.

This is making an analogy between Newton’s First Law: Inertia (“An object at rest remains at rest, and an object in motion remains in motion at constant speed and in a straight line unless acted on by an unbalanced force.”) and “status quo bias” (a concept that encapsulates various biases such as familiarity heuristic, loss aversion, and the sunk cost fallacy and the behavioral tendency to reduce cognitive load). Both cause something to “remain how it is”. But is human psychology analogous to objects flying through space? No. At least not without an explanation as to why they are analogous. The separate concepts are a hard-to-vary explanation that cover more territory, are more universal, and are thus superior to this muddled mental model.

Technology Evolution (AI Risk)

Flawed AI doomer/safetyist arguments are based on the assumption that the development of a artificial intelligence is analogous to evolution. As I wrote in my piece on AI augmentation vs automation:

The Scaling Laws hypothesize that LLMs [Large Language Models] will continue to improve with increasing model size, training data, and compute.

Some claim that at the end of these scaling laws lies AGI. This is wrong. It is like saying that if we make cars faster, we’ll get supersonic jets. The error is to assume that the deep learning transformer architecture will somehow magically evolve into AGI (gain disobedience, agency, and creativity). Professor Noam Chomsky also called this thinking emergentist2. Evolution and engineering are not the same.

Until they articulate a hard-to-vary explanation of how current AI technology development is analogous to biological evolution they are basing their argument on the supernatural and unexplained.

Coda

Mental models can be useful to get a basic understand of a phenomenon. But that’s where their benefits end. They don’t help you to achieve a deep understanding of reality.   

When arming yourself with concepts that help you to solve problems in the future, side with the deepest most universal theories and discard mediocre self-help books trying to re-package content under a new name.