Nor were they designed as “software development engines,” despite their increasing use in this domain. Every tool or piece of software that uses LLMs for code generation feeds instructions to the LLM — invisibly to the user — that begin with the phrase: “You are an expert software engineer” (or something nearly identical to this).

With no more text to consume, and with fewer sources for their insatiable need for energy, LLMs are seeing the bulk of their improvements these days coming from the tools we’re building around them. This includes “prompt engineering” (knowing the secret grammar that will get them to do what you want), multi-agent workflows (LLMs using other LLMs, ad infinitum), etc.

LLMs’s proficiency with facts — yes, they screw up sometimes, and as such we shouldn’t rely on them for critical matters, but their fact-retrieval is astounding — has persuaded us that behind them is a general artificial intelligence, understanding our questions and rifling through its database of facts to give us answers.

As a result, first-time users of an LLM for code generation will often be perplexed at how poorly it understands their existing code, learning shortly thereafter that providing even a few bits of documentation dramatically improves the results — demonstrating just how helpless these things are for most tasks without our assistance.

Recently, Vercel, a cloud computing company, released an AGENTS.md file (a text file standard for coaching LLMs, written in plain English, with intermittent code examples) for writing code in React, a popular JavaScript framework for building complex web applications. This file is nearly 10,000 words, with an estimated reading time of half an hour.

If you take nothing else away from this blog post, remember this: for all the hype around AI, the state of the art in 2026 includes writing an entire book about React, and telling LLMs to read it before every interaction.

This has had me thinking about what an LLM is, exactly. What do we call something like this?

It may be easiest and most common to refer to it as a tool, but this isn’t quite right. A tool is designed, sometimes exquisitely, to achieve a specific purpose. LLMs aren’t that. Just about all their benefits to us are accidental.

Instead, a better metaphor might be to think of LLMs as a resource that we harness rather than a tool that we use. A resource in the sense that electricity, for instance, or the wind is a resource.

The wind is powerful and useful, but aimless and indifferent. We use the wind to power turbines, to fly kites, to sail. Sailing evolved and improved not because the wind got better at propelling our boats — it wasn’t designed to do that — but because we built better sails and developed better sailing techniques. And this is what we’re doing with LLMs today — building better sails, developing better techniques.

I’m reminded of Steve Jobs’ quip about Dropbox: that it’s a feature, not a product. I see echoes of this in Apple’s reluctance to turn Siri into a chat UI, perhaps because they’re seeing it in the same way.

Kagi just sent me three invitations for three free months of Kagi. I forwarded it to one friend who I thought might be interested. But when I tried to think of a second and third, I was at a loss for names.

For a moment, I imagined being any one of them, imagined how it must sound coming from me to suggest to them that they pay money for something that has been free their whole lives.

And then I pictured explaining to them that not only will they have to pay, but they will have to install browser extensions on any device on which they intend to use it. They’ll have to log in on all those devices, too. And I realized that would probably be more of an obstacle than the subscription fee.

Convenience can be a greater factor in a purchase decision than cost. For Kagi to be successful, it would have to be very nearly as easy to “set up” as Google. “Setting up” Google, of course, requires no effort, as it is the default search engine on the vast majority of browsers in use.

Asking someone to pay money for the privilege of managing and logging into a browser plugin — that’s a tall order.

One of the first things we started doing when ChatGPT became public was, naturally, asking it questions. And for the most part, it gave us some pretty good answers.

But as has been widely demonstrated recently — as Google, Meta, and others have begun grafting “AI” onto their search results — is just how wrong it can get things, and it has had us asking: Is this really ready for widespread adoption? Asking for arbitrary bits of information from an LLM’s entire corpus of text — like Google’s and Bing’s smart summaries — is demonstrably, hilariously, and sometimes dangerously flawed.

Over the last couple years, I haven’t really heard much from OpenAI themselves about what we are supposed to be using ChatGPT for. They seem more interested in creating the technology — which no one could seriously doubt is impressive — than in finding real-world applications for it. The announcement for ChatGPT didn’t tell us what to do with it (though it did emphasize that the tool can be expected to product false information).

I think the misconception about what ChatGPT is purported to be good at can be attributed to the name and the UI. A chat-based interface to something called “ChatGPT” implies that the product is something it isn’t. It’s technically impressive, of course, and makes for a good demo. But chat doesn’t play to its strengths.

The reason any given LLM is even able to wager a guess at a general knowledge question is the corpus of text it’s been trained on. But producing answers to general knowledge questions is a side-effect of this training, not its purpose. It isn’t being fed an “encyclopedia module” that it classifies as facts about the world, followed by a “cookbook module” that it classifies as ways to prepare food. It was designed to produce believable language, not accurate language.

Where it does excel, however, is at coming to conclusions about narrow inputs. Things like Amazon’s review summaries; YouTube’s new grouping of comments by “topic”; or WordPress’s AI Feedback — these take specific streams of text and are tasked with returning feedback or summaries about them, and seem to work pretty well and have real utility.

These examples demonstrate two similar but distinct facets of LLMs: Their role as general knowledge engines, or “oracles,” and as input processing engines, or “analysts.” When we ask ChatGPT (or Google, or Meta) how many rocks we should eat per day, we are expecting it to behave as an oracle. When we ask it to summarize the plot of a short story or give us advice for improving our resume, we are expecting it behave as an analyst.

Signs point to Apple using LLMs primarily as analysts in the features to be announced at today’s WWDC, processing finite chunks of data into something else, rather than snatching arbitrary knowledge out of the LLM ether.

The allure of ChatGPT as an oracle is of course hard to resist. But I think if we recognize these two functions as separate, and focus on LLMs capabilities as analysts, we can wring some value out of them. (Their environmental impact notwithstanding.)

Amazon, on the other hand, since they had gotten into the original content market, would be incentivized to give priority to their TV shows over others. Is my Amazon Fire TV recommending Sneaky Pete to me because it’s similar to stuff I watch, or because they stand to gain from it?

When Apple TV+ was announced, I worried that Apple’s having a stake in what specific content I watch would taint the Apple TV experience.

Boy, has it.

Below is a full view of the contents of the macOS TV app. The TV app on the Apple TV streaming box is roughly identical.

Blocks highlighted in magenta are Apple TV+ or Apple-owned content. Blocks highlighted in blue are third-party content that requires purchases or subscriptions from Apple.

The top hero carousel, shown here displaying Killers of the Flower Moon, has twenty-five slides in it. The first nineteen are for Apple TV+ content.

(Post continues below.)

What exacerbates matters is that much of this content relies on a subscription rather than a one-time purchase. If Apple had gotten strictly into the movie industry, for instance, and offered their movies as purchases rather than via a forever subscription, I think this incursion wouldn’t have been quite so aggressive and insidious. Subscriptions, or “services,” are where all the money in tech is now. Get the customer in the door, and keep them paying. Streaming now makes up the vast majority of digital music revenue, as is evident in the languishing of the iTunes app and ecosystem.

Subscriptions are making Apple so much money that they have resorted to user-hostile behaviors like what has happened to their TV app.

Say what you want about Apple — their products are too expensive, they nickel-and-dime you on accessories, everything is harder to repair — but it used to be that once you owned whatever product it was that they were selling, the experience of using it was unparalleled in user-friendliness and delight.

This is no longer the case. Splurge at the Apple Store, open the box at home, turn the thing on, and you will only be pestered by relentless solicitations.

Elon: What this advertising boycott is gonna do, it’s gonna kill the company. And the whole world will know that those advertisers killed the company. And we will document it in great detail.

Andrew: But those advertisers, I imagine, are gonna say, “We didn’t kill the company.”

Elon: Oh yeah? Tell it to Earth.

Andrew: But they’re gonna say, Elon, that you killed the company because you said these things and that they were inappropriate things and that they didn’t feel comfortable on the platform.

Elon: And let’s see how Earth responds to that. We’ll both make our cases and we’ll see what the outcome is.

Andrew: What are the economics of that for you? You have enormous resources, so you can actually keep this company going for a very long time. Would you keep it going for a long time if there were no advertising?

Elon: If the company fails because of an advertiser boycott, it will fail because of an advertiser boycott. And that will be what bankrupted the company, and that’s what everybody on Earth will know. Then it’ll be gone. And it’ll be gone because of an advertiser boycott.

Andrew: But you recognize that some of those people are gonna say that they didn’t feel comfortable on the platform. And I just wonder, and ask you, and think about that for a second–

Elon: Tell it to the judge.

Andrew: But the judge is gonna be–

Elon: The judge is the public.

Something really struck me about the language he was using here. “Blackmail”? “Earth”? “We will document it in great detail”? Blackmail him to do what? What does he imagine this is about?

Of course, one way to read this is that he is shifting blame from himself to advertisers for the potential failure of X, and I’m sure that’s part of it. But he also seems to deeply believe that X is a heroic and final bulwark against creeping Orwellianism from politically correct elites.

To him, the disappearance of Twitter is a threat not just to himself or to the satisfied users of the social network, but an existential threat to “Earth.”

The way this plays out in his mind is that advertisers kill X, and because their abandonment of the platform will have been “documented in great detail,” Earth — Earth itself, not just the users of X — will retaliate. “Free speech” will die, there will be no medium left for expressing yourself without censorship, and the societal cost of this will be so great that these advertisers will themselves face extinction.

This hyperbolic view of things has been on display before, as when advertisers started expressing caution when he first took over in November 2022:

“They’re trying to destroy free speech in America.” Note that he doesn’t believe “free speech” will simply be an unfortunate casualty of the passive cowardice of advertisers fleeing X; he believes that advertisers’ fleeing X is a deliberate attempt to “destroy free speech.” They are invested in the death of X so that they can further their agenda. X is a threat to them, and they are trying to end it.

The “blackmail” he referred to at DealBook is, to his mind, blackmail to join their program of speech suppression. The elites want people to feel afraid to say certain things, and they want X and Elon on their team. By denying them this, he is martyring himself for the preservation of intellectual freedom for our entire species.

He has a messiah complex, and curiously it extends beyond his electric cars and plans to colonize Mars. Even a dumb social network is about his role in the fate of the planet.

I’m also amused by — and I haven’t seen any mention of this — the fact that Elon clearly seems to expect some kind of laughter/applause after his “go fuck yourself” line. He sees himself as a populist, and has gotten so used to being constantly fellated by his X subscribers that it shocks and confuses him when he isn’t showered with adulation for his irreverent “antics.”

Increasingly, it feels like paid memberships for web content are not only a viable alternative to surveillance-driven ad revenue, but one that readers are eager to embrace. The success of Substack demonstrates this. This success is often framed as a preference for reading in the inbox rather than on the web, which is some feat considering how much people have come to loathe email in general over the last couple decades.

But I don’t think it’s the inbox per se that people like; it’s that the inbox gives people what the web used to — and no longer does, but could — give them: an ad-free and distraction-free reading experience. Medium was supposed to do this, but has, perhaps predictably, caved and started showing upsell popups and “related content” sidebars all over its article pages to get your money and keep you on the site.

Patreon, like Substack, has seen a lot of success, but for whatever reason isn’t really thought of as a place for longform writing. Its content tends to be audio, video, art, and access to Discord communities.

All of these services — plus Substack clones like Buttondown and Ghost — require that creators use the associated platforms. This comes at some obvious costs, such as not controlling the means of distribution of your content (although for many, of course, not having to worry about that is a benefit), but also some non-obvious costs, such as having all your content held by a company whose politics or business decisions you may learn later you vehemently disagree with, or unwelcome changes in pricing structures. Migrating this content can be an enormous headache, as the platforms benefit from this lock-in.

(Ghost, because it is open-source and self-hostable, I should add, can be used in a way that doesn’t hold your content hostage on someone else’s servers.)

Paywall-like services that allow you to supply your own publishing platform include Flattr, which takes a monthly subscription and distributes it across the participating sites you visited based on articles read. These sites, presumably, notice your Flattr membership and abstain from serving you ads. For a number of reasons, this hasn’t really taken off, likely a combination of the payout to content creators being pennies, and the unclear value proposition to readers.

For sites running WordPress, Memberful is a popular option. Six Colors, MacStories, and Relay FM, Comedy Bang Bang, and Jason Kottke all use it. For written content, its focus is WordPress, but it does offer an API for people interested in rolling their own integration with their CMS of choice.

But for all the control that Memberful gives you, I think it lacks something that the more ubiquitous Substack and Patreon lack: a unified billing portal.

Memberful’s product page boasts: “If we’ve done our jobs, your members won’t even know our name.” I understand why this might be desirable from a branding angle for a large and well-known organization. But for small creators, this may actually be a detriment.

Memberful memberships are all independent of each other. I subscribe to several sites using Memberful, but I have no single Memberful login. If I find another site I’d like to subscribe to, and they use Memberful, I have to create a new login and supply my payment info again.

On the other hand, if you’re a Substack reader, and you come across a newsletter that looks interesting to you, subscribing is quick and feels safe. You don’t need to re-enter any payment info; the subscription is billed to your Substack account. The writer to whom you’re subscribing doesn’t see any of your billing info: card number, full name, address, etc. If you want to cancel, you know you can do so easily on the Substack account page. And you can see and manage all of your subscriptions in one place. Patreon works the same way.

This is something I’ve grown to love about subscribing to content through the App Store. All of those subscriptions are visible and manageable from a single place on my phone, and I never have to give additional third parties my payment information. Whenever possible, I will choose to subscribe in this way.

What I would love to see, and what I don’t believe currently exists, is something like that, but for independent publishers: some trusted third party that manages subscriptions to many different sites on many different platforms; Substack, but for WordPress/Ghost/Drupal/Craft CMS/whatever. This would have some of the benefits of Substack and Patreon, and fewer of the drawbacks.

Counter to Memberful’s value proposition, I would want readers (were I a small, little known publisher trying to sell subscriptions) to know and trust the third party through whom their subscriptions were being handled. And as a reader, I would enjoy having a central place where all my website subscriptions could easily be managed.

I’d love to know if something like this exists, or if it’s been attempted before. What would some of the drawbacks be? Who would be well-positioned to do this? Why do Memberful subscriptions operate in such a fractured way? Could this be a Substack/Patreon thing, or are they too reliant on lock-in to explore this?

The cases included are:

1. Subjective (“Yesterday, she went outside.”)
2. Objective (“I went with her.”)
3. Possessive pronoun (“The idea was hers.”)

These three cases mirror the common “she/her/hers” structure used to communicate pronouns on social platforms like Twitter and Zoom; as such, they may seem complete, but they aren’t.

The missing case is the possessive adjective: “It was her idea.”

This may seem redundant, because in the declension of feminine pronouns, the possessive adjective is the same as the objective: “her.”

But in masculine pronouns, the possessive adjective is the same as the possessive pronoun: “his” (as in, “It was his idea”).

This isn’t a problem when interacting with people — we know how to decline the common feminine and masculine pronouns, so we know which form to use for the possessive adjective.

But because the iOS 17 UI doesn’t have a field for the possessive adjective, the OS and apps that have access to these fields — which have to behave programmatically — can’t know what to use for that case.

If I’m writing an app and I want the UI to say, “Would you like to call her?” in reference to some contact of yours, I can use the objective pronoun field. But if I want the UI to say “Her birthday is coming up” (the possessive adjective), I don’t have the necessary information.

I can try to guess, based on what I know about the declension of these common pronouns. But what about uncommon ones, like Ze or Xe? Anything I want to use will have to be hard-coded.

This is not even to mention the reflexive case — “She was proud of herself” — which is entirely absent.

The emphasis on both (a) semantic HTML and (b) backwards compatibility means that, as stated by the W3C themselves:

The b and i elements are widely used — it is better to give them good default rendering for various media including aural than to try to ban them.

So: What to do with those letters? <b> can’t mean “bold,” and <i> can’t mean “italic.” What do they mean?

A few years back I noticed that MDN reported that the “b” in the <b> tag stood for “bring attention to,” and the <i> for “important.” Naturally, I checked the spec itself. Neither the W3C nor the WHATWG specs made mention of “bring attention to” or “important.”

I tweeted about this to Eric Shepherd, who — because MDN used to be a proper wiki with page edit history¹ — I could see had been the one to add the text about “bring attention to” to the <b> article on MDN. He responded:

MDN still defines <b> as “bring attention to,” though it has changed its definition of <i> from “important” to “idiomatic.” These terms are still unique to MDN, nowhere to be found in the W3C spec. It’s a shame, because up until this discovery I had assumed — as I believe many people still assume — that MDN is nearly as canonical as it gets. It remains a tremendous resource, but just be sure to double-check with the W3C on its finer details.

¹ It’s now hosted on GitHub, where its history is also visible, but only as far back as when it moved from a wiki to GitHub. It’s also evidently opted out of the Wayback Machine

I’ve been thinking about getting more into “Personal Blogging” elsewhere, and I figured these would be useful in getting words out.

In order to get these prompts on a self-hosted WordPress site, you need to do three things:

1. Upgrade to JetPack 11.7 or higher
2. Add the following line of code to your wp-config.php file:
   

   define( 'JETPACK_EXPERIMENTAL_BLOCKS', true );

3. Go to “Settings > Writing” in WordPress admin, and check the box labeled “Show a writing prompt when starting a new post.”

I don’t know when this is supposed to become available without Step (2), but at that point it might be a good idea to turn off Experimental Blocks by removing that line.

Two days later, another, far more viral tweet joked that the lyrics can be read as the interruption of an act of witchcraft. (This joke has been cribbed in countless subsequent social media posts.)

I was immediately skeptical at the hearing of the lyric as “moon” rather than “mood,” but it was helpfully pointed out to me that the official @PaulMcCartney Twitter account had indeed posted the lyrics as “moon.” (Original tweet link, Wayback archive)

Last year, in December 2021, @PaulMcCartney again tweeted the opening lyrics (backup link), but this time with “mood” rather than “moon.”

I began to hunt more seriously for answers.

That same winter I bought both physical and digital copies of the expanded edition of McCartney II. This album does include the song, but turns out not to have the song’s lyrics in its liner notes.

Unhelpfully, the song was originally released as a 45RPM single, with no printed lyrics.

I discovered that the official Paul McCartney YouTube channel had posted the music video in 2019, and its captions read “The mood is right.”

Liveright published “Paul McCartney: The Lyrics,” a massive two-volume box set of printed lyrics for “154 of his most meaningful songs.” Evidently “Wonderful Christmastime” is not meaningful, as it is not included in this set.

Nevertheless, due to its making much more sense, I had personally concluded that it is, in fact, “mood” (though I never seriously doubted it).

Then, earlier this week, the author of the witchcraft tweet followed up with a notice that @PaulMcCartney has since deleted its 2019 “moon” tweet, and that McCartney himself has addressed the controversy in a recent interview:

No, it’s ‘the mood’. And you know what, I’m thinking about Liverpool Christmas parties, that’s really all I’m doing with that song. “The mood is right, let’s raise a glass, the spirit’s up” – you know, all the stuff you do at Christmas.

Thanks, Paul.