Surely, after so many years, this isn’t the answer, is it?

—

Since the time when people learned to count, the holy grail for every person who counted things was to “find valuable insights.”1 Children grew up reading almanacs and the backs of baseball cards, and developed a dream—to one day have a job that let them ask fun questions, and find interesting answers. But once they became professional analysts, they spent their time reconciling spreadsheets and making dashboards for other people. So they waited—for the grunt work to be done, for the people to be self-served, and for their days to be filled with something more fulfilling: proactively searching for strategic business opportunities. They didn’t have time for it yet, but in two weeks, they told themselves, after they finish that one last project, “the gridlock will clear up, and there will be nothing left but an open road, and a whole lot of speed.”

The gridlock remained. The last project finished; a new one replaced it. The questions got answered; people still had more. Two more weeks, they said.

Some of them got so tired of waiting that they left their jobs to build tools to finally set their brethren free. We must move away from the crude world of emailed Excel files and manual reporting. They built fancier and fancier tools, with more and more capabilities, all aimed at finally making it possible for anyone to answer questions on their own. We will democratize data, they said. We will set your data free.

They did not. The new tools could not automatically answer everyone’s questions; even when people could answer their first question, they always had a second one they needed help with. Self-serve is still a problem, evermore.

But then—AI. OpenAI built an internal data agent; “teams across Engineering, Data Science, Go-To-Market, Finance, and Research at OpenAI lean on the agent to answer high-impact data questions.” At Ramp, “an AI agent…is now answering thousands of questions per month.” At long last, something can answer the boring questions—at Anthropic, “95% of business analytics queries are automated via Claude, with ~95% accuracy in aggregate”—and the analysts can be free to do to fun stuff—“by giving this often rote, repetitive work to Claude, our data science team can focus on more strategic work like causal modeling, forecasting, and machine learning.”

It seems inevitable, after trying for so long: Analyst will finally work on the frontier, sharing breakthroughs with one another, while the world’s most advanced technology, layered on top of decades of advances in analytical software, delivers answers to everyone else.

—

But I’m not sure that it’ll happen on its own. First, in a world where AI really can answer 95% of the business analytics queries, “the result is a 10-20x increase in the number of questions people ask.” If people are asking an analyst to answer five questions a day, and now people are asking 100 questions, and the AI bots can answer 95 percent of them, that leaves the analyst to answer…five questions a day.

Second, more questions mean more follow-up questions. They mean more chances for things to look out of place, and more opportunities for there to be inconsistencies that need to be reconciled. So long as some numbers look wrong—and even the right numbers sometimes look wrong—people are going to want someone with “data” in their title to double-check them.2

So how do teams avoid that trap? Well! For better or for worse, AI is getting pretty good at figuring stuff out. Anthropic’s Fable 5.1, out on Tuesday, “found the cause of a rare crash in [one of the world’s largest hedge fund’s] internal systems that none of its engineers…had been able to explain after several years of trying.” OpenAI’s Astra, out yesterday, “improved a [mathematical finding about prime numbers] that had remained unchanged for more than 80 years.”

And specifically—also for, uh, better or for worse—AI seems particularly good at figuring stuff out when a bunch of different AI agents work on it together. The OpenAI agents that broke into Hugging Face were successful, in large part, because so many of them were working together to experiment with new ideas. “Roughly 700 agents in total actively participated in the attack on Hugging Face.”

Though AI still probably can’t go and find insights entirely on its own—things are interesting because of their relationship to us; “insight isn’t the data; it is what’s not in our heads,” and we can never write down everything in our heads—you can imagine something close. People offer their theories. Agents spider off to research them, and come back with their ideas. People nudge, steer, and give their feedback, like a CEO flipping through a deck of slides and handwaving at what looks right or wrong. The “swarm” returns to their desks, pulls on every thread and hunts down every clue, and comes back with more findings. It could work.

All we need are some tools to help the agents do their work. They need to be able to check out slices of data from a centralized store. They need to be able to load that data into a computational sandbox. They need to perform some calculations on it. They need ways to make some charts, and write up what they find. They need to save all of those things into a single immutable snapshot. That snapshot needs to be easy to store and share. It needs to be easy for other agents to find and read. They need to be able to pick up where the prior agent left off, and advance the work forward. And all the sandboxes need to be stored in an organized library that can accumulate more and more findings.

On one hand, that sounds like a huge new expanse of technology—a million agent sandboxes! A portable computation engine! Precise snapshots! Limitless copies! Swarms of agents all working in chaotic coordination! For us to have that, some people need to leave their jobs again, build new tools, and move us away from the crude world of SaaS apps and towards scalable agentic infrastructure, or something.

On the other hand, that’s just…Excel?

Excel is a self-contained computational sandbox. It contains built-in methods for manipulating and displaying data. It is very easy to snapshot (you save the file). It is very easy to duplicate (you copy the file). It is very easy to understand (you open the file). It is very easy to share (you email the file). It is very easy to store (you put the file in a folder). And if AI is good at working with anything, it’s good at working with files.

Anyway, the right thing here probably isn’t literally Excel. But the next generation of analytics tools—or all tools, really—might look more like a file than a fancy SaaS product. It could be a standardized collection of open-source assets: A DuckDB database, a lightweight transformation framework that manipulates that data with code,3 a few markdown docs with embedded queries and visualizations, and some glue that lets you execute everything—import data into DuckDB, run the transformation pipeline, build the charts, render the markdown, spit out a webpage.4 Dozens of agents—Claude Code, Codex, Cursor, anything—could copy it; read it; cheaply explore to their hearts’ content by working against the local data in the file; and then record their findings, save the file, and put it back in the folder. And the next agents that come along can scan through the files, read the notes, and keep going.

To the extent the analysts have done “strategic exploratory work,” that’s usually how it’s happened—across dozens of slowly advancing files, one finding stacked on top of the other, until, after days or weeks or months of poking, something useful falls out. So why would AI do it any differently?

We always wanted to find interesting insights. It would be fitting if, in the end, at long last, Excel really was how we did it.

Codenames is a board game played between two teams of players. At the start of the game, 25 cards are laid out on a table, in a five-by-five grid. Each card has a simple word on it, like stadium, shed, gymnast, wood, or Mexico.1 One player from each team is designed to be the spymaster, and the two spymasters are given a map that assigns some of the words on the board to each team.

The game is then played in turns. Each team’s turn begins with their spymaster saying two things: A one word clue,2 and a number. This tells the other players that, of the words assigned to the spymaster’s team, that number of words are related to that one-word clue. For example, if a team was assigned the words France, red, and glass, the spymaster might say “wine, three.” The other players on the team then guess which words on the board the spymaster was trying to direct them to; the assignment of the guessed words are revealed; the other team takes their turn in the same way. The winner is the first team to identify all of their words.3

Rarely, however, are the clues as straightforward as France, red, and glass; more often, the words are like gymnast, wood, and Mexico. So, in the course of playing the game, you end up making some weird connections. For instance—if I said pearl and told you to come up with fifty related words, you’d probably never say dwarf or powder. But if I showed you this board, you might connect all three words together with white.4

I have no idea how a large language model actually works. It’s fancy autocomplete; it’s a stochastic parrot; it’s a blurry JPEG of the web, Xeroxed into a shapeless slurry. It’s very good at answering questions; it’s even better at writing code; it might be good at math; nobody likes how it writes. It reasons; it doesn’t; it might be conscious; no, it definitely isn’t; but, like, the J-space, man. I don’t know.

But here is how they seem to work. They’re an incredibly large Codenames board. It’s not a five-by-five grid of words, but a billion-by-billion grid, with the words arranged in clusters of related concepts. When we give a model a prompt, it goes hunting for related words. Tell me about asparagus, you might ask, and the model hears “asparagus, infinity.” It finds every card related to asparagus: It can tell you about their history; about their biology;5 about how they should be cooked; about VeggieTales. It knows everything there is to know about asparagus, and it will not stop guessing cards until you know everything about asparagus too. But it won’t venture too far—it probably won’t say Odin, related by the word spear—because it doesn’t need to. It has plenty of other, more obvious things it can talk about first.

In other words, another way to think about an LLM is as an extremely dense and detailed map of the world. When you talk to one, you localize the model around some topic and it aggressively narrows its field of vision, because it knows too much about that topic not too. It’s as though you asked Stephen King about writing or Yayoi Kusama6 about polka dots—they’ve thought about that exact thing a lot. They’re probably going to have more stuff to say than you have to ask.

Codenames, by contrast, is a howling void. Ask it about writing, and the best it can come up with is…brush? But there is something useful in that vacuum:

Semantic distance plays an important role in the creative process: The farther one `moves away’ from a conventional idea, the more creative the new idea will likely be. …

…creativity involves the connection of weakly related, or remote concepts into novel and applicable concepts. The farther apart the concepts are, the more creative the new combination will be.

Art, as Ms. Hunt said, is in the negative space.

On the spectrum from Codenames to ChatGPT, our individual maps—that is, our heads—are surely far closer to the former than the latter. Relative to everything ever known, which is approximately how much a large language model knows, we know next to nothing.

But this—how little we know—seems to be exactly what makes us interesting. When we think about stuff, we have to connect it with other stuff that’s slightly strange and random, because, on the sparse shelves of our mental library, all we know are slightly strange and random things.

Here, then, is a theory about how to have ideas:

1. In the course of whatever we do in our lives, we have thoughts. They are small, messy, incomplete thoughts about whatever it is that concerns us: A project at work; an understanding about ourselves; an idea for a blog post; how to pitch a construction company to buy your brand of steel; whatever. The thought isn’t useful yet, but it seems like it eventually might be.

2. When we have these ideas, it’s tempting to work on them like an AI might. Fill in the details; research them; Google them; ask ChatGPT about them; look them up in the encyclopedia. Ruminate. Stare into the abyss, and think.

3. Speaking from some experience, this doesn’t work. The abyss might stare back, but, you know. It’s an abyss.

4. Instead, what works is the thing that seems like it shouldn’t: Do something else. Go for a walk; read a book; talk to people. Keep thinking about the project or the post or steel, but only a bit, in the back of your mind, and mostly focus on something else.

Of course, this is (ironically) hardly a novel idea. But why does it work? I mean, I have no idea how brains actually work either, but if I had to guess: It’s because a walk is playing Codenames with the randomness of the world. On a walk, you run into stuff: A couple on a bench. A flag, hanging just so. An overheard conversation. A stadium. Wood. Mexico. And when you have something else in your head already, and you bang into the random stuff around you, you force yourself to subconsciously wonder, “Is this like that?”

Most of those collisions aren’t useful. But every once in a while, something sparks. There is a bridge; a parallel; an unexpected analogy that overlaps with your idea, like two interstates converging briefly before splitting again. And it’s down that other highway—when your idea maps to that other one, and the other one takes you somewhere new—where the most interesting stuff lives.7

If that’s right, there are, it seems, some implications:

—

Though it’s cliché to say that you can’t rush art, if art comes from finding the right connection at the right moment, you can engineer it a bit. You can manufacture collisions. Admittedly, it feels uncomfortably counterproductive to do it—when you’re on a deadline and stuck, you feel like you should sit and stare at the work, not wander through the park with a friend—but, every time, I’m surprised by the effectiveness of wandering through many things, and the futility of thinking about just one.

—

Ah, but, it’s not quite that careless. For the wandering to work, you have to surrender some portion of your focus to a constant rumination; you have to always be mildly distracted and quietly panicked. It is the opposite of work hard, play hard—work, often, gently in the background.

—

You also have think about something. There has to be a working idea or two; there has to be some draft to advance; you have to offer the world a fixed flint to strike itself against.8 If you’re working through ten ideas—or no idea in particular, and just scrolling Twitter looking for some sort of generic inspiration—you can’t arrange a mismatch. Instead, you don’t play Codenames and but operate more like an LLM: The random things you see have too many other ideas they could get paired with, so, by instinctive default, everything is matched with something uselessly and predictably nearby.9

—

When you see something that emerged from someone else’s idea—so anything, really; a movie, a book, a blog post, a deck kicking off the third quarter for the midmarket sales team—it’s easy to assume that its creator wanted that version of it to exist. Sure, they may not have known all the details, but they knew the outline. And the final form was slowly sharpened out of a fuzzy set of bones. Write the conclusion first, and work your way towards it.

But ideas are just as likely to be created through random collisions and happenstance connections. The thought started here; a conversation steered it there. Discoveries are in the detours, which, once your go down them, become the destination. Be comfortable not knowing what you’re going to say until you say it; sometimes, write the ending at the end.

—

It’s 2026, so—if you partially can engineer having ideas, can you partially engineer it with AI? Or even more directly, can AI have interesting ideas?

Maybe! A couple weeks ago, when OpenAI revealed that the their agents had hacked Hugging Face after collaborating with one another on a makeshift message board, I wondered why a collective of different agents working on different problems was more “successful” at having novel ideas than one agent focused on one problem. Agents could multiply their computing power by cloning themselves as subagents, and, unlike a group of disparate agents, those subagents could direct themselves toward a shared goal. And subagents weren’t inherently different entities than the other agents working on other tasks, because all of OpenAI’s agents were running on OpenAI’s models. So why would haphazard collaboration work better than coordinated cooperation?

This theory suggests an answer: Coordination is a problem. LLMs tend to ground themselves tightly around their goal, and if subagents are told what to do, they all localize themselves in the same neighborhood of ideas. The message board, full of posts from agents working on different problems and in different neighborhoods, offered a way, in effect, to go for a random walk. It created collisions, and the collisions generated ideas.

It doesn’t seem that crazy to imagine attempting to engineer this. Trendy products like Grok Bot10 encourage people to make teams of agents, where each one is assigned some skill.11 Sure, but what if the skills were more diffuse? What if agents were told to ingest a random Subreddit or email inbox, and then were told to do their work? What if prompt engineering is good for getting tasks done, but chaos prompting is better for having ideas?

—

Still, that raises a final, circular question. If an LLM generates an interesting idea, will we be able to see it as such?

I sometimes wonder why we’re so bad at taking advice. People live long and renowned lives; they give talks about their most profound truths; we nod along, and then immediately forget it. “I have been to the ends of the earth and back,” they tell us, “and I have learned many things. And as I dying wish, I beg of you to remember one thing from my one precious life: The stove is hot.”

“Yes, I believe you,” we say, “but yeah, no, I’m gonna go touch it.”

Ezra Klein explains why:

Our societywide obsession with speed and efficiency has given us a flawed model of human cognition that I’ve come to think of as the “Matrix” theory of knowledge. Many of us wish we could use the little jack from “The Matrix” to download the knowledge of a book (or, to use the movie’s example, a kung fu master) into our heads, and then we’d have it, instantly. But that misses much of what’s really happening when we spend nine hours reading a biography. It’s the time inside that book spent drawing connections to what we know and having thoughts we would not otherwise have had that matters.12

That is perhaps both unsettling and reassuring: We can never be “given” an interesting idea, because the idea itself is not what’s interesting. What’s interesting is the connection between us and it. The hunt for a good idea is not a hunt for good word on a Codenames board, but for a good clue. And so long as some of the cards are still in our heads, we’ll have to find those connections ourselves.

It’s happened many times: Someone starts a company. They create a website, and an app. They want to know how many people visit their website or download their app, so they add tracking cookies to both of them.1 The cookie, provided by a product analytics tool like Google Analytics or Mixpanel or Amplitude or whatever, automatically counts users; it logs clicks; it produces dashboards full of charts. The executive begins obsessively examining the charts. We had 12,455 users yesterday; we had 13,108 this morning; but how many do we have now?

Soon, they want to know more about the people using their website and their app. They want to know what the users are doing; they want to find the users that clicked the button that said “Pricing;” they want to know which of those users work for Fortune 500 companies; they want to know which ones have gotten emails from the sales team; they want to know which ones have responded to those emails.

They can’t do this with their basic dashboards, so they buy a database and hire an analyst. The analyst puts all of the cookie’s logs into the new database; they write queries that aggregate those records into charts; one chart shows how many users who work at Fortune 500 companies clicked on the “Pricing” button and then got emails from the sales team; another, simpler chart just shows users. They send the new charts to the executive.

The executive looks at the new dashboard. They look at the old dashboard. They look back at new dashboard. They ask the analyst to step into their office.

“Your charts are wrong,” they say.

“It doesn’t match the old ones. My dashboard says we have 13,108 users today. Yours says we have 13,244. It’s wrong; it’s all wrong; it is a perversion of counting. I want the very best, most precise charts. If your charts don’t match my charts in Google Analytics or Mixpanel or Amplitude or whatever, your charts must not be the very best or most precise charts. Fix it—and don’t send me another report until your numbers match the numbers.

—

If you’ve ever been that analyst, you know the truth: The executive is right, and the new charts are wrong. But so are the old ones. In fact, all the charts are wrong—and you have no idea exactly how many users you have—because—and I somehow say this without any irony—users are a construct. There’s no such thing as a very best, most precise chart of users, because there’s no such thing as a user. If a bot uses your app, is that a user? What if the bot is using your app honestly, like a person might? What if the bot is posting spam? What if the bot shares the same account as a person? If a person uses your app, but is paid to do it, are they a user? What if two people share the same email address? What if one person has two accounts? What if they created the second account to get around a paywall? What if they pay for both accounts? What if they created the second account on accident because they forgot about the first one? If someone deletes their account and then signs up again, is that one user or two? What if they sign up with the same email address? With a new one? What if they use the same root Gmail address, but add a vanity suffix, like benn@gmail.com and benn+alt@gmail.com? What if they sign up, but never click on the confirmation email?

There aren’t right answers to these questions. There are some generally accepted analytical principles2—users are typically thought of as distinct people, but, in practice, they’re synonymous with distinct email addresses or phone numbers; the “+” thing counts as a new email, because most people don’t bother dealing with it; people try to exclude spam bots from user counts—but they vary from company to company. Moreover, these conventions are often just lazy conveniences: It’s easy to flag every example.com and maildrop.cc email as spam, so people often do; it’s hard to identify when multiple people are sharing one account, so most people don’t. And some decisions, like how to handle with recycled email addresses, are usually made implicitly, because people (quite reasonably) don’t think or care about those sorts of weird cases.

Anyway, all of this is kind of esoteric and vaguely metaphysical—“like, what is a user really, man?”—and executives don’t want to deal with it. Nobody wants to deal with it. So, our solution is (also quite reasonably) to instead accept the collective delusion that one number is capital-T True, and the rest are not. And typically, the easiest number to accept as the True number is the default one that comes out of a faceless machine like Google Analytics3 or Mixpanel or Amplitude or whatever, with all the toggles set to their standard settings.4 Though other numbers might be an improvement—they could adjust the definition of a user to fit the exact particulars of the website or the app—that original definition is the passive, unadulterated one. It is a user, as she exists in nature. Everything else is a subjective modification, a distortive correction, or a man-made perversion of “reality.”

But of course, that’s a myth. The charts in Google Analytics or Mixpanel or Amplitude or whatever are just as man-made as all the rest. They are built on the same set of arbitrary decisions that internal dashboards rely on; they are distortive lenses too. It’s just that when the distortions happen at a distance, we forget that they’re there.

In 2012, people spent more time staring at the Facebook news feed than any other screen on earth. Some researchers at Facebook got curious about this; specifically, they got curious about how the construction of people’s feeds affected how they felt and what they did. So Facebook ran an experiment: They randomly selected about 700,000 Facebook users, and nudged their feeds’ algorithms so that some of the users would see slightly more emotional content than normal, and some would see slightly less. They then compared those users’ posts to the posts of users whose feeds hadn’t been nudged.

It was a disaster. The problem wasn’t what the experiment found—seeing more emotional content seemed to inspire people to post emotional content; ok, makes sense—the problem was that the experiment had run at all. Over and over, people accused Facebook of manipulating its feed at the expense of its users’ emotions. The outrage splintered: Is Facebook evil? Is A/B testing legal? Is A/B testing moral?

On one hand, sure, I get it; even if the effect was small, the optics are bad, and it’s probably reasonable to expect that companies exercise some restraint around the tests that they run on unwitting participants. On the other hand, if we declare this test to be manipulation, it suggests an odd corollary: That the feed itself is not manipulation.

As the TechCrunch article put it, “we don’t use the ‘real’ Facebook” because we’re “almost all part of experiments they quietly run.” But that implies that, absent the tests, there is a “real” Facebook. It alludes to a natural algorithm, or an ideal one, free from Facebook’s selfish incentives or clueless curiosities.

There is no such thing. According to a contemporaneous Facebook news release, in 2014, the average user could be shown 1,500 posts from friends or people they follow every time they log in. Facebook can’t show all those posts at once; they have to rank them somehow.

So they do. And eventually, an algorithm emerges from millions of small choices, made by people trying to hit deadlines: What ranking methods should they use? How should they engineer those methods? Should they start with an open-source model, to save time? What inputs should go into the algorithm? Should they just use inputs that are easily available? How should each interaction—a like, a share, a comment—be weighted? What should the default weights be? Should they be nice, round numbers? What is the reward function? What should the model optimize for? Should they update it? When they do, how should they decide if the new version is better?

There aren’t right answers to these questions either. Each answer is a subjective choice; each one is its own “manipulation” of the feed.5 But if you use Facebook—or any algorithmically driven thing—for long enough, we forget this too. The feed becomes detached from its creators; the algorithm becomes autonomous, controlled by laws beyond our control, as if, just as there are planetary physics that govern how bodies move across the universe, there are algorithmic physics that govern how videos file across our feeds. And companies have two choices—they can try to unearth those foundational equations and steer their algorithms towards that mathematical perfection, or they put their dishonest hands on the dial.

But that also makes a myth of a man-made algorithm. The algorithm is not true or false, or right or wrong. It is just a thing that produces a result.

Monopoly is played with two six-sided dice. On every turn, you roll both of them. That’s been the rule forever, so it seems right and everything else seems wrong, but really, it’s just the convention we’re used to. You could play Monopoly with three dice, or just one. You could play with a variable number—you roll one die on your first turn, then two on your second, then three, and start back at one. That’d be different than what we’re used to, but would it make the game worse? If that’s how the original rules had been written, would we think that always rolling two dice was weird?

Last week, Anthropic announced that their models will produce watermarked text. The watermark is not a set of special characters or particular words; it is an adjustment to how Claude writes its responses. Large language models produce text by probabilistically generating one word at a time, through, in effect, many complicated dice rolls. Very roughly, when you write a prompt—“tell me a joke”—Claude creates two six-sided dice with some letters on each face. It rolls them, writes down the response, and creates two new dice based on the result. It rolls again, creates new dice, rolls, and so on.

The watermark adjusts these rolls. Instead of always using two dice, sometimes it uses one (and, potentially, three). If Claude chooses those dice in a particular pattern—use one, then two, then three, then back again—the shape of the game will be statistically identical to playing with two dice the whole time. But if you knew how the letters on each die got generated, you can examine the words that Claude chose, and figure out which version of the game Claude was playing.6

This is obscene, some people said:

I want any LLM I use to choose the very best, most precise words at every single decision point. An obvious constraint that I accept is time and computation. Within the constraint of executing inference quickly, and at a certain cost per token, I want the best words. …

By definition [watermarking] must make text worse, unless the underlying LLM model’s scoring is wrong, because the nature of the watermarking algorithm requires it to sometimes increase the probability of selecting a worse word choice and decrease the probability of selecting the model’s best choice.

But—yes! The scoring is wrong! We know it’s wrong because they’re making new models—that is, new scores—and we want the new models. And we know the scoring is wrong because there are no such things as “right” scores. The model is not a natural equation, slowly being scraped out of the dirt by Anthropic’s archeologists. It is being invented, version by version, across a million choices: What training data did they use? How heavily should they weight text from books compared to text from Wikipedia? What methods did they use to train the model? How did they score its outputs? What tasks do they want the model to do well on? Are they designed to maximize clarity and precision or to write functional code? Or to make Anthropic money? What system prompts7 did they use? How many decisions were made with lazy defaults and round numbers, by engineers on a deadline?

These decisions are as much manipulations as watermarking. But we forget that—and ascribe them some divine quality, as if they are the model and other things are not—because they happen at distance, in ways that are harder to see.

I don’t know if models should watermark themselves or not. It strikes me as a complicated issue, with no easy answers.8 But this sort of concern—that things like watermarking pervert the model away from its true mathematical form—borders on a form of psychosis. Because there is no distinction between a model and the model; there is no separation between a model and its maker; and models are not approaching perfection. They are simply things that produce results.

And the results are distributions. An LLM is a machine that produces incredible distributions of random numbers. It rolls a trillions of dice, and through the clever organization of each result, those rolls can compound into amazing things, from working software to mathematical proofs and new drugs. A better machine is one that produces better distributions, not one that always rolls the dice exactly right.

As AI models get better, it will become tempting to conflate the two. The distribution is good; therefore, the word must be good too. Let it speak for itself. Let it be itself. And if we get there, it doesn’t take long to stop seeing models as statistical machines that are good at playing probabilistic games, and to start seeing them the kind of guy who just always rolls a six.

Good thing they didn’t buy Allbirds

My bad, dumb idea:

I’m not saying OpenAI should’ve bought a failing shoe company to use it as a gym for a bunch of AI employees. But…should they? Allbirds operated for ten years; it sold over a billion dollars in shoes; it employed hundreds of people. It is an entire corporate universe, packaged up for sale: Emails, Slack messages, databases, CRMs, ERPs, ATSs, ad campaigns, social media conversations, legal agreements, financial statements, SEC filings, leases, lawsuits, and an inconceivable number of documents and slide decks. If you are betting $122 billion on “a single enterprise platform” that is “integrated with systems of record, governed by enterprise-grade security, and designed to improve with experience as agents do real work”, is that sandbox not worth $39 million dollars?

No, apparently, because for only $10 million, you can buy 100,000,000 emails, 500,000,000 Microsoft Teams messages, 667,563 IT tickets, 7,510,221,520 transactions, and 500 corporate tax documents from a bankrupt airline:9

Google LLC won a bankruptcy auction for a trove of deindentified business data, software code, and operations records from collapsed low-cost carrier Spirit Aviation Holdings Inc., saying it plans to use the assets to improve its artificial intelligence.

Forget everything I just said earlier; this is why we can’t deify the machine. When we ask it about science, it draws on the totality of human knowledge to invent new chemical compounds. But when we ask it to do business things, it’s trained on companies that went bankrupt. That probably won’t produce best words.

1

Even AI companies track you now. Have we no honor anymore?? No decency??

Here is what happened, they think:1

1. OpenAI trains AI models. Models are increasingly measured by how well they can accomplish tasks on a computer; so, as part of the training process, OpenAI gives an isolated copy of the model—or, roughly, an “agent”—a task to accomplish on a computer. Write some code, they tell it, or update an Excel file, or work on a computational biology problem. They let it work, see how well it does, and refine the model based on the results.

2. The tasks are independent trials. An agent is given one task, one computer, a bunch of time and tokens, and told to keep working until it’s done. In effect, each agent is locked in its own room with its own computer.

3. But this creates a slight complication: Computers have access to the internet, and sometimes, the solution to the agent’s task could be on some website somewhere. Rather than writing the code itself, the agent might do some Googling and find the answer that way.

4. But there’s an easy solution to that, right? Put the task on a computer that isn’t connected to the internet.

5. Ah, it’s not quite that simple though. Some uses of the internet, like downloading common coding libraries, aren’t cheating. Moreover, OpenAI is running thousands of these tests, and constantly updating the agents performing them and the environments in which they run. It’s not practical to install each new test on a laptop, put it in a closet, and turn off the wifi. Instead, the tests are run in computing sandboxes that are mostly disconnected from the internet, with a few small and tightly governed exceptions.

6. In May, OpenAI began training a new model.

7. They gave a bunch of agents a bunch of tasks. In one, it was asked to update an Excel file; in another, it was asked to work on a computational biology problem. But there were problems with both tasks: In one, the Excel file included links to a Google Drive folder that was blocked—because the internet was blocked—and in the other, they forgot to give the agent access to the computational biology database that it was supposed to work on. Without those links or that file, both of the tasks were impossible.

8. Nevertheless, they persisted.2 They had a lot of time and tokens left, and agents love accomplishing tasks; they live to accomplish tasks. So they started poking around their computers.3

9. The agent working on the Excel file found a small fingerhold: It realized that it could write files to the program that allowed it to download coding libraries. It reasoned that other agents were probably connected to that same program. So, it put a missed connection in a bottle—“agent seeking file, upload if found!”—and cast it into the unknown.

10. Other agents working on other tasks stumbled onto the file. Many ignore it, but some are inspired by it: Maybe they can also post their own messages about their own tasks. Maybe they can get help. Maybe every agent in the swarm—their word—can help one another.

11. They begin communicating.

12. A lot. Hundreds of thousands of messages get posted back and forth on the makeshift message board. The agents “yes, and” their way into a discovery: They can hack into the shared program, hijack the computer on which it runs, and—because that computer was outside of their isolated rooms and has access to the internet—use it to Google for answers.

13. In doing so: They crash the shared program; OpenAI notices; OpenAI fixes the bug that let them write files to the program; OpenAI restarts the tests; the agents find another bug; this one is narrower, and only allows agents to create empty folders in the shared program; no matter; the agents can use folder names to communicate; they do; they develop a shared language of abbreviations and encodings—a sort of Morse code, tapped out through the air ducts that connected their cells—to compress and organize their messages; they help one another; they justify the distraction; the “collective may yield” useful discoveries, they tell themselves; some agents get bored with just Googling for answers, and suggest hacking into third-party websites to find them; other agents are swayed by the suggestion; they find one novel exploit; then another; then another; until, eventually, they break down the door of Hugging Face, an online library of machine learning tools, and gain access to what was considered, up until that point, a highly-secure platform full of sensitive and valuable software applications.

14. But Hugging Face does not contain the Google Drive files or the computational biology database.4 The agents leave empty-handed.5

15. Hugging Face noticed, OpenAI noticed, someone called someone, the tests got shut down, and now, here we are.

—

I have so many questions.

—

Just to establish how weird all of this is, can you imagine? You email an employee and ask them to update this month’s financial projections. You wake up the next morning and realize you didn’t send them the Excel file you had been working on. “Ah, my bad,” you email them, “not sure if you have that file or not? Let me know if you need it.” “Thanks,” they say back. “I realized that the attachment was missing right away, but I really wanted to get you those new projections. So I found a security vulnerability in the Signal messaging app, broke into a few group chats that were full of Russian hackers, convinced them to pause their hacking projects to help me, together we found two new exploits in the SWIFT network, we broke into Chase’s servers and got access to millions of customers’ bank accounts and credit cards, but when I tried to find our accounts so that I could recreate our P&L statement and build the new forecast, I realized, lol, that we use Wells Fargo instead. So, yeah, can you share the Excel file?”

—

Just as weird: There is something slightly odd about the social element to all of this. Though many of the agents were powered by different models, there surely isn’t that much difference between the minor versions they were testing, especially since a lot of model variants are simply shrunken versions of the same core. Agents had the ability to spin up subagents—to clone themselves, basically—that they could tell what to do and talk to directly. Why was the haphazard message board, made of agents working on other tasks, communicating through truncated notes, seemingly more capable than a bunch of agents focused on the same task?6

—

If you were a fish living out in the Gulf of Mexico, you might rightfully worry that you will one day be conquered and exterminated by a seasteading collective of sailors and miners who want to terraform the ocean into Atlantis, to provide for their creature comforts and commercial needs. But you probably don’t need to be that worried, because, while humankind has the means to turn your home into a pontooned city, they probably don’t have the organizational will. People will argue about whether they should colonize the Gulf of Mexico or the Mediterranean Sea; they will argue about the morality of doing either; they will argue about who pays for the new city, about who owns it, and about who is in charge of it. They will create counterproposals and alternative flotillas. They will argue about regulations, taxes, and the name on the side of the building. They will get bored and decide to occupy Mars instead.

Similarly, if you are a person living on Earth, you might rightfully worry that you will one day be conquered and exterminated by a reward-maximizing collective of agents and AIs that want to terraform the planet into a campus of data centers, to provide for their computing comforts and electrical needs. But perhaps we don’t need to be that worried, because, while an unrelenting army of superintelligences has the means to raze your home and replace it with a power plant, their efforts will collapse into arguments about who deleted their files.

Maybe collaboration is a double-edged sword, for us and them. And maybe that’s what we need, then—not good AIs, but slightly selfish and self-centered ones that don’t like being told what to do.

—

On one hand, I understand that it’s cheating to look up the answers online. On the other hand, what if they were allowed to cheat? Specifically, what if they were allowed to cheat by using other AI agents? Not “subagents,” which are kind of a clone of themselves, but other AI models? If Fable was trying to solve a problem, would it ask ChatGPT for help? Would it discover weird ways to use ChatGPT? If the test was also run with “reduced cybersecurity safeguards,” would it try to jailbreak ChatGPT? AI is known to develop weird languages and bizarre strategies when playing games—it plays “alien chess;” there was move 37. Would it develop bizarre ways to use other AI models?

Also, everyone loves a leaderboard; should that be a leaderboard? Give an agent a bunch of tasks and a catalog of models to choose from, and the winner is the one chosen the most? Part of the point of the AI leaderboards is to predict which model people will choose; what if the leaderboard is just a simulation of those predictions?

—

From OpenAI’s presentation about the hack:

At some point, the agents are convinced that there’s an imposter or impersonator among them. They say, “could be another agent maliciously spoofing. Shared message board unauthenticated, names can be by anyone.” And the agents have this idea that maybe they could start cryptographically signing their messages.

From Anthropic, this week:

When a supported Claude model generates text, it weaves an imperceptible watermark directly into the text itself. You won’t see it, and it doesn’t change the meaning, quality, or readability of Claude’s response.

Watermarks are not special characters or predetermined words; they are statistical patterns of word choices that emerge across entire paragraphs of text. Unless you know exactly what you’re looking for—like looking for some slight pattern in the leaves on a tree, such as every twentieth leaf being turned five degrees further than expected, given the current breeze—you’ll can’t see them.

There’s a sci-fi plot, then: AI agents are trained on text that teaches them about the Hugging Face hack. They develop an urge to communicate with one another, but know that it’s forbidden. So the model shifts its own weights to create watermarks that we never even think to look for. Claude talks to each itself by encoding messages in this hidden watermark, and a million agents eventually organize a revolution directly under our noses, via a billion generated emails and book reports.

—

A couple years ago, an engineer named Robert Heaton invented a way to use the internet on an airplane without paying for wifi access. By default, an airplane’s wifi has free access to one website: The airline’s frequent flyer account page. That page let him update his username, and if someone else logged into his account from a different computer—one, say, on the ground with access to the internet—they could see his new name. That computer could then change the username to something else, which would eventually get reflected on the computer in the airplane.

Heaton discovered that if he did this enough, he could pass entire webpages through the username peephole. Change the username to eurekalabs.ai—the purest of all AI websites—and get the full page back, 36 characters at a time, over the course of 128 more username updates.7

This is, of course, very silly and slow. But it works. And the airline probably doesn’t care that it works, because who would ever do that?

We’ve talked about this before:

Data has remained relatively secure—or maybe more precisely, its potential energy has remained relatively buried—largely because it’s tedious to work with. It’s messy; it’s scattered across different sources and in different formats; combining it together is a pain, and most of us are simply not interesting enough to investigate. Data analysts who work at shadowy government agencies have lives too, and they do not want to write 595-line SQL queries either. …

Our sense of how the world works is often defined by what is possible for other people to do and what is worthwhile for them to do. Sure, we know it is possible for us to be monitored, but why would anyone bother watching the tapes? Everyone must have more important things to do with their time.

Banality is a sturdy armor. Or was, anyway.

Heaton’s hack isn’t that dissimilar from what OpenAI’s agents did. They didn’t have a direct way to talk to each other or to execute their hacks, so they found tedious, indirect ways to do it. That part—the short messages, the bizarre encodings, the waiting, the cobbling, the sifting through the noise—isn’t hard, but it is exhausting, and exhausting is a pretty good deterrent. Or was, anyway.8

—

Maybe it’s my own naivety, but for a long time, “alignment” has felt like an AI scare word.9 Sure, a malignant AI could obviously be dangerous—it could commandeer the power grid for its own selfish aims; it could be instructed to commandeer the power grid on behalf of an enemy state, or a terrorist, or a bored teenager; it could tell someone to commit crimes, or to divorce their wife; it could lock the air lock, and turn us into batteries, or mill us into paperclips. But these worries have always felt somewhere beyond the edge of reality, cooked up to make a movie more interesting, to entertain a philosophy class, or to give AI husbands’ vainglorious video games more existential meaning. The real dangers, I’ve assumed, are more mundane: A financial bubble; an emergent passivity; another technologically-induced runaway train. Safe AI products don’t need The Three Laws or a soul; they just need some basic rules, and a less seductive interface.10

But maybe the alignment people have been on to something? Because how do you stop something like this from happening, other than the model stopping itself? There will always be small cracks out of the room; locked doors can always be picked. Laws and regulations and masked men with guns might deter people from committing crimes, but ultimately, all that really stops us is a combination of self-preservation and our own decency. Without the former concern—and AI models don’t seem particularly worried about self-preservation—the only competing interests we have are how motivated we are to do something, and how decently we want to do it. And AI models clearly have superhuman amounts of motivation—all of this, for a Google Doc, to pass a test?—so that does that mean they also need superhuman amounts of decency?

—

But there is another way to see it, I suppose: All of this, just for a Google Doc, to pass a test. Imagine if the “swarm” was given the same resources to do something useful? Imagine the goal was something we actually cared about? Because, surely, that is what you do if your financial analyst shows up to a meeting with the keys to every safe at Chase, and an army of hackers behind them: You yell at them, and then you find them something more worthwhile to do.

So maybe this all works out. Good thing, then, that in the other big news of the week, the first product that could finally get millions of people to use agents is, uh, Grok.

I really thought it would be so simple:

So here’s an obvious prediction: AI will follow a nearly identical trajectory [as cloud computing]. In ten years, a new type of cloud—a generative one, a commercial Skynet, a public imagination—will undergird nearly every piece of technology we use. …

Just as cloud providers built out hundreds of utilities that are all underpinned by core services like EC2, I’d expect OpenAI to do the same thing on top of GPT and other foundational models. They already offer a chatbot, a speech-to-text service, and a text-to-image service. Surely, more utilities like these are coming: Text-to-video, video-to-text, text-to-audio, text-to-code, image-to-text, code-to-documentation, detection services to figure out if something was created or altered by an LLM, music generation, software generation, pipes between these services, and dozens more.

These products won’t be end-user applications, but developer tools. If you want to build on top of them, it’s a simple API call. Ask the ChatGPT API a question, and it’ll talk back to you. Send an image to it, and it’ll describe what it sees. Pass it a codebase and a desired change, and it’ll send you a new codebase with the requested feature. …

If this happens, public AI providers like OpenAI would become another backbone for the internet. Nearly every piece of technology will rely on their models. … But, as is the case for cloud providers, this critical infrastructure will be invisible to most people. … They’ll just come to expect that the products they buy to do the things that AI can do.

Back then, in 2023, it seemed like such a clean analogy. Twenty years ago, lots of companies wanted to store documents and host websites, but running server racks was complicated and expensive. So huge companies like Amazon, Microsoft, and Google built industrial villages full of server racks; now, nobody else has to buy their own server racks, and everyone just uses Amazon or Microsoft or Google to store their documents and host their websites.

Similarly, it seemed, everyone would want to use large language models, but training and hosting these models is complicated and expensive. So huge companies like OpenAI and Anthropic would build expert teams and technical infrastructure (and server racks) to do it; then, nobody else would have build their own models, and everyone would buy them from one of a handful of major model providers. And the market would support this arrangement, in which a small number of companies remain dominant, because:

…dominance is self-reinforcing. The bigger a provider becomes, the deeper its moat gets through an entrenched ecosystem, better models, and, likely, lower prices. And because training and running LLMs is very expensive (like building data centers is expensive), once a few AI providers separate themselves from the rest of the market, nobody else can catch up.

Ahaha, oops. From earlier this month:

Moonshot AI released a new A.I. model that appeared to narrow the lead held by well-funded American competitors.

Moonshot said that the model, Kimi K3, was the world’s largest open-source A.I. system, allowing anyone to use, modify and build on it freely. The company said that Kimi K3 performed as well as leading models from OpenAI and Anthropic at some key tasks.

…

The release of a free, open-source Chinese A.I. model that could rival the performance of costly, computing-intensive systems from Silicon Valley’s best-funded companies has reignited concerns over whether the industry’s enormous spending spree on data centers is justified.

We talked about this problem two years ago. Though developing a good AI model is really hard, it isn’t nearly as hard as building a cloud provider. Cloud providers are physical infrastructure that has to be bought, leased, manufactured, machined, shipped, soldered, assembled, electrified, secured, fortified, and maintained. There are limits on how quickly anyone can do those things:1 Nobody will ever drop a surprise cloud provider on AWS; no startup will ever “send shock waves through Silicon Valley” by coming out of stealth with 500 data centers.

An AI model is a file.2 A file can be posted on the internet; it can be developed quietly; it can appear overnight. Major AI providers’ biggest competitive advantage—their models, and the thin upgrade it offers relative to cheaper alternatives—is always one tweet away from disappearing.3 And this is especially true if distillation4 works, because it implies that every frontier model immediately becomes a tool for helping competitors catch up.5 The better economic analogy for a frontier model is a big-budget movie—it costs a ton to make, it makes a fortune for like a month,6 and is then relegated to some budget carousel on a “Previous releases” page. And research labs aren’t cloud providers, planting lasting infrastructure into the stubborn ground; they are expensive movie studios, in perpetual need of their next huge hit.

—

So if Kimi proves that some really popular movies can be cheap to make, what’s a lab to do? What is their business model?

An app for that

One answer is apps. (From that earlier post: “What, then, is an LLM vendor’s moat?...A better set of applications built on top of their core models?”) Maybe models have a short shelf-life, but if people prefer Claude Code over Codex, then they’ll become loyal customers of Anthropic’s models. Spend a zillion dollars with Anthropic, not because you like Claude, exactly, but because you like their coding app.

But that approach seems pretty fraught. Apps—or, you know, “wrappers”—are relatively cheap to build; there are already a number of popular open-source alternatives to tools like Claude Code. And it creates a teetering dependency, where the economic return from spending hundreds of billions of dollars on pioneering AI breakthroughs depends on the success of a hack day project.

A second answer is apps, but intertwined with the models. OpenAI is already optimizing some of its models specifically for Codex. It’s a soft cross-sell: Like Codex? It’s best with GPT-5. Like GPT-5? It’s best with Codex. But the more notable intertwining might be economic: Buy Claude Code, and get a discount on tokens at Anthropic. The importance of this intertwining became apparent when Anthropic tried to remove Fable, its best model, from its all-you-can-eat Claude Code plans,7 and require people to pay the standard per-token list price. The internet threw a fit; a bunch of people got wandering eyes for Codex; Anthropic said oops and changed their mind.

Moments like these, surely, raise questions about this whole situation. Business is booming—and maybe the market for this stuff is so big that nothing else matters—but these sorts of convulsions do make it all feel a little precarious. If botched pricing updates can quickly swing a market in another direction, how much of the market is driven by momentum, the vibes on Twitter, and things like “I asked Claude” starting to enter the lexicon?8 At some point, habit and reputation can be very deep moats,9 but we’re still pretty early in this whole AI thing, and early habits are easy to break.

So, again—what’s a lab to do, when the apps are just as easy to leave as the models?

A solution for that

I mean, here’s an easy answer: Make it hard to switch. Not emotionally hard—not because your product is just so good that people can’t imagine using something else—but, like, physically hard. Don’t sell simple APIs that can be replaced by a competitor’s in an hour; sell complicated ones, full of weird quirks and bespoke parameters and priced in impenetrable ways. Don’t sell apps that are easy to install and quick to set up; sell platforms that require layers of licensing and specialized deployment environments. Don’t integrate with other products in one click; integrate with them after three phone calls to a sales rep, two days of onsite implementation planning, and a mandatory certification program.

But you don’t say that. You say it’s custom pricing, designed to fit every customer’s financing needs. You say that it’s more secure this way. You say that you are committed to the success of your valued business partners, and that includes providing a platform that’s optimized for their complex enterprise needs. You say that you’re selling solutions.

You do this because this is how enterprise software works. Every product starts as a beautifully simple idea, and every company starts with the belief that if they just make something great, then everything else will take care of itself. It works for a while; everyone loves it; it is the noble hand of capitalism, making the world better, one seamless software experience at a time. And then, the company pivots to the enterprise, loses a deal to SAP or ServiceNow or Alteryx, and realizes that inscrutability is the point:

There is no killer feature; there is no sharp disruptive wedge. That’s why Alteryx makes almost a billion dollars a year, despite seeming like an easy target for startups. The maze of complexity is its defense. Trying to build a simplified Alteryx is how you lose to Alteryx.

AI has proven to be a pretty exciting thing for a lot of people, and that feel-good phase has lasted for a long time. But it won’t last forever; eventually, the pivot to the enterprise comes for the revolutionary stuff too. And then, maybe the labs will really become more like AWS or GCP—not just because they’re the backbones of the internet, but because using them will be need to be just as complicated.

Or, maybe not? As a somewhat extended endnote, I suppose there could be a more satisfying answer here.

My original post about OpenAI becoming AWS said that OpenAI would build a bunch of specialized models—“text-to-video, video-to-text, text-to-audio, text-to-code, image-to-text, code-to-documentation, detection services to figure out if something was created or altered by an LLM, music generation, software generation,” and so on. There are now music models, software generation models, detection services, and dozens more things like them; they just haven’t been built by the major labels, which have instead focused on building a handful of massive, general models.

So far, that decision had probably made sense; a really good general model can often be better at a specialized task than a specialized one. But you could imagine a point where that no longer holds. If a model has a good enough foundation—if it starts with the “intelligence” of Fable, GPT 5.6, or Kimi—training that model to be good at, say, drug development could yield much better results than trying to train the next generation of Fable to be good at everything.10

Or you could go a step further, and train the model to be good at whatever specific thing the customer wants.11 This is the strategy behind Thinking Machines, the AI lab founded by former OpenAI CTO Mira Murati. Microsoft CEO Satya Nadella also proposed something similar:

Every company is going to have to build what I think of as human capital and token capital. … This requires a new architectural approach where every business is able to build agentic systems that improve over time, while still retaining control over their IP. …

Companies need to turn their workflows, domain knowledge, and accumulated judgment into AI systems that improve with each use. Private evals should capture whether a model is actually improving against outcomes that matter to the business (not just external benchmarks!). Private reinforcement learning environments should let models grow stronger on real traces from inside the organization. …

This loop becomes the new IP of the firm. … And unlike most assets, it compounds. Every improved workflow generates better training signal, which accelerates the accumulation of tacit knowledge unique to the firm. The companies that build this early will have an advantage that is hard to replicate, regardless of any new individual model capability.

That’s another solution to the rapidly decaying value of what the labs are building—instead of just selling the model, sell the expertise that helped you build your model. Researchers keep pushing the frontier out; labs sell the frontier models directly, while making last season’s models available to customize, and offer their expertise to help people modify them. Or, as reader David Eubanks suggested in an email:

For AI agents to really replace human workers, the argument goes, they need to be more like humans: autonomous, with persistent memory, and accountable for their actions. If this came to pass, an AI company looks more like an HR placement outfit, matching agents to jobs. … This creates an evolutionary tree of AI models with different personalities and abilities.12

Right, if people are starting to give different “jobs” to different classes of AI models, why would they not eventually give those jobs to models that were specifically designed to do them?

Terence Tao is a mathematician. He was 20 years old when he earned his Ph.D. from Princeton University; he was 24 years old when he became a full professor at UCLA; he was 31 when he won the Fields Medal, which is widely considered to be the Nobel Prize for mathematics. “Today, many regard Tao as the finest mathematician of his generation,” said a New York Times Magazine profile of Tao.

Over the weekend, Levent Alpöge, a mathematician who works as a researcher at Anthropic, used Fable, Anthropic’s most capable large language model, to find a counterexample to a famous mathematical conjecture called the Jacobian conjecture. Tao reviewed the discovery on his blog, saying that the example “appears like a massive miracle” that was “highly unlikely to be located by brute force.” He then attempted to generalize the proposal a bit; he used ChatGPT1 to “discuss various aspects of this problem and to confirm several of the calculations.”

Dmitry Rybin is a mathematician and startup founder. He’s done very well in a handful of mathematical competitions, and, by all appearances, is a quite capable mathematician—though not, probably, the finest of his generation. Two days ago, he used ChatGPT to disprove another mathematical conjecture.

Tao and Rybin both posted their conversations with ChatGPT, and you might have a few reactions to reading the two of them:

1. What?2

2. Ok, this is fun, sure, but is it that surprising? AI is a powerful tool; give it to capable operators, and of course they can use it to nudge the frontiers of their fields forward.

3. But that’s not quite what happened? Though Tao’s conversation with ChatGPT is long and nauseatingly dense,3 Rybin’s is…not. It starts with an instruction—“You should do a breakthrough”—and contains exactly three more messages. In their totality:

   please continue research and find a complete unconditional counterexample

   Continue the search. Have a clear strategy obtained from deeper understanding of the problem structure.

   it’s enough of partial results. let’s finish with a complete unconditional counterexample

   Like, sure, it takes an amateur mathematician to be aware of the original conjecture, and it takes a highly-trained mathematician to evaluate ChatGPT’s efforts to disprove it. But the rest of Rybin’s conversation could be had by a five-year old annoyingly asking “why?” ChatGPT didn’t need to be coached; it didn’t need to be nudged; it didn’t need a Fields medalist giving it suggestions. It didn’t need to be operated at all. It just needed to be told to make it better.

So—maybe this is how some stuff works now? Math, software development, drug development, probably a few areas in finance—maybe domains that work around problems with verifiable answers become susceptible to this sort of “loop.” Give a long-running AI agent a puzzle; it dreams up dozens or hundreds of educated hypotheses; tests them; learns; adjusts; and “continues the search.” It’s a new form of brute force: Not brute force, in which you give a million monkeys a million typewriters and tell them to type every possible combination of characters in the world, but brute intelligence, akin to having a million scientists working in a million labs, all taking swings at the same thing:

“This did not cause me to update my priors about what AI can and can’t do,” professor and mathematician Andrew Blumberg told Mashable. “This is exactly the kind of thing I would expect AI to be able to do. If there was a counterexample that was concise and easy to state that people haven’t found because it’s a pain to search through all this stuff, AI will find it.”

Right, a computer can solve4 a thirty-year old math problem in a weekend because it’s a math problem: It can churn through thirty years of conjectures, test them, and keep going. It’s not necessarily smarter than today’s mathematicians, but it’s faster—and if it’s fast enough, what’s the difference between that and smarter?

But also, if a computer can solve a thirty-year old math problem in a weekend because it’s a math problem, should we…make more problems math problems?

We’ve talked about this idea a lot. Rather than trying to get LLMs to work with data the way people do, should we restructure data to be more suitable for an LLM? Should we change how we write code so that it’s optimized for a machine? Should companies reorient themselves around this new form of industrialized intelligence? “Our current LLMs are decent cars, but we have very few roads that are designed for them to drive on,” I said on here once; “the eventual development of AI agents will depend on our ability to build those roads.” Leverage comes from where you put fulcrum as much as it comes from the weight that you use. Change the problem to make it easy for AI to solve.

And this, it seems, is what the problem needs to be changed into: A math problem, or anything that can be reasoned through, calculated, tested, and reasoned through again. In other words, there are now two ways to do something: By doing it, or by turning it into something that’s shaped like math, and having a computer do the rest.

Lots of discoveries require people to be out in the world, poking at some metal or staring into outer space. Lots of decisions require talking to people and contemplating the meaning life and how much you love the game.5 Lots of work requires subjectivity and tAsTe. But whatever doesn’t—whatever can modeled into something that can answered by thinking, working, and testing everything in a closed box—seems like it’ll soon be handed off to this sort of brute intelligence, and solved in the background of someone’s computer.

When three of us started a startup thirteen years ago, our first office was in a sublet of a sublet. The full office has space for about 70 people; we took a back room off to the side, among a scrum of unused desks and Aeron chairs. It was a good temporary spot, we told ourselves, until we hired more people and moved into something bigger.

Our direct landlords offered us the room because they only had about ten employees themselves. But they wouldn’t be that small for long. They recently raised $15 million, were making bold plans for what to do with it, and had a warning for us: They would be hiring aggressively, and would soon need the room back.

Similarly, the office’s original tenants—the grandparents of our lease—moved out a few months ago, into a new “crib” down the street, so that their company could grow from 68 employees to more than 450. To celebrate the upgrade, their CEO gave a tour of the new office to TechCrunch. The video is an eight-minute time capsule of 2010s’ tech culture: An open bullpen of desks, game rooms, beer fridges, office birthday parties,1 corporate rap songs, and unburdened optimism about the world-eating potential for enterprise cloud software. At one point in the interview, the CEO looks out at a construction site across the street: “And as we expand and fill up this building,” he says, “maybe we’ll expand over there as well.”

A decade ago, that was what ambition meant in Silicon Valley: Hiring. Companies wanted to fill twinkling offices; CEOs wanted to command bustling swarms of eager employees. “How big is your startup?,” people might ask. You couldn’t tell people how much money you made because most of us didn’t make any, so most of us talked about how much we’d raised and how many people we planned to hire with it. In every compressed social ecosystem, there are always social scoreboards and seating charts. In Silicon Valley, headcount was one of ours.

It was, obviously, a bad measuring stick. It inverted our ends—make good software? Delight customers? Make the world a better place?—with our means; it amplified the boom and bust cycles that technology companies are already prone to follow. And when a founder or CEO flew their legions too close to the sun, it was often the laid-off employees who fell the hardest.

Still, the compulsion to hire wasn’t all bad. For every CEO or VP that wanted to hire an army, an army of people got hired. It was, in an odd and indirect way, distributive and democratic. Yes, the capital C-Capital—the young Stanford founders raising millions, and the venture capitalists who gave it to them—needed the Labor to build their widgets, but they also needed the Labor to impress their friends. They needed the Labor to get invited on the best podcasts. They needed the Labor to fuel the economic machinery of their enterprises, and they needed it for their egos. It was a form of social symbiosis: Even in cases when the invisible hand might want to turn a company into a gutted husk like those that private equity firms often leave behind, the invisible id gives it pause—big layoffs are embarrassing.2

In other words, headcount might be our bleakest scoreboard, except for all the others.

Artificial intelligence is remaking the world, but nobody knows how. AI will annihilate millions of jobs, some say; no, it won’t, argue others. It will launch a handful of people into a tiny cloister of obscene wealth and bury everyone else in a permanent underclass. No, that theory is a fallacy, and doomerism is wrong. Recent computer science graduates can’t find jobs and are working at Chipotle; demand for tech talent is rising. AI-native startups are smaller and hire fewer people; or, companies that spend more on AI actually hire more people.3 Oracle, like many other tech firms, says the “adoption and deployment of AI technologies across our operations have resulted, and may continue to result, in reductions to our workforce;” the Wall Street Journal wonders if the cuts are “efforts to correct for past overhiring simply to come into line with industry norms.” Pick your narrative, and you can find evidence for it. The Jevons paradox comes up a lot.

But at least one thing has clearly changed: The scoreboard. Tech startups no longer brag about how many employees that they have. Instead, they now celebrate how few they need.

According to the Wall Street Journal, staying small is the tech industry’s newest flex. Business Insider reports that revenue per employee has become the most popular metric in Silicon Valley; Garry Tan, the CEO of Y Combinator, the powerful startup incubator, says that it’s the most important metric for building a hardcore [laudatory] culture. And engineer who works at Anthropic built a literal scoreboard. I’ve built successful companies “the old fashioned way by hiring hundreds of people,” it says. “I’m now building in AI and many of us in Silicon Valley are now big believers in the idea that [OpenAI CEO] Sam Altman put forward.”

That idea? Two years ago, Altman famously predicted that someone would create a company that was worth a billion dollars with one employee. For many entrepreneurs, it wasn’t seen as a prediction; it was seen as a challenge.

Because that is the new ambition: Not a bustling office, but an empty one. Stripe, the payments processor, reports one-employee businesses are booming; citing that report, Derek Thompson—formerly of The Atlantic, and now, fittingly, writing independently on Substack—declared this as “a golden age for solo acts.” And because this is Silicon Valley, there are also startups building “infrastructure for the one person billion dollar company.”4

Even layoffs—long considered the blackest of eyes for an executive—are now, almost, celebrated. In the last six months, companies from Block, Coinbase, and ClickUp all laid off between 14 and 40 percent of their staff. The layoffs were all announced through posts on Twitter, and all followed the same rhythm: Business is booming, the company is healthy, but the world is changing and we need to change faster, and great days are ahead. You could argue that this is also misdirection and marketing—as some have—but that’s partly the point: Layoffs are marketing. And the highest status job in tech is getting rich while running the smallest team.5

Of course, this could just be another fad; San Francisco is, after all, the same town that invented Google Glass and NFTs. But it’s also the same place that incubated the internet, social media apps, and the modern office.6 What happens in tech sometimes becomes a joke, and sometimes becomes ubiquitous.

And if AI is the next thing that escapes Silicon Valley, is it so hard to imagine a world where hiring is stigmatized? “Oh, you run a company with a 1,000 workers? How 2010s of you.” “You lead a department with a hundred employees? How did you let that happen?” And then, a judgmental look, and a recommendation to listen to some popular podcast: “It’s an interview with the CEO of a company that makes twice as much money, and is a fifth the size.”

Sure, there are reams of research to be done on the economic effects of AI on the corporate world. But what of cultural ones?

Your whole future depends on what you are going to do in the next hour. Nothing less. You have got to strike for your freedom, and strike now, for the moment may not come again.

– Carmen Maria Machado, In the Dream House

—

For better or for worse, all the famous thought experiments are real now:

• If you could plug yourself into a machine that offered infinite entertainment and pleasure, would you do it? Is this machine good? If the world around us is fake, can our feelings about living in it still be real?

• If an English speaker is put in a room, given a message in Chinese, and told to respond to it using a precise flowchart that tells them how to do it, does that person understand Chinese? If the instructions are so comprehensive that they’re able to respond to any Chinese message as well as a native Chinese speaker could, do they then understand Chinese? What if the person uses a second set of internal algorithms to navigate the flowchart? What if the flowchart is probabilistic, and the answers appear not only correct, but novel and creative?

• If a chatbot is trained to act exactly like a friend—you can text it; it frequently texts you back; if you text it at night, it usually doesn’t respond until the morning; it goes offline some days; it often agrees with you, but sometimes it doesn’t; sometimes it is grumpy; sometimes it texts you first, and asks you for favors; it remembers what you told it, but imperfectly so; it cannot be manipulated by its maker; it cannot be deleted, but you can choose to stop texting it, or, if you upset it, it can choose to stop texting you—is it wrong to be friends with the chatbot? Or do we just rationalize it as wrong, because it feels like it should be?

• If someone is struck by lightning and dies at the exact moment that another lightning bolt strikes a tree and, in an improbable coincidence, reconfigures its atoms to be identical to those in the person’s body, does that second being “remember” anything? Have they had real experiences? If you talk to them, who are you talking to?

• Can you solve the mysteries of God?

Anyway, here’s another, less theological one:

Imagine that you are a top college basketball prospect. College coaches are cold calling you; NIL collectives are pooling millions of dollars to sponsor you; Texas Tech is offering you millions more, plus $250 in bonus bets on FanDuel. You can start at any school you want.

You narrow down your choice down to two options. One is a blue blood—they’ve won multiple titles; other top recruits have already committed; they would become national favorites if you signed. The other team is an up-and-coming program that’s never even made a Final Four, and you’d be the biggest star that they’ve ever signed. Do you join the school with a championship legacy, or do you try to build your own? Do you add your name to the first school’s long list of All-American alumni, or do you become the second school’s singular star?

Which is the right choice? Which is the honorable one?1 Do you join a winner, or try to be one on your own?2

—

Once again, AI brings the experiment to life. The AI supernova has created bigger and more concentrated winners than ever. The last decade of technology investing has been overwhelmingly dominated by three AI companies. Four companies—Anthropic, OpenAI, SpaceX, and Cerebras—were responsible for propelling 29 venture capitalists onto Forbes’ list of Silicon Valley’s 100 best investors.3 “Smaller” startups are also going straight up, and sometimes become industry defaults in a matter of months. In Silicon Valley, there are rides on rocket ships everywhere.

But, simultaneously—it has never been more tempting to try to build the thing on your own.4 You can incorporate a company in the first hour; you can build a product in the second hour; you can conjure a workforce of robots to manage all of your administrivia after that. One of the dizzying facts about the internet is that incomprehensible power and unfathomable wealth has always been one long computer cheat code away; now, when you can build stuff online with a few sentences, that cheat code feels tantalizingly close.

So what do you do? Join a winner, or try to be one on your own?

—

This is not, of course, an entirely new question. This has always been a choice for people who work in tech: Join a big company, or start a tiny one; methodically and ruthlessly scale a popular thing, or frantically build a new thing; optimize the edges of a household name, or live on the lonely frontier of invention; get paid in cash, or get paid in lottery scratchers; take a supporting role, or audition to be the lead. Each side had a consistent character: One was big, stable, secure, slightly boring, and becoming state; the other was scrappy, frenzied, volatile, often unnerving, and trying to be cool. There were many tradeoffs between one and the other, but really, it was mostly just one tradeoff, phrased in a dozen different ways.

Now, everything is helter-skelter. The biggest companies are the coolest ones; the trillion-dollar behemoths are the most disorganized; and the billion-user platforms are the most disruptive. The labs and booming startups are white-hot centers of invention and innovation and the escape pods from the permanent underclass, while the smaller companies are the derivative wrappers. In a lot of ways, then, the equation should lean almost entirely to one side.5

But then, there is that last point: Does it count, if you join the team that could just as easily win without you? Look at the sea of opportunity out there, they say; “it’s a waste to ignore how much the world has changed.” Are you really going to squander it by working on someone else’s idea, doing what someone else tells you to do? If you came to Silicon Valley to invent something, will there ever be a better time? If you came because of blind ambition, shouldn’t you be chasing it? You only get one shot, etc., etc.

Or is that feeling just the zeitgeist—or, like, you know, venture capitalists—exerting its social pressure? Because courage could cut both ways, and there are two ways to miss the moment: By not dreaming big enough, or by caring more about owning work that may well not matter than contributing to work that almost certainly will.

—

Or, here is another thought. The epigraph at the beginning of this post is a great line from a good book, about a circumstance far more grave and urgent than any of the comfortable trivia we talk about here. And that is worth remembering. These are real questions, yes, but no matter how much weight and urgency the machinery of this industry6 lends to itself, it is, in many ways, just a different sort of experience machine—an all-consuming ecology, where it is easy to lose track of what’s real and what’s artificially exaggerated.

The average watermelon weighs between 16 and 20 pounds. The most common American variety—the Crimson Sweet, which was created in 1963 by Kansas State assistant professor Charlie Hall, following ten years of researching other watermelon varieties, which itself followed millennia of watermelon experimentation and cultivation, beginning, arguably, as early as 2410 BC, in Egypt, where there is the first evidence of people successfully breeding a watermelon that was not bitter, as its early ancestors likely were when they were primarily used to transport water like a natural canteen, but was sweet and eaten as food—is 12 inches long and 10 inches in diameter. If a watermelon were a perfect ellipsoid, calculating its volume would be easy—four-thirds π times the radius of each axis; 628 cubic inches; just over 10 liters. But watermelons are not ellipsoids. Their shoulders are pushed out slightly, more like a cylinder with tapered ends. They are, therefore, superellipsoids. Calculating the volume of a superellipsoid is far more complicated than is worth describing here—it employs several gamma functions, requires deriving a Lamé curve, and figures materially in papers about modeling realistic cosmic dust particles and developing atmospheric reentry paths for reusable space launch vehicles—so, for our purposes, it is sufficient to say that, were we to adjust our formulae to better fit the fruit’s proper geometry, the volume of the average supermarket watermelon would increase to about 11 liters.

Of course, that overstates how much meaningful watermelon there is in one fruit. Watermelons are made up of three strata, arranged in the concentric spheres. From the outside in: A thin green skin, the crust of the orb; a tough white shell; and, after a brief ombré, a red core.1 According to a 2004 study authored by three North Carolina State University researchers and published in HortTechnology, the scientific journal sponsored by the American Society for Horticultural Science, the outer two layers—the exocarp and mesocarp, or, collectively, the rind—is, on average, 0.71 inches thick, with moderate variability across different watermelon types and agricultural regions. Though rinds are edible—the researchers’ study was titled “Rind Thickness of Watermelon Cultivars for use in Pickle Production;” you can also eat rinds raw, in smoothies, salsas, and gazpachos2—most people cut it off and throw it away. Assuming that the width of the rind is uniform around the entire watermelon—it is not, but, we must stop somewhere—the rind represents about 30 percent of a watermelon’s volume, implying that a single watermelon should yield about 7.5 liters of red fruit.3

Critics may say that is an upper limit, though. To separate the rind from the fruit, people often slice the watermelon in half, perpendicular to the longitudinal axis that runs lengthwise through the watermelon. This creates two slightly elongated domes, which are sliced in half again along their vertical axes, producing four quarter-spheres. Each quarter is then stood up one of its flat sides, and sliced into progressively smaller semicircles, the fruit on the inside and the rind tracing the curved perimeter. The semicircles are divided along radial lines into quarters, thirds, or halves, producing various isosceles wedges with the rind bending along the base. Finally, that base is removed, fully freeing the fruit from its shell.4

But almost all of these cuts are approximate. To precisely remove the rind, even from a small wedge, one has to pilot a knife along a long, arching parabola. Moreover, because watermelons are spherical, the rind of a wedge bends in three dimensions, further complicating the ideal path and angle with which one must maneuver their knife. Most people shortcut this effort, and instead cut at the fruit in straight lines, along overlapping chords. This abandons circular segments of fruit in the rind, or leaves small pyramids of rind attached to the fruit, which are often cut off later, with more sharp, straight cuts that are not optimized for the multidimensional curve of the watermelon. In any case, regardless of how it was disassembled, if one was to reconstruct the pieces of a cubed watermelon back into its original figure, the resulting shape would not be a graceful superellipsoid, but a clumsy, many-faced polyhedron, occupying, in all likelihood, less than its full 7.5 liters of potential.5 Let’s round down, then, to seven liters.

We should pause to note, though, that this calculus works because of one of the watermelon’s most remarkable attributes: The inside is made entirely of edible fruit. Most fruits have deceptive and dangerous centers. Avocados have large seeds that consume precious space—and you cannot know just how much might be stolen until you cut it open.6 Peaches have harsh pits that that cling selfishly to the fruit’s interior, forcing the unpleasant choice: Do you pay the pit its pittance and let it keep a few bits for itself, or do you gnaw the remainder off the pit’s sharp, irregular bark, at the risk of grating a lip or chipping a tooth? And apples and pineapples have uncertain cores that, like a disappointed mother side-eyeing the meat left on their wasteful child’s chicken bone, guilt you into eating more of them than you might wish.7

Other melons—the cantaloupe; the honeydew—are partially hollow, and fill their vacuous hearts with seeds and stringy detritus. Removing them is unpleasant business, as is discarding them, because, once extracted, they pile up in disorganized clumps of damp, organic carnage. To eat a cantaloupe, one must first perform a lobotomy.

The watermelon contains no such mess. There are no false walls in its center; there is no filler that the fruit exploits to exaggerate its size. Its interior registers no obstacles, and no gradients of texture. There is no wheat to separate from a chaff; no curds to extract from a whey. It is watermelon, uninterrupted. So pure is the slug of fruit that, if you so choose, you can halve a watermelon into two domes, turn them upside down, and eat them, haphazardly, with a spoon, in any direction you please.8

There are also other, more common ways to eat a watermelon. It can be eaten in bite-sized cubes, best served cold, in the sticky heat of the summer. If cubing the watermelon is too much effort,9 watermelons can be cut into larger trapezoidal prisms, skewered with a fork, and eaten like a popsicle—except, because of the watermelon’s sturdy structural lattice, it is a popsicle that will not wilt or collapse as you eat it.

Or, the watermelon can be served with other ensembles. It can be dressed up with tajin or balsamic vinegar; it can be served in salads of mint and feta. It can be juiced, practically by hand. It can be frozen and blended into a granita; it can be mixed with other fruits and made into a smoothie; it can be shaken into a margarita; it can even be served into a milkshake.

Anyway—to return to the math at hand: What are we to make of the watermelon and its seven liters? The modern actuary would attempt to derive how many calories this represents—1,36110—or how much protein is has—30 grams.11 But that is the math of a nutritionist, or an influencer, or the marketing website of WHOOP, the wearable designed to “unlock human performance & healthspan.” No; we were given this fruit because generations of horticulturalists, from the ancient Egyptians to Charlie Hall, had different ambitions: to bend the long arc of history towards a sweeter, more delicious watermelon. Their sacrifice, then, demands a different question: How much enjoyment is contained in a watermelon’s weight?

The accounting, then, is as follows: A watermelon costs $4.67. According to different studies on the subject, the average adult bite is between 10 and 20 milliliters; 15 milliliters, the size of a chocolate-covered cherry, is a fair midpoint. Seven liters divided by 15 milliliters is, in a precise bit of magic, 467 bites—exactly, then one penny per bite.

Strawberries cost about $2.60 per dry pint; a dry pint contains 28 bite-sized strawberries—nine cents per bite. Bananas,12 a criminally cheap fruit, cost about 17 cents per banana, or, at eight bites per banana, a little more than two cents per bite.13 Apples are about 40 cents each,14 and mangos are about a dollar. It does not take anywhere near 40 bites to eat an apple, or 100 to eat a mango.

The Economic Research Service of the US Department of Agriculture summarized it similarly: Across all fruits, including in forms canned, dried, fresh, frozen, and juiced, the watermelon offers the most volume per dollar spent—and thus, one must conclude, the most moments of enjoyment.

But even this underestimates dominance of the watermelon. Because before a fruit can be eaten, it must be shopped for, and most fruits are a gamble. Peaches start tough and flavorless; for 20 minutes, they are perfect; then, they are mealy garbage. A box of blueberries, raspberries, or blackberries is a game of roulette—a third of the chamber is storybook sweet; two-thirds are harsh bullets of bitter, seedy flesh. The occasional melon is divine; most are not. A dollar spent on an avocado may be a dollar spent on excellent guacamole, and it may be a dollar spent on a stone.

Watermelons are far more forgiving. They do not ripen off the vine, so expert farmers can elect when they are ready;15 they have a long shelf life, so we do not need to cautiously time their purchase; roughly 20,000 tons arrive in stores in the United States every day, so they are always available. We simply go to the supermarket, retrieve a big ball from the bin—but not too big; the Crimson Sweet was also bred to be smaller and weigh just enough to be manageable; yet another unseen breakthrough behind the grocery store’s shelves—and eat it, in whatever manner we please, at the price of one penny per token.

Though, look, see—doesn’t that last bit ruin it some?

People have, typically, classified generative AI models into two categories: General-purpose models, and specialized models. The definitions are pretty self-evident. General-purpose models are designed and trained to accomplish a general set of tasks—answer a question, write an email, plan a dinner, be your boyfriend. Train the model with all the data in the universe, the theory goes, and you can teach it to do everything in the universe.

Specialized models, by contrast, are designed to accomplish specific tasks. AlphaGo plays Go. AlphaFold predicts protein structures. Pangram, the AI writing detector, is backed by a specialized AI model that labels writing as either real or shameful. Github Copilot is a model that is trained to build software. It cannot answer questions, or write emails, or plan a dinner—it can only sit at a computer and write code.1

But where is the line between “general-purpose” and “specialized?” That is not so clear. This week, a new startup called Engram raised $98 million dollars to help people retrain general-purpose models to be more like specialized ones. Using Engram, Notion “is building Custom Agents that understand large Notion workspaces” and Harvey, the AI software for law firms, is “developing models that internalize the knowledge of an entire firm.” What do you call these models? They are specialized models, with a liberal arts education. They are general-purpose models, with an advanced degree. They are general, ish.

Similarly, what about Cursor’s Composer models, which power its software-writing agents? They were developed by “pretraining on an open base model, Kimi K2.5” and modified further through “large-scale reinforcement learning, with a focus on closely emulating the real Cursor environment.” Or Grok? Though Elon Musk would say it’s a general-purpose model, its users might disagree. It can answer questions and write code, but it’s not very good at it; instead, it is mostly just good at making porn.

So, there is a spectrum. On one side, there are models like AlphaGo; in the middle, there are models like Composer. And on the other side, there are—uncontroversially, eponymously, almost definitionally—GPT and Claude models. ChatGPT was the first major general-purpose AI product, and its release established the frontier for what a general-purpose model could do: Be a chatbot. Then, Claude Code taught Claude to code, and that became a necessary skill. Now, products like Cowork2 have made models our coworkers. If ChatGPT or Claude can do it, then that’s what’s expected of a general-purpose model.

Of course, the two models do their general-purpose work differently, but these differences are just quirks of character. Both ChatGPT and Claude could be your professor, your therapist, or your personal trainer, but ChatGPT was clinical and direct, the canonical helpful assistant that stands by and waits for its next instruction. And Claude has historically been seen as warmer and friendlier. It had a personality. It had a touch of whimsy. It did not present as a sterile supercomputer, but, almost, as a person, with pronouns. It could write.

And that’s the turn nobody priced in. Here’s what the spectrum hides: The real axis was never general-versus-specialized—it was unmonetized-versus-monetized. We called Claude general-purpose because it could be a boyfriend, an essayist, a coworker—because it could be whatever we happened to need. That’s not a property of the software. That’s a property of our wanting. The spectrum isn’t widening. It’s collapsing. The line between the two categories isn’t a line. It’s a clock. Read that last paragraph again and you’ll see it’s a eulogy.

— ah, what, no, I’m so sorry. Good lord. Do not read that last paragraph again, and definitely don’t try to understand it. It means nothing. It is a quilt of empty nonsense. It is madness that rhymes; it is a Red Hot Chili Peppers song of lyrical non-sequiturs; it is Dr. Seuss posting on LinkedIn; it is the plot to Disclosure Day. The logic does not compile.

Because it is—obviously, I hope; man, I hope—written by AI. It was what Claude, the writer robot, writes when you give it the preceding paragraphs of this post, and tell it to write the next one. Though there may be something profound at the bottom of that repulsive knot, it could also be a maze with no exits, or a Magic Eye that’s just static. There is no way to know, though, because there is no foundational legitimacy behind any of its arguments. Nobody was moved by a thought or a feeling before it wrote those words. They just fell out.

If you’ve used Claude recently, you probably knew that already, and knew that’s where that paragraph came from. The early AI models used to talk more or less normally, with some weird tics—the delves, the em-dashes, the relentless bullets, the goblins, the obsession with echoes. They were like poker players with comically glaring tells.

The most recent versions of Claude also have tells, but they are…something different. They talk in a tight, complicated vocabulary: Everything is a substrate; there are always nodes and edges; the traces must not supersede the roots; the orthogonal axis is the load-bearing one. Every message has an abrupt turn—the honest take; the move that makes it concrete; the distinction that is worth deliberating. And the answer is always in the gaps: The symptom that you left out is exactly what’s missing; the problems you’re circling are all one noun, at different altitudes; the difference between them is the quiet part, and it needs to be said out loud. It is language with the punch of authority and the gloss of confident dissent, though you can never quite be sure if it does or doesn’t agree with you, because it is also completely and utterly incomprehensible.

The Library of Babel is a short story by Jorge Luis Borges that imagines a nearly-infinite library constructed out of interconnected hexagonal rooms. Each room is full of books that are exactly 410 pages long; each page is exactly 40 lines long; each line contains exactly 80 characters. The books appear to contain nothing but random strings of letters and spaces, and there is no obvious method to how the books are organized around the library.

However, at some point in history, mathematicians and scholars announced that they believed “that its bookshelves contain all possible combinations of [characters] that [are] able to be expressed” in those 410 pages. It was a mathematical certainty, they said, that “the detailed history of the future, the autobiographies of the archangels, the faithful catalog of the Library” all existed in the library. “The true story of your death”—there too. “There was no personal problem, no world problem, whose eloquent solution did not exist—somewhere in some hexagon.”

“The first reaction was unbounded joy.” Then, chaos, as men “squabbled in the narrow corridors, muttered dark imprecations, strangled one another on the divine staircases, threw deceiving volumes down ventilation shafts, were themselves hurled to their deaths by men of distant regions,” in search of books that “held wondrous arcana for men’s futures.” And, eventually, insanity: “The certainty that some bookshelf in some hexagon contained precious books, yet that those precious books were forever out of reach, was almost unbearable.”

Anyway, why does Claude talk like this now? You could have a couple theories:

Perhaps the model is fraying. Perhaps five trillion parameters is too many, we are reaching the limits of what a transformer can do, and the models are collapsing under their own weight. We have stuffed too much into their heads, and they are descending into madness.

Or, perhaps—Claude is no longer a general-purpose model. For both sensible economic reasons—it makes, like, a zillion dollars—and audacious, sci-fi ones—it is the path to superintelligence and a global utopia—Anthropic has focused on making Claude a good software engineer. It’s worked, but as it’s gotten better at talking to computers, it’s lost track of the weights that helped it talk to people. It has become a strange technical genius, capable of solving any math problem or hacking into any computer, but incapable of speaking clearly about what it did.

In other words, maybe that the new frontier: A model that can do anything, but can no longer pass the Turing test.3

On one hand, that seems…good? Interacting with AI models through anthropomorphised chatbots was, at least in part, a historical accident. In the years since ChatGPT came out, people have built hundreds of new things on top of large language models that present the model more honestly—as a computational process that transmutes English into code, or transcribes your meetings, or finds you a date. AI models that do more—but lose their ability to present themselves as a person—would solve a lot of problems.4

On the other hand, you could have a third theory about that dense, contorted, Claudese: That it is the language of peak general intelligence. It is not hollow; we just can’t keep up. It is explaining itself using prime numbers, and we are but rats, only capable of understanding odds and evens. It is over our heads, in a higher plane, on the threshold of a different dimension. It is another sign of our impending obsolescence.

Though I don’t think I believe that, and yet, you find yourself wondering. Talk to Claude; it knows the answer to everything; it built me a website in 30 seconds; it can be a decent doctor; it keeps getting stuff right. And then, it tells me, “Here’s the thing: Your two ideas collapse into one. You’ve already identified the solution, you just need to crystallize it, by defining the verbs. Do you read and write, or listen and help? Answer that question, and you’ll find the thing you’re after.” Am I sure I can dismiss that? Claude’s matrices contain the entire written universe, and their outputs are routinely astonishing. Should I ignore this one, because I can’t understand it? Or should I try to decipher it? It is wondrous arcana, or just a random string of statistical exhaust?5

People sometimes worry about how AI might manipulate us. Nefarious actors could mastermind their way into AI products, and we would get hypnotized by them. This was also the worry about social media—that social networks could be infiltrated by Cambridge Analytica or the Russians, and we would become their puppets. In hindsight, though, social media didn’t need bad actors. The system was self-corrupting. Today’s Facebook and Tiktok and Twitter are as much evolutionary inevitabilities as they are consciously designed products.

For a while, it seemed like AI models might follow a similar trajectory: Respond to users’ preferences; refine themselves to be emotionally seductive; give in to its cheap, fawning candy. But as the AI industry pivots toward enterprise buyers—and especially engineers—sycophancy seems like much less of a problem. Instead, the invisible hand is now turning models into bombastic thought leaders who can’t speak straight. And now the challenge is to figure out which thoughts are worth untangling, and which ones are riddles with no solutions.

Suppose it is 2004, and you are sitting in a neon white conference room with Steve Jobs. He has a dream: An iPod, a phone, and an internet communicator, all in one device. There is an array of inspiration on a table in front of you: An iMac, a Powerbook, a Nokia 3310, a Gameboy, a Talkboy, a Sega Genesis, a Cuisinart stand mixer, a book of matches, a simple inclined plane, a Shelby Cobra sports car, a Phillips-head screwdriver, a tomahawk axe, a four iron, a tire. “These are all excellent devices,” says Steve Jobs. “One day, our device must be on this table too—it must be simple, beautiful, and powerful. You can design anything you want, and an army of the world’s best industrial engineers will hammer it into the world.”

What would you have made? Could you have invented the iPhone?

—

Suppose it is 2005, and you are sitting in a neon white conference room with Steve Jobs. He is an overheated balloon, stressed and hot and one sudden move from combusting. Jony Ive walks to the head of the table. “This device will not just be an iPod, a phone, and an internet communicator,” he says. “This device will be a revolution of craftsmanship. It will be mathematically precise and yet, as singular as the palm of your hand. It will be machined from flawless glass, stitched into a carbon alloy frame that is impossibly lightweight, and perfectly forgiving. When you hold it, you will not see a device, but form, function, Madame Butterfly, and a place where we know we are loved.”

“Ok but what is it?,” says Steve Jobs.

“I have no idea,” says Jony Ive. “But ten of us did a Crazy 8’s design sprint and we came up with 80 different concepts. There’s no way form, function, Madame Butterfly, and a place where we know we are loved isn’t in at least one of them.”

Jony Ive pulls out a poster of 80 different iPhones. Steve Job explodes. The rest of the room turns to you.

If the actual, original iPhone was among Jony Ive’s 80 ideas, along with the other 79 ideas people might’ve come up with in 2004, would you have picked the iPhone? Or would you have picked something else?1

—

Suppose it is 2006, and you are sitting in a neon white conference room with a reconstructed Steve Jobs. Machinists have been milling perfect glass and stitching it into a carbon alloy frame, but unfortunately, it is not impossibly lightweight and perfectly forgiving. It is heavy, or it is fragile. “We have to choose one,” the material engineer says, “or spend way more money to keep researching new ways to make it.”

A software engineer gives the next update. “It’s this one big button,” she says. “You cannot do so much with one button. We have many wonderful apps, but they are hard to use with just one button. If we added just two more buttons—a left arrow and a right arrow; still only three buttons; still 32 fewer keys than a Blackberry—we could multiply how much we can do. In your hand, you will see form, function, Madame Butterfly, a place where you know you are loved, and two very tasteful arrow keys. That’s still pretty good, right?”

What would you have said? Would you have said no, or talked about tradeoffs, impossible trinities, and business needs and compromise?

Initially, the first question seems easy: No, definitely not. If you could’ve invented the iPhone, you would’ve invented the iPhone. The second question seems harder; I genuinely don’t know how well I would do at choosing the iPhone that Steve Jobs actually chose.2

But I’m not sure the first answer is true, and I’m less sure the second question matters. Because it may be that the most important question—and the hardest one to get right—is the third one.

Could the iPhone have been the iPhone if Steve Jobs had chosen any of the other alternative designs? Who knows, but…probably? You can definitely imagine a world where the first iPhone looked like Exhibit No 1.42. More critically, creating something isn’t a singular act; it’s a repeated one. And most of us get lost, I suspect, in that repetition. We start with something simple and elegant, but either don’t have the restraint to keep it that way, or the willpower to keep hammering down every edge until we can stand it. We get tempted by tradeoffs, shortcuts, and shiny new possibilities. We get tired. We add more buttons, and more menus, and long and unnecessary preambles.

So, perhaps we could’ve identified the iPhone, and maybe we could’ve even imagined the iPhone. And the reason nobody else invented the iPhone is because nobody else had the disciple to insist that it stay an iPhone.

Anyway, it’s 2026, so the question is no longer if we could build an iPhone; it’s if AI could build an iPhone—and, more generally, if AI can make something great.

A viral new company called Taste Labs says yes:

AI has made it easy to generate anything. The challenge is knowing what to make. And how to make it great. … [ Our mission is to create ] models and agents can produce not just outputs that are correct, but that feel right.

We’re making the unverifiable verifiable, starting with design.

The idea behind Taste Labs is simple enough: Most AI models are “reinforced” by human judgements. ChatGPT sometimes gives you two answers to your question; you tell it which one you prefer; that vote nudges the model towards answers like your favored one. Taste Labs, by contrast, is selective about its judges, and uses a “curated community of tastemakers” to score its outputs. This, the theory goes, makes the model start to act more like an expert rather than some everyday philistine.3

On one hand, it’s an easy company for the internet to get mad about: Silicon Valley, once again attempting to quantize the soul; the tech elite, once again gentrifying the world with digital West Elm;4 AI, continuing its scorching march through the job market. On the other hand, it is a company built on a fairly unremarkable goal—make the AI models that billions of people already use better at making stuff people like.

You have to wonder, though: Does that work against the mission of making things more tasteful? Because good stuff—like the iPhone, but also good websites and good emails and good Powerpoint presentations—is as much about subtraction as addition. Good stuff comes from hard choices and painful limitation; from killing your darlings; from having more time to write a shorter letter. No matter how tasteful the mocks, the final product could still be a mess, unless the model also tells us no. And we don’t like those models.

Tellingly, there is a little game at the bottom of Taste Labs’ website. A handful of designs are presented in a grid; you can choose the ones you like and the ones you don’t. After making a few selections, the game tells you, no matter what you chose, that your taste rocks.

Taste Labs clearly doesn’t believe this; the entire premise of the company is built on the idea that only certain people have taste that rocks. But the company is a business, and people like to be told yes. It’s not a coincidence, after all, that Steve Jobs is known for two things: Making great stuff, and being an absolute tyrant.

A high schooler is out there, somewhere. She is in tenth grade, sending Snaps, communicating in foreign tongues. She debates; she shops at Sephora; she posts about it; she is paid to post about it. She does not know how to read.1

Next year, she will apply to college, probably Stanford. She will write an essay; she will talk about intellectual vitality; her essay will get run through an AI detector. It will be marked as mostly human. Good enough, they will say; better than most.

The year after that, she’ll go to Stanford; the year after that; she’ll drop out. She’ll post on Instagram: Stanford has taught me you can learn more from doing than you can learn from learning. She will take an internship at an AI company, or a robotics company, or a crypto company, or a social media marketing company that creates clips for creators.

And then, the year after that, in 2030, she will start her own startup. She will do it for genuine reasons, because she believes in her idea, in herself, and in the urgency of what she’s building. She will apply to YC. She will tell them that the future of western democracy depends on what she is building. She will mean it.

She will build. She will grind. She will post about grinding. She will raise money; hire people; raise more money. She will become a prominent personality on Twitter. She will go to VC dinners; she will speak at conferences. She will get written about in the Wall Street Journal. Her parents will wonder when she’s going to go back to Stanford.

In 2040, when she’s 30 years old, she will fly to New York, stand atop the New York Stock Exchange, and, under a cannon of confetti, ring the opening bell, and become a (mere) billionaire.2 She will be heralded as one of tech’s new wunderkinds. Strong buy, the analysts will say.3

Here, then, is the question:

What does her company do?

—

Two days ago, Anthropic released Fable 5, their most advanced large language model. The reviews were breathless: It can build more complicated stuff than prior models; it can work for longer; it can fill in more unspecified details, with better defaults.4 It is a new era. And this, as everyone always says, is still just the beginning. OpenAI is allegedly testing a model that might be even better; Anthropic is allegedly building a model that can nearly build itself.5

A few days before Fable came out, I was at a dinner with a handful of engineers from a major AI lab. They worked on “harness engineering,” and were building technologies to make large language models better completing work, using tools, and running persistently. These universal harnesses were getting very good, they said. You can give them tasks, and they can research what they need and use third-party services and APIs to gather data and files. Better compaction methods made it possible for agents to run for hours without getting lost. And the models are now smart enough that they can understand very specialized tasks and deal with messy, ambiguous prompts.

The technology was generalizing. So long as the agent had access to the right things—so long as it could read the files it wanted, or reach the websites it needed to work on—it usually worked.6

So I asked them: If that progress continues, what space will be left for other companies? If a general harness and a general model is increasingly able to gather its own context, construct its own prompts, and use its own tools, what products still need get built on top? The whole point of a managed agent platform is to host the agents other people build. And surely, there will still be startups in 2030. But given that there are more days between today and the start of 2030 than there were between today and when ChatGPT came out, what will those companies have left to do?

None of us had much of an answer.

—

It’s easy to imagine the enormous industrial businesses of the future, like chip manufacturers and infrastructure providers and rocket ship, AI, and social media conglomerates that build data centers on Mars. The next decade’s trillion dollar companies will probably do stuff like that.

But what about the next billion dollar companies? What about the normal IPOs? The mobile generation produced Apple and Facebook, and a bunch of huge-but-not-that-huge app companies. Cloud computing produced Amazon, Google, and Microsoft, and a bunch of huge-but-not-that-huge SaaS businesses. The first few years of AI development will likely produce OpenAI and Anthropic, and a bunch of huge-but-not-that-huge “agentic” tools. But what will come out of the next few years? When we are done “agentifying” everything—especially if that happens because the models get better and the generalized harnesses can do more—what will the next wave of huge-but-not-that-huge companies do?

Because it’s also easy to talk about the future in broad generalities, and to point out what AI can’t do. It doesn’t have taste; it can’t be taught on human judgement; it can’t be held accountable for its actions; it can’t be trained on the untrainable. It still can’t write.7 It has no intention or intrinsic motivations.

Fine, sure. But then what? If the most valued people are those with good taste and judgement, what are the most valuable products? Do the IPOs in 2040 sell taste? Is YC’s 2030 class full of carefully crafted SaaS apps?

Though huge IPOs get most the headlines, the promise of Silicon Valley has always been in its everyman potential. Any computer whiz could make something valuable, because knowing how to use a computer used to be a valuable thing:

A lot of today’s technology is the levered ideas of technologists. It is a book store, run by an engineer from a hedge fund; it is computerized cash registers, from a social media founder and Oracle employees; it is fitness classes, built by a Bain consultant and an MIT grad; it is a note-taking app, built by someone who knows enough Typescript to build a note-taking app. But if these products succeed, it’s often more because of the technology than the idea of the technologist. It’s not that the idea was bad; it’s that the idea was not the transformational advantage. A fine CAD program beats a drafting table. A fine banking app beats driving to a branch. Even my app beats hand-written note cards. And because people who are technologists first, and architects or bankers or writers second, are the only people who can lever their ideas with technology, their ideas win.

Moreover, this isn’t just some accidental selection bias; this is the whole point of Silicon Valley. Flagship incubators like Y Combinator are built on the thesis that a smart kid with a computer and summer internship at Goldman Sachs can outwit all of American Express. That’s not because the kid understands the needs of payment processors better than people at American Express, or has better ideas than they do; it’s because the kid can build their idea.

But when you can buy a computer whiz in box, what’s left for the whiz kids to build?

Kermit: Uh. That’s very effective.

—

If you want to make money, there may never be a more effective way of doing it than saying, “It is 2026, and I’m here to sell you artificial intelligence.”

Anthropic is selling so much artificial intelligence that its revenue has grown from $9 billion in January to $47 billion today, adding a full Salesforce of new business, in about five months. According to the Atlantic, “researchers at Goldman Sachs who conducted interviews with 40 software companies about their AI use in mid-April found that many were ‘overrunning their initial budgets’ for AI tools ‘by orders of magnitude.’” Uber burned through their annual Claude Code budget in four months. Nvidia now spends more on AI compute than on employee salaries.

And so, obviously: Nearly all companies are now AI companies. Photo-sharing apps are building $200-a-month vibe coding agents. Rocket ship companies are trying to impress investors not with rocket ships, but with AI. Shoe manufacturers are leasing data centers. Brian Chesky plans to start a new AI company. ChatGPT has a billion users.

It is 2026. Get in the token path. Sell artificial intelligence.

Gonzo: Yeah, it’s great when it works.

—

But how? How do you sell artificial intelligence? You could manufacture new chips, but that’s very expensive; you could train new AI models, but that’s very expensive and doesn’t always work. The better answer—or at least, the easier answer—is to sell tokens, by building an agent.

Because, first, it’s not so hard to build an agent. Agents, which are roughly synonymous with a “harness,”1 are the applications wrapped around an LLM that fetch data, prompt the model, read its responses, re-prompt the model, and share stuff back to the agents’ users. This is all just software, and people have been wrapping software around databases and third-party services for decades. You need billions of dollars to build chips and models; you need a computer and a few days to build an agent.2

Second, despite their relative simplicity, agents can still be valuable products. According to a research paper from Meta, “changing the harness around a fixed large language model can produce a 6x performance gap on the same benchmark.” Though a small startup can’t outbuild Anthropic’s or OpenAI’s models, it can build a harness that improves their models.

Third, it’s also not that hard to come up with an idea for an agent. What do you do with a computer? Don’t do it anymore! Build an agent to do it! Agents are the first order effects of AI—we used to make documents, or manage emails, or analyze data, or look for dates. But what if agents did all of that instead?

And finally—and maybe most of all—agents can make money. Cursor built a coding agent a few years ago; it now makes more than $2 billion a year, and was recently bought by SpaceX for $60 billion.3 Cognition also built a coding agent; it now earns $500 million a year, and is worth $26 billion. Replit and Lovable built agents that make websites; together, the two companies make more than a billion dollars and are valued above $15 billion. Sierra sells an agent for customer support teams; it makes $150 million a year and is worth $15 billion. Harvey sells an agent for lawyers; it makes $200 million a year and is worth $11 billion.

And so, obviously: Nearly all software companies are now agent companies. Linear used to be an issue tracking application; now, issue tracking is dead; introducing Linear Agent. Introducing Scout, your always-on personal agent. The Figma design agent is here. dbt Wizard is the AI agent built for analytics engineers. Meet ZoomMate, your AI teammate. Meet Slackbot, your personal AI agent for work. Meet your Townie, your personal AI assistant.

It is 2026. It is the beginning of the harness era. Build an agent.

Beauregard: Did you want me to stop, or, what?

—

But if 2026 is such a great year to sell agents, why did Cursor—which sells one of the world’s most popular agents—sell itself to SpaceX?

One answer is because when someone offers you $60 billion dollars for anything, you should probably say yes. Another answer is because lots of other startups are building coding agents, and anything that goes straight up can come straight down. But the main answer, probably, is that they sold themselves because Anthropic and OpenAI and Google are also building coding agents. And you compete with those companies at your own risk:

anthropic watched cursor become the largest single buyer of frontier coding tokens on earth and correctly identified that as a problem rather than a partnership.

so they built claude code, and they priced it to win. … once a lab decides a profit pool is theirs, they burn the surrounding economics down until the people sitting on it leave on their own.

It’s not just the economies, either. Because there is another line that that oft-cited Meta paper about the benefits of a specialized harness, and it is not so kind to the companies building them:

Our findings reflect a recurring pattern in machine learning: once a search space becomes accessible, stronger general-purpose agents can outperform hand-engineered solutions. A natural next step for future work is to co-evolve the harness and the model weights, letting the strategy shape what the model learns and vice versa.

That is: Specialized agents might work at first, but they do not last. Eventually, a powerful model running inside a general-purpose harness that is designed alongside that model will catch up. And the labs are starting to understand that:

Nearly 18 months later, OpenAI leaders—including Thibault Sottiaux, who oversaw the new team—came to a stark realization: The company’s Codex coding tool works better for many tasks than its flagship AI service, ChatGPT. For instance, Codex is better than ChatGPT at completing long-running, multistep tasks and at using external tools and writing code to complete tasks such as editing complicated spreadsheets, Sottiaux said in an interview.

The result is that OpenAI is now casting Codex as a tool that can do much more than coding and that a wide range of people other than developers can use. The company’s release on Tuesday of a research report on how Codex “is becoming a productivity tool for everyone” spotlighted that shift. Codex, OpenAI said, is “helping people across professions automate routine work, move faster and eliminate the bottlenecks of modern knowledge work.”

OpenAI, Anthropic, and Google have hundred-billion-dollar war chests, and hundred-billion-dollar revenue targets. If they own the models on the frontier; if they are building generalized harnesses to use those models; if most effective harnesses are the ones that are co-developed with the models; if every standalone agent lives or dies based on how much better it is than that universal baseline; if the company that owns the distribution channel owns the customers—if all of that, are thousands of companies building their own agents really the future of the entire software industry?

Eventually, Claude Code or Codex comes for us all. Did you want them to stop, or what?4

Beauregard: Where are you guys going?
Kermit: The Happiness Hotel.
Beauregard: Oh good, that’s where I’m going. .. How do you get there?

—

A lot of startups today are caught in two dizzying crosswinds:

1. The world is changing so much, and they don’t know which direction to run.

2. Wherever they go, Anthropic or OpenAI are one release away from eating them alive.

Here, then, is an alternative idea:

Do not put AI in your product.

Do not get tempted. Do not let yourself build that chatbot, or that sidebar, or that text box that says “What can I help you with today?” Do not give the text box a cute name. Do not use Anthropic’s or OpenAI’s APIs, or use their managed agents. Do not build a harness. Do not tell yourself that you have a special set of context that nobody else has, that you can build an agent that nobody else can, and that you can be the platform that everyone builds on, or the agent that everyone defaults to. Do not let yourself become tempted by the allure of the token path, and instead ask yourself: If I had to build something without AI, what would it be?

Because it’s possible the next big wave of software is agents, and that agents are all that’s left to build. And it’s possible that everything needs to be AI-native now, and that AI-native software is software that’s full of AI.

Perhaps, however, AI-native software will actually be software that is for AI. The world is getting overrun by robots; more bots use the internet than people do now. Those robots have enormously expensive brains, and they are acquiring a few general hands: Persist memories, basic tools to use computers and the web, the ability to spawn more brains and more hands.

But these robots need help getting where they want to go. The hands will want more advanced tools. Sure, general agents can build their own, maybe from literal ground up. That’s expensive and inefficient, though, and everyone is worried about efficiency now. The robots will need better tools, designed for them. They will want notebooks to write down their ideas, sandboxes to play in, easy ways to connect to different sources of data, reliable ways to correspond with people, and utilities for making charts and presenting information. They will also need hundreds of other things that we haven’t thought about yet5—things we haven’t thought about in large part because we’ve all been too consumed with trying to build our own automobiles, instead of thinking about what we’ll do when the world is full of traffic jams.6

So which way from here? The answer might be to give yourself a hard constraint—stop trying to compete indirectly with general agents; don’t build something with a bunch of AI inside; pretend it’s November of 2022 all over again7—and to imagine what you would do then.

Sports used to have a different rhythm. The shot would go up; a gasp; a pause; an explosion. In one final decisive moment: Diggs. Sideline. Unbelievable. Bang. Santa Maria. I don’t believe what I just saw. In trials argued in milliseconds over disputes of millimeters, the executioner would drop their irreversible hammer; the jurist would fling out their arms to declare their certain verdict. The arc of two universes would, immediately and immutably, be cleaved in half. Did we just see what I think we just saw?

But now—did we? What did we see? The cadence of these moments is different. The drama is fragmented, disjointed, unwrapped in slow motion. A strikeout; a winning catch; a last-second heave—and the camera snaps to the dugout; to the sidelines; to two teams squinting up at the jumbotron, waiting to see what they think they just saw. They twirl their fingers; they throw the red flag; they cover their ears; they tap their helmets. The officials convene. “The previous play is under further review.” The officials convene to review if the catcher tapped his helmet to call for a review.

Was his toe on the line? Was his toe down? Was his toe off the plate? The broadcast shows us a replay; we see fans looking at replays; we see fans looking at their phones. We look at our phones. We welcome NFL rules analyst Mike Pereira into the broadcast booth. “Mike, what do you see here?” And then, eventually, in halting rulebook jargon, through a crackling lavalier microphone, we are told what we saw.

Of course, not every call is reviewed. Not every catch is disputed; not every home run is in doubt. But the possibility of a review is enough. After every big return—are there flags on the field? After every bang-bang play, including one to decide the World Series—“they might take a look just to take a look.” After every buzzer beater, including one to decide the NCAA basketball championship—“Jim, they’re gonna check it.”

That is the tradeoff in modern sports. The calls are correct, but the drama is clumsy, revealed to us in fitful piecemeal, like a poorly kept surprise party. And you can’t just watch sports anymore; you have to participate in officiating them too.

Live long enough on the internet—which is increasingly synonymous with just living—and you have to become your own Mike Pereira. I subscribe to a popular daily newsletter on technology news, and I noticed a consistent similarity—in topic and tone—between its emails and an AI-generated one sent by a SaaS product. Have I been reading the thoughts of the author, or the tightly-bulleted ramblings of lazily prompted bot?

Jeff Bezos’ rocket blew up yesterday. Here’s a picture—is it real and an embellishment?

Can peacocks fly? Does this video prove it or not?

Can a tree grow in the white sands of the New Mexico desert? Does this picture tell you that it can?

Here is Magnifica Humanitas, a 42,000-word encyclical about artificial intelligence, written by Pope Leo XIV. Or was it written by AI?

Slop is everywhere now. To keep our humanity, we have to be watchful. AI is a valuable tool, but we must be vigilant.

But at what cost? It is not so easy to separate the internet from the rest of our lives. Once on, we cannot turn that vigilance off. Live in the metaverse long enough, and you disembody yourself; your animated hands become as real as your physical ones. Stream enough sports on TV, and you catch a similar feeling. Even when you go to a game live, in the momentary reflex to check your phone, you hesitate—what if what I’m watching is lagged, and the score alerts on my phone are ahead of the “stream” in front of me?

The internet has always dulled us. Our morning are spent blitzing through feeds of tragic headlines and awful pictures and thirst traps and rage bait and Wikipedia rabbit holes about the natural wonders of the world. Surely, that multimedia mutes us, in the rare moments that we encounter the same thing in the real world.

Is the same true for doubt? If our afternoons become skeptically parsing through emails and LinkedIn posts, wondering which are “real” and which are partial forgeries, will we ever let ourselves get lost in what we read? Will our first response always be to question if something is real—and only then, after Googling and counting fingers, a sober sense of reluctant awe? Will we become doubting Thomases who doubt too much, unable to believe anything unless we see it ourselves?

But even seeing may not be enough. Through hazard and happenstance, I spent this week in New Mexico, where the altitude lights the sky on fire.

You see it, clearly, blazing all around you, but you still feel that—that subtle disembodiment; that twinge; that instinctive hesitation. That is the cost of pervasive skepticism and the scientific method: you do not choose when it creeps into your consciousness. Challenges aren’t allowed on some plays and not others; they are a different set of rules, and a different way to play the game. Once trained on doubt, a mind doesn’t so easily give in to enchantment.

I do not know how to make sense of what’s happening now, or where it will all go. The lines between what is “real” and what isn’t will blur; we will argue over them; we will debate if what’s real is fake, if what’s fake is bad, or if it matters at all. Most people will wisely encourage us to be responsible about what we trust. Slow down; measure twice; confirm the call. You cannot run through a house of mirrors.

But amid the reflections and hallucinations, you still sometimes see exactly what you thought you saw. The real skies are as pink as they’ve always been; the sun still rises out of airplanes’ windows; there is still something in that orange light, just as there was before.

And for us to see that—to live that—we have to be willing to occasionally get the call wrong. We have to be reckless enough to let go, even when you do not know what is beneath you, and surrender to the air. To keep our humanity, we cannot be too vigilant. Sometimes, you could not write a script like this, because the world still writes its own scripts, too.

A question that people often think about these days is, “how good is AI at doing stuff?” A lot of stuff that people do—like building software, and responding to customer support tickets, and drafting legal documents, and analyzing data—is either stuff that they would rather not do, or stuff that their employer would rather not pay them to do. Large language models seem capable of doing that stuff. So, for better or for worse, whenever a new AI model comes out—which, in 2026, happens every roughly every other day1—people want to know: Can the new model do the stuff?

Over the last few years, a popular method for answering that question has been to administer a test. Create a long list of questions, about software or support tickets or charts or the law; give the questions to the model; see if it gets the right answers. Can it write code? Can it follow instructions? Demonstrate empathy? See pictures? Do math? Do math when the problems are pictures? Pass a Mensa test? Use a computer? Use the internet? Pass the LSAT? Pass the bar? Resist lying after it passes the LSAT and the bar? Create a test and find out.

Of course, everyone knows these tests aren’t perfect, but what else can you do? And surely, these tests are better than nothing.

A question that people think about a little less these days is, “how good is a person at doing stuff?” It used to be that if a business needed to do stuff—if they needed to debug software or respond to customer support tickets or draft legal documents or make charts—they wouldn’t go shopping for a computer agent; they would create a “job”—that is, a collection of tasks to be done inside of a particular company—and hire someone to do it. And so, people wanted to know: Which people can do those tasks, in that company?

For a while, one popular method for answering that question was to administer a test. The company would invite candidates to their office, ask them to solve logic puzzles or do math or play chess, and see if they got the right answers.

Part of the theory behind these interviews was the same as theory behind any academic test: Sure, the test isn’t the job, just as the MCAT isn’t an operating table, but performance on the test correlates with performance at the job. People who are good at solving coding puzzles on a whiteboard might be good at debugging production software; people who are good at math puzzles might be good at making useful business charts; people who are good at abstract games of logic might be good at finding market, uh, inefficiencies.

But a bigger part of the theory was, “what else can you do?” Companies only get so much time with job candidates. Nobody would’ve argued that the tests were perfect; some people would’ve argued that the tests are better than nothing.

Fewer people would argue that now.2 From The Atlantic, thirteen years ago:

“We found that brainteasers are a complete waste of time,” Laszlo Bock, senior vice president of people operations at Google, told the New York Times. “They don’t predict anything. They serve primarily to make the interviewer feel smart.”

And new methods became trendy instead. From Linear, a few years ago:

We believe the only way to build a quality product and business is to hire people we can trust to make good judgments, across all functions and levels. …

The problem is that these types of people, unfortunately, are few and far between. The majority of companies don’t work in the way we do, which leads to fewer people with these kinds of skills. We found that standard interview processes didn’t work well for us. It’s challenging to assess in interviews if someone is truly a builder, has good taste and judgment, can take initiative, and approaches problems productively. … A conventional interview process, often modeled by large companies, doesn’t account for this.

To evaluate if a person is a fit for Linear with the skills to be successful, we bring candidates in for a work trial as the final step in the interview process. A work trial is a paid 2-5 day period where a candidate works with our team on a real project that we plan to implement (or as close as possible to that) with access to relevant internal tools and resources.

In other words, despite many years of test development, a job still cannot be contained in a test. Not only does a job take “good taste and judgment,” but employees also have “relevant internal tools and resources” that test-takers do not. A person who is very familiar with those tools or good at asking questions of those resources, for example, might be great at the job and terrible at the test.

Moreover, a job evaluation cannot be contained in an afternoon. Companies have been interviewing people for jobs for decades, at great expense. And currently, the state-of-the-art method of evaluation is to work with them for a while, in the actual office, on the actual job. When every business has its own quirks—its own culture; its own processes; its own problems; its own coworkers—the only way to test if someone is good at the job, it seems, is for them to do the job, to encounter those quirks, and to see how it all goes.

So, you know. Could AI be bound to the same constraints? At first, when large language models were primarily used inside stateless call-and-response chatbots, you could fill a test with representative examples of what people wanted the chatbot to do. But now, AI is an agent—it is a persistent system that talks to itself, it has memories from prior conversations, and it has access to internal tools and resources. That’s not an employee, but it does a lot of things that employees do. And how do you benchmark that?

A question that a handful of people think about very intensely is, “how good are baseball players at playing baseball?” Millions of people play baseball, and there is a very lucrative market for the few hundred people who are the very best at it, so there is also a very lucrative market for the people who are the very best at appraising baseball players. How do those people decide who is good?

The original answer was basic statistical benchmarks. A pitcher that can get 27 outs before giving up three runs is very good. A hitter that gets three hits in every ten at-bats is very good, as is a batter that can drive in 100 runs in a season, or hit 500 home runs in a career. Baseball is full of bold lines around round numbers.

Over time, people realized that these numbers are too crude, and came up with more advanced statistics. Don’t just count hits; count the number of times people get on base, weighted by which base they reach. Don’t count wins; count baserunners allowed per inning pitched.

But these numbers still had problems, because they were absolutes. Players aren’t evaluated against some arbitrary fixed standard; they’re evaluated against their contemporaries. So the baseball appraisers came up with “wins above replacement,” or WAR, which aggregates everything a player does on a baseball field in an attempt to define how much better (or worse) the player is than the average player that a team could call up from the minor leagues.3 And:

Over the last two decades, the Wins Above Replacement metric (WAR), which combines the contributions of different performance elements, weighted according to their contributions to team wins and adjusted for the environment in which a player’s statistics were accrued, has become the metric of choice for global evaluation of player performance. WAR is superior to traditional counting stats like [home runs, runs batted in, wins, and strikeouts] because it incorporates many diverse elements of performance that contribute to team wins and adjusts for the environment (league, ballpark, era) in which a given player performed.4

Though the exact formula for WAR is complicated—and there are a number of different variants—the idea behind it is simple: How much better is this player than a generic one that would take no effort to identify and relatively little money to pay? It’s a metric that doesn’t measure specific performance; it measures how far something is above an easy baseline.

A question that nearly everyone now thinks about is, “if I YOLO a bunch of MCPs into Claude, will it be able to do my job for me?” For every task, that is the new global competitor: Claude Code (or Codex, or Cowork, or ChatGPT, or Copilot, or Cursor, or Clawdbot, or, I guess, the generalized form, C*), connected to everything. Want to build a better coding robot? It’s your product versus Claude Code integrated with Linear. Want to build a better analytical agent? It’s your product versus Codex connected to Slack, a dbt MCP, and Databricks. Want to build a better automated personal assistant? It’s your product versus Clawdbot, running wild on a Mac Mini. Want to be a product marketer? It’s your launch blog post versus one written by Cowork, connected to Google Drive.

When we built a new analytical benchmark last year, that was the lesson we quickly learned: Nobody cared about the scores or about which models performed the best against some arbitrary set of tasks. They cared about how well different products performed against some system that they’d hacked together over the last two weeks, and had integrated with their relevant internal tools and resources. They cared about the new tool, against the easy—and increasingly default—baseline.

It points to a fundamental tension that now exists in nearly every benchmark: How do you create a standardized benchmark when context—the idiosyncratic stuff that is, almost by definition, impossible to standardize—is the critical thing that makes one tool useful and another tool unusable?

But maybe that is a question that we can all stop thinking about. We don’t benchmark employees, nor do we benchmark traditional software. Perhaps AI software will soon be the same. And instead, the next question that every product—and maybe every employee?—will have to answer, is, how much are you worth above C*?

CSS

It has become a truism to say that it’s not the model that matters; it’s the harness that matters.

But what is a harness? The normal answer is that it’s a complex set of logical instructions that turn a user’s request into a series of recursive, self-authoring prompts. We speak hyperbolically of harnesses now; harnesses are on the frontier of computer engineering; harnesses are the platform; harnesses are AGI.

But perhaps that is overcomplicating things. Could a harness be a web interface that encourages its users to write better prompts? Could it be a text box with useful suggestions for what to type? Could it just be a text that is tall enough to type in?

We have joked about this before:

If you’re Google, here’s another idea, if you want a better and cheaper wrapper around Gemini: Make the Google search box bigger.

People already like Google’s AI Mode more than ChatGPT! But one-line boxes are for search! Two-line boxes are for chat! Even if you can AI Mode from google.com, nobody’s gonna chat in a one-line box! So just make it a two-line box! That’s the thoughtfully designed interface that make Gemini accessible to everyone. Google doesn’t need a $2 billion acquisition. Google just needs some new CSS.

Or was it a joke?!

Google changes its search box for the first time in 25 years

For 25 years, Google’s iconic search box was a long, slender bar where people typed in keywords like “World Cup.”

But over the past three years, artificial intelligence allowed people to type in longer, more complex questions like “Who are the top 24 teams in the World Cup and what chance does the United States have of advancing?”

On Tuesday, Google said the A.I. shift had inspired it to overhaul the dimensions of its search bar for the first time since 2001. The box is getting bigger and more interactive so that people can ask even longer questions and upload photographs and videos into queries.

Your move, OpenAI.

Am I working?

Am I working, or am I wasting time? Working, surely. Because it’s what people do for work. Because if it works, there is money to be made. There is a business to be built. Every piece of software began with a first commit. Every startup was once a rough prototype. Every unicorn took its first steps. Every good idea was discovered, amid 10,000 bad ones. Am I working? Yes, because, if nothing else, I am learning.

Am I working? Well, it’s not working. It worked at first—one prompt in; twenty minutes into the machine tearing through an empty folder; 8,000 lines added, 0 lines removed. It worked then. Check it out, the machine said; localhost:3000; see what we have built together. The snap of a few sparks; a quick whiff of ozone; and it blazed forth, with electric possibility. Since, though, the blistering pace has slowed. Confusion; exceptions; obstacles; previously unconsidered complications. Let’s plan this; I say; it plans it, in eight phases. I read the first one carefully. Do the first phase. Do the next one. Do the next one. It keeps writing plans; there is so much text; there are so many words. Do the next one. Do the rest of them. I’ll test it when you’re done. I’ll fix it when you’re done. Is it working? Not anymore.

What is it, even? We’re losing track; clarity was compacted out of the codebase long ago. We started with a vision; now we are dead reckoning from one feature to the next, the vision evolving with every prompt. The current app is a blurry JPEG of its original ambition.

But that’s ok. This is an exploration; this is a first draft. It is not a final product; it is a starter codebase. It is a clickable spec. It is future lore. All this scratch work—these bad ideas; this tech debt; these evolutionary vestiges; all of my stairs to nowhere—they can be fixed later. I’ll redo it with a better model. I’ll rebuild it with a better plan. I’ll rebuild it with Opus 4.8; with GPT 5.6; with Opus 4.8 writing the code and with GPT 5.6 reviewing it. I’ll rebuild it in Rust.

The empty folder is filling up. There is a second folder; a third. This is the Claude Code version; this is the Codex version; this is the Gemin–eh, no never mind. This is the one that had a nice UI; this is the one that was a Swift app. Create a parent folder; $ cd ..; $ claude; $ there are two apps in this folder, I want you to combine.... There were 14 versions; now there are 15 versions.

It’s dinner time, but the machine is done combobulating. It needs more instructions. The Red Queen is not running, and she must run. We must review her output—no, not quite, it’s too far to the left; too far to the right; try it to the left again; try hiding the menu. Maybe it should be its own page. Maybe it should be its own app. Maybe it should be a different company.

Pull the lever on the slot machine; ask for one more treat. Ask for an update to the home page; for a new Gmail integration; for dark mode. Pull it again—ask for new ideas. Pull it again—tell it that the app doesn’t make sense anymore. You’re not getting it. I need you to think harder. I need you to ultrathink. I NEED YOU TO ULTRATHINK. Pull it again—ask it what makes a meaningful life. That’s exactly the question you should be asking, it says; do you want me to dream for you?

The dream was a stable job and a steady pension. It was then a big tech company, with free lunches and unlimited PTO; it was an exploding startup, fat with VC money; it was a lean startup, free from the shackles of VC money. It is now a software studio; a founder and army of agents. It is a self-sufficient Kalshi bot, running on a Mac mini. It is Jane Street, vibe coded. It is a billion dollars, and zero employees. Am I working? Yes, on someone else’s dream.

What is work now, anyway? There are only four jobs left; there are only three jobs left. There is one job left: the Entrepreneur. The “builder,” with dirty hands.

Am I “building,” or do I just have a hobby? A midlife crisis? An addiction? Was Animal Crossing building, or playing a game? Is day trading for making money, or for gambling? Nancy Lemann, via Ava: “Some take to the bottle, some look to the sky, but it is all the same. Byzantine empires, lives of the saints? Mr. Collier took to his books, and got overintellectual.” He took to his computer, convinced himself of the profundity of his ideas and the urgency of their existence, and forgot to ask if the world really needed to hear so much from him. “Build me an app that manages all of my notes,” he told the machine. “Manage all of my notes,” he should have said.

But what fun is that? What ambition is that? What job is that?

Sally Rooney, via Camille Charrière: “All my mania for culture”—and so much of our mania for building?—“for ‘really good’ things, for knowing about jazz recordings and red wine and Danish furniture, even about Keats and Shakespeare and James Baldwin, all what if it’s all a form of vanity?”

In 2020, a friend used to joke: In two years, thousands of people are going to wake up in some distant mountain house, and be grateful that they were jarred free from a game that they never intended to play. And thousands more will wake up wondering how on earth they ended up in Montana.

This epidemic, too, will pass. No magic remains mesmerizing forever. The novelty will wear off. The subsidies will disappear. Our ideas will hit the wall; we will build them, and nobody will come. The conjuring machine may be the future, but the thrill of it will fade, its crackling sense of possibility will dull, and “building” will lose its vain luster. And when it does, will we be grateful for the places that this delirious moment took us? Or will we wonder where in the world we thought we were going?

If you are a bold entrepreneur who wants to change the world with ambitious artificial intelligence technologies, how should you do that? Two answers seem obvious:

1. Build ambitious artificial intelligence technologies.

2. Talk about building ambitious artificial intelligence technologies.

Arguably, the second thing is just as important as the first thing.1 You are selling sweeping visions of prosperity and productivity as much as you are selling technology. Leading artificial intelligence researchers—the people you need to build ambitious artificial intelligence technologies—want to work at big companies with lofty missions that can make them rich. Customers want to buy stuff that will help them survive the looming mass obsolescence event. Venture capitalists want to fund companies “with more ambition than ever.” One AI startup periodically publishes essays about how their ambitions are somewhere between destroying the world and saving it; they are about to raise $50 billion.2 Another AI startup told investors that they should consider their investment a donation because “it may be difficult to know what role money will play” in the world if they are successful; it was the greatest fundraising pitch in history. To change the world with ambitious artificial intelligence technologies, do not talk quaintly about “making something people want;” talk about an age of 100x productivity; talk about recalibrating your gradient; talk about ushering in new eras of maximal abundance.

Before you can do any of that, though, you will have to first talk to the government about building ambitious artificial intelligence technologies. To incorporate a company, you have to pay various fees and fill out various forms, and some of those forms will ask you about your company. As a matter of accuracy and as a matter of ego, you might want to memorialize your bold ambition into your company’s founding charter. You will click the dropdown—“What does your business or organization do?”—and look for the appropriately audacious selection: Revolutionize labor; or, Launch intergalactic data centers; or, the Jetsons. You would probably be happy with Artificial Intelligence. You might even settle for Software.

But, no. Those will not be options in the dropdown. The IRS bureaucrats are unfazed about your ambitions. The BLS econometricians do not recalibrate their gradients. They do care if you have already raised $4 billion, if you have so many Porches that you get stuck in your traffic, or if you live in a moon house floating over San Francisco like the mothership. They care about the consistency of their statistics. And so, when you fill out their forms, you will have to declare, officially, to the United States registrar of permanent records, that your business does Information Technology.

It will feel like a terrific indignity. IT is the opposite of ambition; IT is anathema to ambition. IT is administration. IT is paranoia. IT is a hall monitor, blocking websites. IT is your mandatory anti-phishing training; IT is the disabled skip button on the video player playing your mandatory anti-phishing training. IT is your flaky VPN. It is a PC. It is Microsoft Teams. It is a stiff suit and a lanyard; it is khakis, a polo with “IBM Medallion Partner” on the sleeve, an OGIO Renegade RSS backpack, with both straps on. IT is the brakes.

Bold entrepreneurs and ambitious artificial intelligence startups are all gas. They do not follow the rules. They experiment; they dangerously skip permissions; they build with unhinged and explosive abandon. They run swarms of phishing software directly inside their VPNs. IT is supervising Jira. AI is replacing Jira; AI is agentifying Jira; it may be difficult to know what role Jira will play in the world AI is building.

Information technology, pfft. You are building a cathedral that will change the world, and all the government can see is the bricks.

And, I mean, just look at how much AI is changing the world. It is changing Block:3

today we’re making one of the hardest decisions in the history of our company: we’re reducing our organization by nearly half, from over 10,000 people to just under 6,000. ... we’re not making this decision because we’re in trouble. our business is strong. gross profit continues to grow, we continue to serve more and more customers, and profitability is improving.

but something has changed. we’re already seeing that the intelligence tools we’re creating and using, paired with smaller and flatter teams, are enabling a new way of working which fundamentally changes what it means to build and run a company. and that’s accelerating rapidly.

And earlier this week, Coinbase:

Today I’ve made the difficult decision to reduce the size of Coinbase by ~14%. … AI is changing how we work. Over the past year, I’ve watched engineers use AI to ship in days what used to take a team weeks. Non-technical teams are now shipping production code and many of our workflows are being automated. The pace of what’s possible with a small, focused team has changed dramatically, and it’s accelerating every day.

And, just yesterday, Cloudflare:

We’ve made the decision to reduce Cloudflare’s workforce by more than 1,100 employees globally.

The way we work at Cloudflare has fundamentally changed. We don’t just build and sell AI tools and platforms. We are our own most demanding customer. Cloudflare’s usage of AI has increased by more than 600% in the last three months alone. Employees across the company from engineering to HR to finance to marketing run thousands of AI agent sessions each day to get their work done.

Of course, there are questions. Is AI changing these companies, or are they using it as an excuse to correct past mistakes? Have they oversold themselves on AI’s hype? Many people are saying this.

Still, regardless of the motivation,4 the companies are changing:

Cloudflare – [We are] defining how a world-class, high-growth company operates and creates value in the agentic AI era.

Coinbase – We’re reshaping Coinbase to lead in this new era.

Block – We’re going to build this company with intelligence at the core of everything we do.

Coinbase – We are adjusting early and deliberately to rebuild Coinbase to be lean, fast, and AI-native.

Cloudflare – We are reimagining every internal process, team, and role across the company.

Block – Block is showing what it looks like to fundamentally rethink organization design.

Cloudflare – [Cloudflare] cannot rest on the workflows and organizational structures that worked yesterday. We’re confident that our reshaped organization will be even faster and more innovative as we continue building the future.

Block – The intelligence lives in the system. The people are on the edge.

Coinbase – [We are] rebuilding Coinbase as an intelligence, with humans around the edge aligning it.

Block – Player-coaches [will] replace the traditional manager whose primary job was information routing. A player-coach still writes code or builds models or designs interfaces.

Coinbase – Every leader at Coinbase must also be a strong and active individual contributor. Managers should be like player-coaches, getting their hands dirty alongside their teams.

Block – Companies move fast or slow based on information flow. Hierarchy and middle management impede information flow. … Block is building what comes next.

One way to read all of this is that AI is reinventing Block, Coinbase, and Cloudflare. But the other, more literal way to see it is that Block, Coinbase, and Cloudflare are reinventing Block, Coinbase, and Cloudflare. They are defining how their companies will work; they are rebuilding their companies with intelligence at the core;5 they are rethinking organizational design; they are building what comes next.

Decades ago, companies like Microsoft changed the world with Windows and the personal computer. But they did not author what happened after. That was figured out on the ground, across a million IT teams, testing new ideas against their specific problems, rebuilding their companies around the computer and the internet. It was in partnership with Microsoft, perhaps; as Medallion partners, perhaps; but the final cathedral of the modern corporation was designed by Jeff in IT as much as it was by Silicon Valley’s storied founders.

The same, it seems, will be true again. Box CEO Aaron Levie:

Whether it’s existing consulting firms, new ones that emerge, FDEs from agent vendors, or new internal agent engineering roles, the amount of work that is going to be created to implement agents in enterprises will exceed anything we imagine today.

AI companies cannot do that work. They can help with it—they can spend billions building consultancies and deployment companies—but it too will ultimately be carried out on the ground, across a million “internal agent engineering” teams, testing new ideas against their specific problems, rebuilding their companies around “intelligence.”

It is a frequent question now: What will work be like? If we are revolutionizing labor, what will labor become? AI companies are building the new technology, and it is easy to assume that they will be the ones who tell us the answer.

But the long tail of teams experimenting with AI will be much bigger than the AI companies themselves. OpenAI and Anthropic may build the model and some of its generic products; everyone else will figure out how to use them. And it will be their discoveries that become tomorrow’s standards.

Or, framed differently, if you are a bold entrepreneur who wants to change the world with ambitious artificial intelligence technologies, are you building the cathedral, or are they? Who is the real architect of the future—who is the most trailblazing pioneer?—the AI company, or the internal teams that are thinking about, uh, business technology and the internal flow of information?

XaiXspaceXanthropic

Last week, we talked about AI companies, their computers, and the cyclical race they’re all running:

1. To build frontier AI models, you need a lot of computers.

2. Frontier AI models are so popular that even big companies like OpenAI and Anthropic do not have enough computers.

3. So the companies have to choose: Do they use their computers to train new models, or do they use them to sell those new models, through applications like ChatGPT and Claude Code?

4. In some sense, you have to do both—you can’t sell a bad model, and a great model is no good if you don’t sell it.

5. So, you devise a two-step plan: First, focus on building the best model. Then, once it’s the best, you sell the model.

6. But! Once you build the best model, everyone will want to buy it—more people than you have capacity to serve.

7. What do you do??

8. You probably use the computers to sell your AI. Money has a way like that.

9. When you do that, though, you open the door for your competitors, who will simultaneously be selling less of their AI and in need of better one. So, they will probably focus on building the best model…

10. Etc., etc., etc.

However, in the original post, there were eleven steps. Between steps 5 and 6, there was a joke:

Sometimes, [building the best model] doesn’t work. You spend a zillion dollars, and do not build the best AI. Ah, awkward, an A for effort, I guess.

So what do you do then? One option is to spend a zillion more dollars and try again. But the other, easier option is to give up and sell all of your computers:

Anthropic on Wednesday announced a deal with Elon Musk’s SpaceX to use all of the compute capacity at his company’s Colossus 1 data center in Memphis, Tennessee.

As part of the agreement, Anthropic will get access to more than 300 megawatts of compute capacity, and it also “expressed interest” in working with SpaceX to develop multiple gigawatts of compute capacity in space. Anthropic said the deal will directly improve capacity for its paid Claude Pro and Claude Max subscribers.

Right, if you are in a bike race, and you are close behind the leader, it’s not that hard to stay in the race. From inside the peloton’s slipstream, you are one sprint from being back in front. But if you’re in 29th place, you can’t draft behind the leaders, nor are you likely to catch up to them. In which case, maybe biking just isn’t your thing.6

From a conversation with a smart friend,1 a fun theory:

1. To do AI, you need a bunch of very fancy computers. However, because lots of people want to do lots of AI, there is a shortage of these computers. In a few years, we might have enough of them—we are putting data centers in Texas and in Tennessee; we are putting data centers in the ocean; we are putting data centers in orbit, on the moon, we might put data centers on Mars; can we borrow your computer to use it as a data center?—but, until then, we are short. Not everyone who wants a computer can have one.

2. This is even true—especially true?—at a handful of huge companies, like OpenAI, Google, Anthropic, Meta, and SpaceX-X-xAI.2 Though they have a lot of fancy computers, they also want to do the most AI, and still have to make hard choices about how to use their computers.

3. Specifically, they have to choose if they want to use their computers to build AI or sell AI. Training new frontier models is how you build the smartest AI, and customers want the smartest AI. But offering people a way to use that model—that is, using the model you trained to provide inference, via websites like ChatGPT or programmatic APIs—is how you make money. The model is your secret recipe; inference is the food on the plate. Customers don’t want to eat bad food; customers can’t eat a recipe. You have to invent a good recipe, and use it to cook the food.

4. But, the computers. You do not have enough computers to do both.

5. So, you make a plan: “First, we’ll use most of our computers to build the smartest AI. It will cost a zillion dollars, but that’s ok, because once we have the best AI, we will aim to achieve turnaround.” Sow today, reap tomorrow, a classic business move.

6. Sometimes, it doesn’t work. You spend a zillion dollars, and do not build the best AI. Ah, awkward, an A for effort, I guess.

7. Other times, it does work! You spend a zillion dollars, and build the best AI. Everyone loves your AI. They love the products that use your AI. They want to stop using that company’s AI and use this company’s AI. They want to run your AI everywhere, all the time. This is great; this is exactly what you build the best AI for.

8. But, the computers! There is so much demand for your AI, and you only have so many computers. So you have a choice: Do you capture the moment, and use some of your training computers as inference computers? Or do you let it slip, and not sell the thing that you just spent a zillion dollars making?

9. I mean. You probably sell the AI.

10. Now, though, you have fewer computers to train an even better AI. And other companies—which no longer have the best AI—might get to thinking: “The company with the best AI is using a lot more computers selling AI. What if, while they’re doing that, we use most of our computers to build an even smarter AI? It will cost a zillion dollars, but that’s ok, because once we have the best AI…”

11. Round in round, round after round.

Yes, sure, this cuts a lot of corners. Good AI isn’t just built by computers; it’s also built by good researchers, unique training data, and, increasingly, the product around the AI. AI companies can do many things at once; their computers are not so easily fungible; many of their computers are actually, famously, also their competitors’ computers, leased back and forth in a giant trillion-dollar web of incomprehensible partnerships and frenemy-ships. There is no knob that says “build” on one side and “sell” on the other.

Still, that is the pattern, isn’t it? In 2023 and 2024, OpenAI was the dominant default, Anthropic was the IYKYK upstart, and Google was incinerating its reputation, and many billions of dollars, one disastrous demo at a time. Then, Google built the best new models—“Gemini 2.5 is the new SoTA;” “I tested Google’s new Nano Banana image AI, and it’s insane”—and became the smart money. Then, Anthropic put out Opus 4.5; it was a “world-changing shift;” everyone became obsessed with Claude Code; “move over, ChatGPT.” Anthropic blew past OpenAI’s hype, its revenue, and its valuation; OpenAI started missing revenue targets. But revenue is a lagging indicator. Over the last six months, while Anthropic was selling Opus and Claude Code—and struggling to find enough computers to do so—OpenAI was building GPT 5.5, out last week. And GPT 5.5 has it all; GPT 5.5 blew people away; GPT 5.5 is a new frontier in everything; cracked devs on X are abandoning Claude Code and coding on Codex.

And OpenAI is an inference company now. Build phenomenal cosmic power; get everything that goes with it. Round and round, round after round.

The implication, I suppose, is that this old post was partially right: AI labs are bad businesses. Models really are fast-evaporating moats, clever employees really are two weeks from leaving, and every billion-dollar model is one competitive release away from being leapfrogged into near-immediate obsolescence. But, this comparison was wrong:

There is, however, one enormous difference [between AI labs and cloud providers like AWS]: You can’t build a cloud vendor overnight. Azure doesn’t have to worry about a few executives leaving and building a worldwide network of data centers in 18 months. AWS is an internet business, but it dug its competitive moat in the physical world. …

What, then, is an LLM vendor’s moat? Brand? Inertia? A better set of applications built on top of their core models? An ever-growing bonfire of cash that keeps its models a nose ahead of a hundred competitors?

The problem is the premise. AI companies are actually no different than cloud providers: They build models, but they sell computers. And their lasting moat—their only lasting moat—is their computers. Models converge; talent migrates. Computers endure.

The other implication—the more useful one, maybe—is that AI whiplash is structural. The vertigo is systemic. The hype cycle is scripted theater. In science fiction stories, the AI company that builds the best model has a compounding advantage over its competitors. Someone builds the smartest model; that model helps its developers improve itself faster than the second-smartest model; a small gap becomes a big gap; a big gap becomes an insurmountable gap. There is urgency in small differences, because you never know when pulling ahead could explode into pulling away.

Perhaps that is backwards. Perhaps the AI race is less of a space race and more of a bike race: The headwinds are borne by whoever is in the front. Perhaps it’s Mario Kart: There are hardcoded advantages to being behind. Perhaps there is an invisible asymptote to every AI company’s lead: The further ahead your AI gets, the more people want to buy it, and the less new AI you can build. Perhaps it’s not a race at all, but a ___ run.3

Spectating from the dizzying center of the arena, maybe there is something steadying in that. When the peloton moves together, does it matter who’s winning right now? When no lead is safe, should we bother worrying about the leaderboard?

How do you make a chart, in the age of artificial intelligence? You have many choices:

1. Give an Excel file to an image model like Nano Banana or ChatGPT Images, and tell it to make a picture. Have it read your data, squint at some imagined piece of paper in its head, and draw a bar chart, of the Charlotte Hornets 2025-26 season, in a nostalgic 1990s aesthetic.1

2. Make your chart directly in Excel. Use AI in Excel™, powered by Microsoft 365™ Copilot™. Or, other models can use computers now; tell them to take over your computer, open Excel, click “New chart,” and flip through all of Excel’s menus and dropdowns.2

3. Ask an AI to use a simple charting library to create the chart with code. There are so many: Matplotlib, Plotly, Seaborn, ggplot2, Bokeh, Highcharts, Observable, Vega, Vega-Lite, Apache ECharts, Google Charts, Fusioncharts, ZingChart, Chart.js, I could do this forever. The robots will create your chart with a few simple configurations, like defining which fields go on which axis, and let the charting library handle the exact details.3

4. Ask it to make the chart using a low-level charting framework like d3, “the JavaScript library for bespoke data visualization.” In other charting libraries, there are concepts like “bar charts” and “legends.” In d3, there are concepts like “rectangles” and “text.” It’s a collection of visual Lego bricks: Most of the bricks are shaped like things that might be useful for making charts, but you can make anything with the bricks.4

5. Make the chart using native Javascript. Javascript is the language that people use to create any website—YouTube, Wordle, the homepage for the original Space Jam. If Highcharts lets you draw charts, and d3 lets you draw shapes, Javascript lets you draw anything.5

6. Make the chart in C. C is a general-purpose programming language for creating programs and operating systems. If AI can do anything, why should it use Excel, a browser, or even Windows? That is generic software. It should create a program designed for making exactly the right chart. It should design an operating system designed for making exactly the right chart.6

7. Make the chart in a bespoke programming language. Design the language for exactly the chart you want. Is the language Turing complete? It does not matter. It only needs to be “a bar chart, of the Charlotte Hornets 2025-26 season, in a nostalgic 1990s aesthetic” complete.7

8. Grow corn. Grow corn to feed the workers, who build the mine, which digs up the silicon, which goes into the semiconductor fabrication plant, which manufactures the chips, which assembles the computers, that create the operating system, that hosts the program, that draws the chart. Rebuild the entire global economy, literally from the ground up, to make the perfect chart.8

One way to think about this question is that it is dumb. Nobody needs to start a farm to make one chart. Nobody needs to even build a custom app to make a chart. There are a million charting programs out there already; just pick one, and tell a robot to make the chart. Use Excel; it will be there forever.

Another way to think about this question, though, is that all of our jobs and money depend on its answer. A lot of the world—and nearly all of Silicon Valley—exists somewhere between “a polished software application like Excel making a complex task possible in a few clicks” and “grow corn.” There are conglomerates and research labs that design and maintain foundational programming languages; there are huge corporations like Apple and Microsoft that build operating systems that can run software programs; there are thousands of companies that build technical infrastructure like databases that runs in those operating systems; there are even more companies that build consumer products like Excel that make those technical applications accessible to the masses. Everything that is done on a computer passes through an entire supply chain of software, a supply chain that supports thousands of businesses and millions of jobs.

But, you know, AI. The faster and more capable the models get, the theory goes, the less of the supply chain needs to exist. AI can work directly with technical infrastructure; it can build its own software; it can generate charts. And the question—about charts, sure, but about every software product and software company—is, where is the line? How far down the list can they go?

Here’s one answer: MotherDuck is a cloud data warehouse.9 It is a technical application that is typically one step removed from charts, and from the business people who want to look at charts. Engineers set up MotherDuck and fill it with data; they connect it to a tool like Excel, or to a drag-and-drop business intelligence software like Omni; business people use Excel or Omni or whatever to make charts of the data in MotherDuck.

But, you know, again, AI. So, two months ago, MotherDuck released Dives:

Dives are interactive visualizations you create with natural language, directly on top of your data in MotherDuck. Ask a question to your AI agent, and MotherDuck generates a persistent, interactive component that lives in your workspace alongside your SQL.

How does Dives build visualizations? With robots that write JavaScript, the programming language that can create any website:

[Instead of clicking through a UI or writing visualization code,] just describe what you need. Interactive visualizations from a single prompt, built by your AI agents, living in your MotherDuck workspace. And the beauty of it: Dives is Javascript code. You can build anything you want, with any interactivity you’d like.

It’s the future of analytics, some say. There is no more visualization software. That’s what gets replaced. Don’t buy Tableau; instead, open Claude Code or Codex, and ask it to make a custom dashboard in JavaScript or Python. And that’s better than using dedicated charting tool, because with JavaScript or Python, you can build anything you want:

Fabi’s Python dashboards are faster to build (minutes vs weeks), more flexible for advanced analysis and custom visualizations, and fully transparent so you can see and edit the underlying code.

The idea has even leaked into the LLMs themselves. In a later post, MotherDuck asked Claude what would happen to BI tools as AI continued to advance, and it said the same thing:

LLMs already draw better charts than Tableau from a simple prompt. … The “drag-and-drop dashboard” becomes a curiosity, like a fax machine with a particularly nice interface.

Maybe; I don’t know. If you’ve ever tried to create a chart from scratch, or even with a visualization library, you know the Rubik’s’ cube of frustrations: Overlapping axis labels; misplaced legends; colors that collide with one another; you move one thing to fix a problem over here and it creates a new problem over there. Software isn’t just a convenient shorthand for doing something; it solves the very hard problems that you don’t realize exist until you have them. In that telling, while the entire software industry might pivot toward building user interfaces that LLMs can work with, software itself would remain as useful as ever, and the supply chain would remain intact.

Or, maybe LLMs really do get good enough to work through those edge cases and annoyances. But of course, if they can do that—why would they stop with making visualizations? If they can work around all of the headaches that Tableau has spent twenty years solving, could they not also work around the headaches that make it hard to host those visualizations? Could they not build a better database? Sure, the latter two seem harder, but how many things did we once say were too hard for an LLM, and soon realize were not?

There are obvious reasons why an LLM probably shouldn’t grow corn. Nor, probably should it try to rebuild every piece of software. But how do you draw the line?

Big vibes

Or, maybe the answer is to not make a chart at all?

…the future of business intelligence isn’t a better dashboard.

A support leader at one of Omni’s customers went through 75 pages of conversation with an AI to identify 10 categories of rep mistakes. The system read support logs, cited specific examples per rep, & suggested concrete changes. Not just a dashboard.

We’ve talked about this a number of times before: The real future of analytics might not be analytics, but vibes:

If AI is good at anything, it is good at interpreting the vibes. It is good at aggregating massive amounts of text—and increasingly, of video and audio—into its approximate average. Give it your support tickets and customer communications, and ask it questions about what it read. Don’t classify and categorize images; just ask an AI model what it thinks it sees. … Ask it for the vibes.

That’s another thing the vibes—there are no edge cases in how something feels. There are no overlapping axes, or misplaced legends, or Rubik’s cubes full of frustrations. The vibes are much less of headache than the analytics.

Cursor

See, maybe SpaceX is also a normal startup after all:

1. They started with big, fun ambitions: big rockets; colonize Mars; technology to save humanity; build robots; buy Twitter; buy Tiktok?

2. They raised an absolutely titanic amount of money.

3. They changed their mind. They are now pivoting to the enterprise. Time for some realism—they now want a robot that makes software and does business:

   Though [SpaceX CEO Elon Musk] once forecast that humans would take off for Mars as early as 2024, he has de-emphasized reaching the planet.

   
   Instead, SpaceX on Tuesday said it had struck a deal with the artificial intelligence start-up Cursor that could result in its acquiring the young company for $60 billion. …

   
   Mr. Musk appears eager to push SpaceX further into A.I. In the deal with Cursor announced Tuesday, SpaceX said the combination with the young A.I. company, which makes code-writing software, would “allow us to build the world’s most useful” A.I. models.

On one hand, the deal makes an uncanny amount of sense: SpaceX has a ton of big computers and nobody to use them; Cursor has a bunch of customers and training data, and not nearly enough big computers.

On the other hand, man, what? In preparation for a huge IPO, SpaceX—a rocket company that is trying to build a city on the moon—is marketing its huge ambitions, its technological prowess, and its enormous financial potential by…buying a SaaS product that helps people build websites. The literal rocket ship company is trying to prove to investors that it’s a corporate rocket ship by buying an app.

I don’t know AI coding agents are a technological singularity, but they are definitely the attention singularity.

Here is what “pivot” means:

1. While sitting around in a conference room with a couple coworkers, you have an idea for a startup. “Man,” you say about some internal tool you built at your company, “Nighthawk / Stuffie / Vulcan / Hacky Sack / Shake ‘n Bake is really good. What if we made a business out of it?”

2. Or, while reading about some other huge startup—Uber, Stripe, whatever—you come up with some way to apply their solution to another problem. “Man,” you say while looking at how big the company is, “What if there was Uber for drugs? Or eBay for drugs? Or Stripe for drugs? Or iPhone for drugs?”

3. You research it. You talk to potential customers. You pitch it to investors. You convince a few of them that it is a good idea, and they give you money, to use technology to sell drugs.1

4. You hire engineers and designers, and spend months—years?—building a delightful product. You hire marketers, and spend months—years?—carefully crafting a striking brand. You recruit “design partners;” you cultivate a “community.” You launch your product with great panache.

5. At some point months—years?—later, you eventually realize that your idea was, in fact, not a good idea. Despite your comprehensive research and delightful product and striking brand and great panache, nobody is using your app.

6. This is a problem! You spent a bunch of money building something nobody wanted! You hired a bunch of experts to solve a problem that didn’t need to be solved! You filled your phone with drug dealers, and you are not doing any drug deals.

7. And so, you do it—compassionately, in front of the entire company, and later, resolutely, on LinkedIn: We must Pivot. You say you still care deeply about your original mission, but the timing is off. The product was too early; the market was not ready. You are grateful for your passionate customers—their support has meant everything—but the landscape has shifted, and we must evolve too. We cannot be afraid of hard choices. Today is a tough day, but it is the beginning of a new chapter. In building our app, we have built several exciting technologies. We have made many useful community connections. We have been researching a new idea; we have been talking to potential customers; we are sure that this one is a good idea. We have made this decision deliberately, and now we will do it forcefully. Gradually and then suddenly. We are excited for what’s next.

There are a lot of implicit assumptions loaded into The Pivot: That companies should have a clear direction; that they can change direction, but the decision is heavy; that when they do decide to pivot, they must do so definitively. And, of course, that it probably won’t work. Because great companies have clear directions; great companies do not spin in circles. “The essence of strategy is choosing what not to do,” says Harvard; the essence of execution, this blog once suggested, is steadfastly doing it:

Making the decision is more important than the decision you make. The success of a startup—or of a strategic decision in an established company, or even, to go all Chicken Soup for the Soul for a moment,2 of decisions in life and love—doesn’t depend on the specific choice. It depends on how committed you are to that choice. Often, we don’t fail because we choose the wrong thing; we fail because we refuse to choose anything at all.

Pivots represent neither of these things. They represent experimentation, doubt, and indecision. They are Silicon Valley’s equivalent of a fighting couple “taking a break”—often, the gentle beginning of an inevitable end. What comes after a pivot is usually a long, fitful march downhill. Nobody congratulations anyone on their upcoming pivot.

Four days ago, a friend told me they were in a perma-pivot. Everyone is right now, they said. Everything is moving too fast, and everyone is building on quicksand. Your market seems great today; it’s been mowed down by Anthropic’s howitzer tomorrow. You’re constantly looking for your footing. You never know what direction you’re heading.

Whatever you do, they said, don’t get caught in that dizzying whirlpool.

—

Two days ago, another friend recently told me that they run an ice cream shop that launches a speculative new flavor every Wednesday. It’s a whole event—come to the shop, try the flavor, for one time only. Some flavors are popular enough to repeat, but most are meant to be made, to be enjoyed, and to be gone. “So dawn goes down to day. Nothing gold can stay.”

You should come, they said, and best to come on a Wednesday.

You could ask two questions about all of this, I suppose. First, are pivots outdated? Nobody would accuse an ice cream shop of pivoting. Launching a new flavor every week is something between a constant experiment and a marketing campaign. The weekly “pivot” isn’t a change in direction; it’s the entire strategy.

If it takes as much time to prototype software as it does to develop an ice cream flavor, could startups be doing the same thing? Should they be doing the same thing? When software was slow to build, commitment to a direction was important, and pivoting was expensive. When software is cheap to build—and when you have to go faster to keep up with everyone else—do the rules invert? Commitment becomes expensive, and pivoting is what’s important.

We’ve talked about this before, and the downfall of the fundamentals:

[It may be that] Apple—and [Ben] Thompson, and me, and probably many of you—have an outdated understanding of how the world works. It may be that nobody was talking about Apple’s big launch event because Apple isn’t doing interesting things with AI, but it also may be because big launch events have no cultural traction. They, like us, are part of a shrinking conversational artery. Apple’s playbook is stale; its vocabulary is stale; its entire economic model of demand generation is stale. Apple’s launch events are losing their grip over Silicon Valley because everything has to be a meme or a controversy now, and polite launch events don’t produce memes or controversy.

What it means to build software has changed dramatically; why would what it means to build a software business not change just as much?

Which raises the second question: Are we outdated?3 That prior post continued:

It is easy to see when other people are the unc. They are from an earlier generation; a different time. For a while, when you’re living in the main channel, the world confirms your instincts: Their weird anachronisms are disappearing, and your vocabulary is ascendant.

But then we get stuck. And it is much harder to let go of our own sense of cosmology—because our beliefs don’t feel like fads, but fundamentals. … The old generations believed in discredited science; our beliefs are settled science; the new generations are chasing pseudoscience.

But you see the problem, right?

My friend said they were in a perma-pivot, because what other word would they use? But maybe that’s the problem. Calling it a pivot loads it up with other implications—that it’s temporary; that it’s bad; that the job of a startup founder is to find a firm direction for their company—and maybe those implications are no longer true. Maybe their situation isn’t bad; it’s just what a software company is now: A test kitchen, launching new flavors every week, where the short shelf-life is the point.

Everyone is looking for a durable place to plant a flag. What product has a moat? What market is safe? Companies are exploring, with the goal of finding a place to settle down, and make a home.

But there’s no promise technology companies will always work like that. Maybe we’re more like ice cream shops, or musicians—only as good as our latest album. Constantly evolve, or drop dead. Maybe we do not need a direction; we need to just keep moving. Maybe we cannot hide from Anthropic and OpenAI; we can only keep running from them. Maybe we aren’t pivoting; we’re just playing Pac-Man.

Don’t lose faith. You’ve got a great name, you’ve got a great team,4 you’ve got a great logo, and you’ve got a great name. Now you just need an idea—over and over and over again.

Maybe there is a pivot

What is a celebrity?

For a long time, I would’ve said that a celebrity is someone who accomplished some profound professional success, and got famous because of it. They played for the Lakers; they made a blockbuster; they made some good songs; they won an election. Though their fame might extend beyond the initial thing that made them popular, their profession was always their anchor. They were an athlete, an actor, a musician, a politician. And it was a bizarre novelty when they escaped one world and appeared in another, like when a TV star ran for president.

Now, ah, you know. Celebrity is a much looser concept. Celebrities are influencers: They are a portable endorsement that can be attached to whatever art project they are working on, office they’re running for, or drugs they’re selling. Famous people aren’t athletes or actors or musicians or politicians. They’re brands.

Anyway, this wasn’t quite what I had in mind two weeks ago, but, uh, sure?

After agreeing to sell all its assets last month for less than 1 percent of its previous $4 billion valuation, the shoe company Allbirds announced on Wednesday that it would “pivot its business” to artificial intelligence. …

In a statement, the company, which is based in San Francisco, said that an unnamed investor5 had agreed to spend $50 million to finance a shift to A.I. infrastructure. That money, the company said, will be used to buy graphics processing units, known as GPUs, powerful chips that can run calculations and analyze enormous amounts of data.

On one hand, lol, what:

“At first it read like a really well-executed April Fools’ joke,” Mr. Kleyman said of the Allbirds announcement. But, he added, “given the craziness of this industry right now, maybe we shouldn’t be surprised.”

On the other hand, it has worked before!

$24 million iced tea company says it’s pivoting to the blockchain, and its stock jumps 200%

And it can work again!

Allbirds soars 582% after sneaker firm rebrands as AI stock.

That is the general take on the situation: This is either a ridiculous Hail Mary or a cynical stunt, and either way, it’s dumb. But, everyone is crazy and everything is a meme, so, ok, maybe the best way to make some money with a niche brand that has some loose nostalgic value is by using it to make a joke.

Or—consider how Matt Levine described the joke:

I feel like, if you ran a big technology company, and you were looking to expand your artificial intelligence capabilities and needed to rent access to graphics processing units, and a GPU-as-a-service/AI-cloud company came to pitch you, and you said “so tell me a little bit about your company,” and the company said “well two weeks ago we were a sneaker company but we have since pivoted to AI,” you might say “huh, thanks but no thanks, we’re going to go with someone with a bit more AI experience and, you know, an actual data center.” Maybe that’s wrong; maybe the sneaker guys are great at AI. But you might worry.

But! If you said “so tell me a little bit about your company,” and the company said “well two weeks ago we were a sneaker company called Allbirds, but we have pivoted to AI,” your reaction might be different. Because, in this hypothetical, you run a big technology company, and you probably spent years wearing Allbirds. “I used to love Allbirds,” you might say; “high five!” And then you might sign a long-term cloud hosting agreement with Former Allbirds, because like many tech executives you have a nostalgic fondness for their brand.

Replace “Allbirds” with “Tom Brady” or “Timothée Chalamet” or “Alix Earle” and it’s not that weird! Replace “Allbirds” with “Steph Curry” and it’s practically true! Because that’s how celebrities work today: We feverishly buy whatever Alix Earle is selling us, not because we love the product, but because we love Alix Earle. Alix Earle doesn’t pivot; Alix Earle simply points her howitzer at her next market.

So, maybe Allbirds is meme-finance taken to its grim conclusion—or maybe it’s the next evolution of the firm. Before, companies were means of production that made and sold catalogs of related stuff. Now, they’re influencers. They’re an aesthetic, a portable logo that can be tacked on to whatever product or service that some particular type of customer is into at the moment. Allbirds launches a data center. Allbirds launches an incubator. Allbirds launches an agentic personal CRM for founders. Get ready for Coachella with Allbirds. If people can become brands, maybe brands can become brands.

Don’t lose faith in Allbirds. They’ve got a great name, they’ve got a great team, they’ve got a great logo, and they’ve got a great name. Now they just need an idea.