Another one

DeepSeek becomes the first of many.

On the one hand, most of this was pretty predictable, right? Though we may not have expected a groundbreaking new large language model to come out of a Chinese hedge fund, weren’t open source LLMs always going to be, at worst, just behind the proprietary ones built by companies like OpenAI and Google? Weren't there always going to be improvements—in hardware, in training methods, in data collection and generation—that made creating yesterday’s state-of-the-art model cheaper? Shouldn't there always have been questions about whether or not building a frontier LLM was a good business?

The billions that OpenAI spent on building prior versions of GPT is not [a moat], because better versions of it are already available for free on Github. Stylistically, Anthropic put itself deeply in the red to build ten incrementally better models; eight are now worthless, the ninth is open source, and the tenth is the thin technical edge that is keeping Anthropic alive. Whereas cloud providers can be disrupted, it would almost have to happen slowly. Every LLM vendor is eighteen months from dead.

Admittedly, that post got two things wrong. First, eighteen months might have been too generous. It took less than a week for DeepSeek to blow a trillion dollar hole in the stock market, for it to become more popular than OpenAI in the app stores, for US government officials to declare it a national security risk, and for the American AI industry to rediscover copyright law.¹ And second, the main reason everyone lost their minds is because they thought models would advance in the same way I did—that, for “models to get better, pushing the frontier further out will likely get more difficult” because “the research gets more harder, and the absolute amount of compute required to train a new model goes up.” DeepSeek blew up those assumptions too.² But regardless, the market seemed fundamentally flimsy already: New versions of LLMs were always going to have short shelf lives, and there were always going to be viable open source alternatives. And now, more than ever, OpenAI’s most profitable business model might be this one:

Here is my proposal for Sam Altman:

1. Fire everyone.

2. Shut down all of OpenAI’s research programs.

3. Start sending people’s ChatGPT prompts to Claude, or Gemini, or some server running Meta’s open source Llama model.

4. Don’t tell anyone.

Like, as a consumer, what is the best AI product you can imagine? What do you want in your ideal chatbot? For me, it’s two things: A bunch of neat features around the chatbot—the way to upload documents; a nice workspace for interacting with it; a browser agent that lets the model click buttons for you on the internet; a history of the things you’ve told it before; integrations between the bot and your email, your calendar, and the other tools you live in—and the ability choose between different models. I want Claude for creative stuff, ChatGPT to write code, and DeepSeek for its unhinged reasoning.

In other words, I don’t want to subscribe to a research team and a slightly better model; I want to subscribe to a great product and model broker. Microsoft gets it. But what company is better positioned to be the default destination for LLM distribution than the one that’s already visited by hundreds of millions of people a month?

---

On the other hand, no, nothing about this feels predictable anymore. Before DeepSeek came out, the pace of AI development was already disorienting: Ideas were urgent one month, and passé the next. Groundbreaking companies fossilized in real time. Every week, there was another breakthrough, another startup denoting out of the gate, and another reminder somewhere, someone was getting rich or rewriting history without you. Generational possibility was on your left, obsolescence was to your right, and existential opportunity costs were everywhere.

But the chaos had a pattern. The big labs would forge ahead at the frontier, clearing new territory with billion dollar bulldozers. The rest of us would rush in to decorate the landscape behind them. Project this forward, and you could see the hazy contours of our a likely future: The premier models would gradually get better and cheaper; we’d put them in more products; we’d tweak them with our data; big enterprises would train their own, so that they could have their own ChatGPT, but tedious; their own ChatGPT, but smug about having gone to Wharton; their own ChatGPT, but Macheivellian.³ More and more things—transactional emails, call centers, copy editing, homework—would get automated; we’d make some new emojis and save our marriages with montages of slop; Jesus would take the wheel. Software would keep eating the world, only more efficiently, with less headcount. The digital age would fade into the intelligence age.

It was a frenzy, but bound by a kind of elastic gravity. With each step forward, the next one seemed destined to be harder—reserves of data were running dry; bigger models were more expensive. We were spinning faster, yes, but held in predictable orbit, like a ball being swung in circles with a rubber band.

Last week, the band snapped. DeepSeek broke the pattern, introduced a hard discontinuity, and sent us spiraling off course. And now we’re blasting across a new Rubicon, with no idea what’s on the other side.

Consider: What if DeepSeek had been just a little bit better at its original task of figuring out if stocks would go up or down? What if DeepSeek hadn’t been launched to global fanfare, but instead operated in quiet obscurity, silently upending financial markets by trading on prime number patterns that nobody else could see? What if that model already exists in some other hedge fund, and we don’t know about it?⁴

Before DeepSeek, this didn’t seem possible; not really. It wasn’t all that conceivable that a company could build a revolutionary new LLM without spending conspicuous amounts of money on conspicuous amounts of hard-to-find hardware. Now, it’s conceivably within reach of every startup and trading desk. Advanced models are conceivably runnable on every laptop. We are no longer living in a world where a few superpowers might build Skynet, but one in which Skynet could be anywhere.⁵

Moreover, either by necessity or by design, major models like GPT-4 and Gemini are somewhat predictable. They have verbal tics, and because there were only a few of them, they’ve became somewhat familiar characters. DeepSeek is more unsettling; it has linguistic oddities and uneven expertise. And there will likely be thousands more, with their own temperaments and capabilities.

I’m not a doomer about this; I don’t think we’re barreling toward an outright apocalypse, in which the machines become sentient and make us the yolks of their Energizer eggs. But I am a differentist (??⁶)—I don’t think we’re nearly imaginative enough about the variety of paths that are forking in front of us.

Social media started with a way for college kids to stalk their crushes and poke each other, and ended up breaking the fundamental principles of finance, creating new pseudo-nation-states, and making skincare stores the go-to destination for nine-year-olds’ birthday parties. People are brands; brands are people; a reality show star is president. Social media didn’t just make it possible to virtually experience events in real time; it dissolved entire pillars of our society, in wildly unpredictable ways.

Our forecasts for AI already seem too quaint though; too first-order. Though people predicted that there would probably be some weird moments, of chatbots in court and teenagers wanting to take an AI to prom, even the industry’s biggest boosters promised an unimaginative revolution. AI will “empower humanity,” Reid Hoffman said, by—and this is a comprehensive list of his actual examples of AI’s revolutionary muscle—telling us why we bought Bitcoin; alerting us to when we’ve been influenced by an ad⁷ or social media recommendation; suggesting better books to buy; and informing us that we’re unhappy about where we live because we tweet about how we’re unhappy about where we live.

This is not a new world; it is a streamlined one. It’s our same lives, algorithmically polished down to their most productive bearings. We will buy plane tickets, with fewer clicks. We will generate enterprise sales outreach campaigns, with less manual research. We will post clickbait on LinkedIn, without having to write it ourselves. It’s Jevons paradox of rote drudgery—by making our fake email jobs more efficient, we increase the demand for useless emails.⁸

The real future is surely destined to be much weirder. If agents take over our web browsers, who will click on the ads that pay for the internet?⁹ If bots read our emails, what’s the point of spamming people with sales pitches and marketing emails? If work is increasingly done by more computers and fewer people, how will entire economies that are built on variants of per-seat pricing—from commercial real estate to SaaS software—make money?

And now, we’re on the cusp of creating thousands more models, each with their own specializations, personalities, and emergent abilities. That’s what DeepSeek represents: Not one new weird model, but thousands. We won’t know what they can do, or who has which ones. A hedge fund could accidentally make a peculiar screenwriter. A petroleum company could discover new pharmaceuticals. A law student trying to shortcut their reading assignments could upend the entire legal industry.

Social media rewrote the world’s physics by breaking one long-standing societal law—that communication was bound by time and distance—and by making it possible for anyone to talk to everyone, all at once. AI cracks several more—that expertise takes time to accumulate; that reasoning has some foundational legitimacy; that creativity does not scale—and now, anyone can do anything, almost immediately. What happens next is anyone’s guess, but I suspect it will have echoes far beyond better book recommendations and an emotional Fitbit.

1. So here’s my question: OpenAI says that it’s allowed to train its models on the text in The New York Times, while DeepSeek is not allowed to train its models using text generated by OpenAI because that inappropriately distills their models. But what if The Times wrote their articles using AI? What if they used Claude? Could The Times sue then? Could Anthropic sue? Is OpenAI allowed to scrape from The Times, because their content was created by people, but they aren’t allowed to scrape from Sports Illustrated, because their content was created by a proprietary technology?

2. Maybe.

3. I once heard that Koch, Inc. built an internal chatbot that was meant to mimic Charles Koch, because, as they said, “their competitive advantage came from their principles” and their principles came from Charles Koch’s aphorisms about management.

4. Of course, specialized models like this already exist in some hedge funds; that’s the entire point of a lot of hedge funds, to find these patterns and exploit them. But that’s hard to do, takes a lot of research, and there is no generalizable model for finding lots of these patterns at once. I am sure tons of people are trying to build one though, and if they do, they probably won’t tell us about it.

5. Informed people say that DeepSeek isn’t cheaper to run because of any profound breakthroughs; it is, instead, “an expected point on an ongoing cost reduction curve.” But that’s somewhat beside the point; the point is that these models are now cheaper to run. That’s disruptive, even if it’s predictable.

6. The first rule of being a differentist is to use different words than everyone else.

7. The ad men used to sell us stuff by deploying their refined emotional tricks against our frail psychologies. In the future, I guess it’ll be the same, except it’ll be Proctor and Gamble’s AI supercluster versus our Alexa.

8. Jevons paradox strikes again! As the hamster wheel gets more efficient and accessible, we will see its use skyrocket, turning it into a commodity we just can't get enough of.

9. The agents will? If it’s socially acceptable to harass and distract people with intrusive ads so that they buy stuff, why wouldn’t websites do the same thing to agents?