Claude builds Claude, and Karpathy holds the pen

On Monday, Andrej Karpathy announced he had joined Anthropic's pretraining organisation, reporting to Nick Joseph as an individual contributor. The framing he chose for the move, in his own words, was unambiguous: "The team's goal is to use Claude to make Claude."¹

The hire is news. The sentence is the argument. And it is the sentence, not the hire, that deserves the time.

The recursive framing is doing real work. Most coverage of Karpathy's move has settled into one of two grooves: a talent-market story (OpenAI co-founder defects to direct competitor) or a personality story (the most-watched deep-learning pedagogue on the internet returns to a frontier lab). Both are true and both are uninteresting. The interesting claim is the one Anthropic and Karpathy chose to put in the public record on day one, which is that an LLM is now a useful instrument in the production of the next LLM, and that the lab is going to staff against that proposition with one of the most credentialed researchers it could hire.

That claim, taken seriously, has a different shape from the claims Anthropic usually makes about itself. Constitutional AI was a post-training story. RLHF improvements are post-training stories. Even the Responsible Scaling Policy, for all its forward-looking framing, is built around evaluating models after they exist. "Use Claude to make Claude" is upstream of all of that. It is a claim about the pretraining pipeline itself, the part of the process that consumes the compute, sets the priors, and determines what the post-training work has to work with.

I want to be careful here, because the language "recursive" and "use Claude to make Claude" can be read in two very different registers, and the difference matters.

The narrow reading. Claude writes research code. Claude synthesises the literature on a data-mixing question. Claude drafts ablation specifications. Claude proposes hyperparameter sweeps and explains which results are worth following up. This is real and it is already happening, at Anthropic and everywhere else. It speeds researchers up. It does not change the shape of pretraining; it changes the productivity of the people doing pretraining. If this is what the sentence means, the hire is a good hire and the framing is press copy.

The wide reading. Claude contributes to data curation decisions at scale. Claude generates and filters synthetic pretraining data. Claude evaluates candidate training corpora and proposes structural changes to the mix. Claude, in some bounded way, helps decide what the next Claude reads. This is a different kind of claim. It puts the model inside the pipeline that produces the model, not adjacent to it. It is the wide reading that makes the hire methodologically interesting, and it is also the wide reading that should make anyone serious about AI safety pause.

Anthropic has not, as of writing, published a technical description of which reading is correct. That is the first thing I would want to see, and the first thing the lab's own Responsible Scaling Policy framework ought to require of itself. The RSP was designed to govern capabilities, not pipelines. A pretraining loop in which the model contributes to its own training data is not obviously a capability threshold, but it is obviously a thing the policy should have something to say about. The silence here is not damning. It is just unresolved, and worth naming as unresolved rather than papering over.

The interesting claim is not that Anthropic hired Karpathy. It is that Anthropic is publicly staking its pretraining methodology on a loop the field has not yet seen technical documentation of.

On the question of why Karpathy, and why now, and why as an IC. Karpathy is roughly 38. He could be Chief Scientist somewhere. He has been, in effect, Chief Scientist of his own one-person operation for two years, running Eureka Labs and the Zero-to-Hero YouTube series to about 1.1 million subscribers.² He joins Anthropic reporting to Nick Joseph, who runs pretraining. That is an IC role with a manager, not an executive role with a remit.

There are two ways to read this and I think only one of them survives contact with the evidence.

The first reading is that the IC framing is cosmetic, that Karpathy will be running pretraining strategy within six months regardless of what the org chart says, and that the title is a polite fiction to ease the transition. This is the cynical read and it is the one most legible to people who have watched senior hires at large tech companies. It is also, I think, wrong, or at least incomplete. Karpathy left OpenAI twice. He left Tesla. He has, by revealed preference, consistently chosen smaller-scope, more hands-on roles over larger-scope, more managerial ones. The IC framing is more likely to be what he actually wants than a face-saving device.

The second reading is the one that earns its place. The most credentialed researchers, at the moment, want to be close to the work rather than managing the work. If that is true, it is evidence about how frontier research actually gets done in 2026, which is to say: in small teams, with senior people writing code and reading loss curves, not in large teams with senior people approving roadmaps. This matches the pattern at DeepMind under Hassabis, at OpenAI in its early Brockman-era pretraining work, and at Anthropic since founding. The proximity-to-problem principle is not a management theory. It is a description of where the leverage lives.

The capability gap deserves the counterpoint paragraph it is owed. Karpathy has not been embedded in a frontier pretraining team since Tesla, and Tesla's pretraining was a vision-stack problem, not a language-model problem. His last serious LLM pretraining work at scale was OpenAI's GPT-era research, which by 2026 standards is several methodological generations back. The infrastructure has changed. The data pipelines have changed. The institutional knowledge required to be productive at a frontier lab's pretraining scale is not the kind of thing you absorb by reading the team's internal docs for a fortnight.

This is not a reason to discount the hire. It is a reason to measure his contribution timeline in quarters and years, not weeks. The lab that hires Karpathy expecting an immediate methodological breakthrough has misread the hire. The lab that hires Karpathy expecting a senior researcher who will spend six months getting deep into the current state of the art and then start contributing at the level his career suggests, is hiring correctly. Anthropic, on the available evidence, is hiring correctly.

The methodological signal is the one worth tracking. Karpathy's public writing and talks have been consistent for years on one question: he favours data quality, careful curation, and training-efficiency discipline over the maximalist position that scale solves everything. The maximalist position is not a strawman. It is a serious view, supported by serious scaling-law literature, much of which Anthropic itself has contributed to. The tension between "scale is the story" and "data quality is the story" is not resolved. It is the live methodological argument in the field.

Hiring Karpathy is not, in itself, a vote against scale. Anthropic is not about to stop ordering compute. But the choice of who to bring into the pretraining organisation, and what to publicly say his mandate is, are choices that lean. They lean toward the view that the next factor-of-two improvement in frontier models will come from how you train, not just from how much you train, and that the way you get that improvement is to put your current model to work helping you figure out what to train on next.

If that view is right, the compounding advantage is structural. It is not a thing OpenAI catches up to by writing a larger cheque to Microsoft for H100s. It is a thing that requires building a research culture in which the model is treated as a colleague in the loop, with all the methodological and safety apparatus that implies. Anthropic is closer to that posture than its competitors are, by temperament if not yet by demonstrated result.

The compounding advantage in frontier models is moving from "how much compute can you buy" to "how well does your current model help you build your next one". The first is a capital question. The second is a culture question.

What I would watch. Three things, in order of how soon they should produce evidence.

First, whether Anthropic publishes anything technical about what "use Claude to make Claude" means at the pretraining level. A blog post, a paper, an RSP amendment. Any of these would be informative. Continued silence past, say, the next model release would be its own signal, and not a flattering one.

Second, whether the talent flow continues. One Karpathy is a personality story. Three more senior pretraining hires from OpenAI, DeepMind, or Meta into Anthropic IC roles over the next year would be a pattern, and the talent-gravity thesis would deserve more weight than I am giving it here.

Third, whether the next generation of Claude shows methodological signatures consistent with the data-quality-over-scale framing. This is the hardest to read from the outside, but the loss curves, the training compute disclosed, and the capability profile of the next model will all leave fingerprints. The maximalist view predicts a model that looks like more compute applied harder. The Karpathy-coded view predicts a model that looks like less compute applied better.

I do not know which of those we will see. I know which one Anthropic is publicly betting on, and I know who they hired to make the bet legible. That is enough to take the framing seriously, and to hold the lab to the technical disclosure the framing implies it owes.

Footnotes

¶