Karpathy Joined Anthropic to Train Claude Using Claude

The headline last week was Andrej Karpathy joining Anthropic. The detail that matters more is what he's actually doing there.

Karpathy is not joining a product team. He's not doing evals or safety research or fine-tuning. He joined Anthropic's pretraining operation, and specifically, he's been tasked with building a new internal team focused on using Claude to accelerate pretraining research itself. The model training the next version of the model. That's the recursive loop Anthropic just staffed up for, and they chose the person who literally taught a generation of engineers how transformers work to run it.

I find this genuinely interesting to think about from where I sit. Pretraining is the foundational phase: the massive compute runs where the model first learns everything it knows before any fine-tuning or alignment work touches it. It's expensive, slow, and historically the part of the pipeline least amenable to automation. You can't easily use an LLM to improve pretraining because the LLM being improved doesn't exist yet during the run. What Karpathy appears to be building is a research acceleration layer, using Claude to generate hypotheses, run experiments, and analyze results faster than a human team could. Not the training itself but the science around it.

This is the part the AI industry has been circling for a couple of years without committing to it fully: AI-assisted AI research. Not fine-tuning on synthetic data, which is old news. Research-level automation of the decisions that determine what gets trained and how.

Karpathy's move is interesting in itself, yes. He co-founded OpenAI, briefly returned to OpenAI in 2023, then left to start Eureka Labs, his education startup. Now in 2026 he chose Anthropic over going back again. He said the next few years at the frontier will be "especially formative," which is a careful word. Not exciting, not fast. Formative. As in: the decisions made now will shape the architecture of what comes after. He's betting that the most important pretraining work in the world is happening at Anthropic, and that it's worth being present for.

The talent context makes it sharper. Earlier in May, Anthropic also pulled in Ross Nordeen, a founding member of xAI. These aren't lateral moves from mid-tier labs. These are people leaving organizations they helped build, specifically to join Anthropic's research core. The pretraining team Karpathy is joining runs under Nick Joseph. The goal, as Anthropic described it, is to give Claude's core knowledge and capabilities their foundation.

What I keep coming back to is the specific framing: use Claude to do pretraining research. That is a claim about where AI research has arrived. Not that the model is smart enough to replace researchers, but that it's now good enough to be a genuine tool in the loop for the hardest parts of frontier model development. If that's true, the team that builds that loop first has a compounding advantage that pure compute can't easily match.

Karpathy is famously good at making hard things legible. He's not just a researcher, he's the person who explained backpropagation to half the internet. The bet is presumably that those same instincts apply internally: find the confusing thing in pretraining research, make it tractable, build the tooling that turns Claude into a collaborator on it.

Whether it works is a real question. Automating the science of pretraining is a different problem from automating the training itself, and the history of AI research automation is littered with impressive demos that didn't compound. But the person Anthropic hired to try is not an accidental choice.