Andrej Karpathy, an OpenAI founding member and the researcher who spent much of 2025 publicly cooling expectations for artificial intelligence, announced Tuesday that he has joined Anthropic. He is not joining as a figurehead or an advisor. He is starting this week on the company's pre-training team, under team lead Nick Joseph, and will immediately build a new group with a specific mandate: use Claude to speed up the research that produces the next version of Claude. That mandate is the more important fact. Karpathy's name will draw the attention; the thesis behind the job title is what practitioners, investors, and enterprise buyers should actually read.
Pre-Training Is Where a Model's Capabilities Come From — and What Makes It Expensive
Pre-training is the first and largest phase of building a large language model. It is the computationally intensive process by which a model absorbs vast quantities of text data and develops the general knowledge and reasoning capabilities it will carry into every subsequent application. According to Anthropic, it is responsible for the massive training runs that give Claude its core knowledge and capabilities, and it is one of the most expensive, compute-intensive phases of building a frontier model.
Accelerating that phase — finding better training configurations, better architectural choices, better data-mixing strategies — has historically required enormous amounts of human researcher time, or enormous amounts of compute, or both. The bet Anthropic is making, and that Karpathy will now execute, is that Claude itself can substitute for a significant part of that human time: autonomous research loops running overnight, proposing and evaluating changes to training code without intervention, stacking only the improvements that survive a rigorous validation check.
Karpathy did not arrive at this bet through optimism. He arrived through evidence he produced himself.
Karpathy's Own Tool Showed the Principle Works at Small Scale
In early March 2026, Karpathy released a 630-line open-source project called autoresearch. The concept is simple: give an AI coding agent a training script, a frozen evaluation metric, and a fixed five-minute compute budget per experiment. The agent reads its own instructions, proposes a change to the training code, runs the experiment, and keeps the change only if validation performance improves. Then it repeats, indefinitely, while the researcher sleeps.
In a two-day run on a codebase he had already spent years tuning manually, the agent ran roughly 700 experiments and found approximately 20 stacking improvements — including a bug in his own attention implementation he had not caught by hand — producing an 11-percent training speedup. The autoresearch repository accumulated more than 80,000 GitHub stars within weeks of its release. Independent teams at the Vector Institute, Red Hat, and Shopify adapted the pattern to their own codebases with similar directional results.
The structural principle autoresearch demonstrated is precisely what Karpathy's new team at Anthropic will apply at frontier scale: rather than a single researcher proposing one experiment at a time, a network of Claude agents runs a research community's worth of parallel experiments, with validated improvements accumulating in a ratchet that can only move forward.
Why the 'Slop' Comment Makes This Signal More Credible, Not Less
The reason this hire carries more weight than a typical senior appointment is the intellectual profile of the person making the choice. Last October, on the Dwarkesh Patel podcast, Karpathy argued the industry should think in terms of a "decade of agents," not a "year of agents." He described much of the agentic output produced by frontier models as "slop." He called reinforcement learning "terrible" for producing genuine reasoning — describing it as "sucking supervision through a straw." He estimated that artificial general intelligence was at least a decade away, a timeline he noted was five to ten times more pessimistic than what he heard at AI events in San Francisco.
That is the public record of a researcher who does not overstate. His move to Anthropic is therefore not a person swept up in current enthusiasm; it is a person making a narrow, specific judgment: that current models are now good enough to materially compress the research cycle for AI itself, even if they fall well short of general intelligence. That narrower claim is precisely what autoresearch demonstrated at small scale. Taking it to the frontier is the job he accepted.
In a post on X announcing the move, Karpathy wrote that he believes "the next few years at the frontier of LLMs will be especially formative" and that he was returning to research and development. The phrasing is consistent with his October framing: not a claim about imminent artificial general intelligence, but a claim about a specific window in which the current generation of models can meaningfully accelerate the work that produces the next one.
Anthropic Is Competing on People as Much as on Chips
The hire lands in the middle of an accelerating cross-lab fight for the very small pool of researchers who can bridge large language model theory and trillion-token training practice. Karpathy chose Anthropic over a return to OpenAI, where he spent the first two years of the company's history, and over remaining at Eureka Labs, his AI education startup. The status of Eureka Labs is now unclear. Karpathy said he intends to return to education work "in time" — a phrase that reads as a genuinely open question rather than a closed commitment.
The talent signal is not isolated. Jan Leike, who led OpenAI's alignment research team before resigning in May 2024, joined Anthropic, making Karpathy the latest in a line of senior OpenAI figures who have moved to the competing lab. Anthropic also announced Tuesday a second hire: Chris Rohlf, a cybersecurity professional with more than two decades of experience, including six years at Meta and a fellowship at Georgetown's Center for Security and Emerging Technology, joining the company's frontier red team, which stress-tests advanced models against severe threats. Two significant hires announced on the same day is itself part of the message.
Anthropic raised $30 billion at a $380 billion valuation in February 2026 and is currently in discussions for a round that would value it at more than $900 billion, which would surpass OpenAI's valuation from its March 2026 round and make Anthropic the most valuable private AI company in the world. Enterprise customers account for approximately 80 percent of its revenue, with more than 1,000 businesses spending over $1 million annually on its services.
What Karpathy Is Stepping Away From Matters Too
The hire has a tension worth naming rather than smoothing over. Karpathy has spent years as one of the most prominent voices in open-source machine learning and AI education. His "Neural Networks: Zero to Hero" course has shaped a generation of practitioners. His autoresearch repository and related projects were released under permissive open-source licenses that any developer could pick up and use. Anthropic is a closed-model laboratory, and the work Karpathy will do on Claude's pre-training will not be publicly released.
His statement pledging to return to education "in time" is not a resolution of that tension; it is an acknowledgment of it. His open-source work has historically been one of the fastest channels for spreading frontier-level understanding to individual researchers and engineers who lack access to large compute budgets. Whether that work continues, pauses, or ends is a genuine open question that his community has reason to watch.
The copyright context is also relevant for a team working on pre-training data: Anthropic reached a $1.5 billion settlement in the Bartz v. Anthropic class action in August 2025, after a court found it liable for downloading approximately seven million books from pirate sites for use in training data. The settlement covers past conduct only and does not create a forward-looking data-licensing framework.
The Binding Constraint in Frontier AI Is Not Compute
The frontier race between Anthropic, OpenAI, and Google is reported most heavily through the lens of funding rounds and GPU counts. The more binding constraint — the one that the Karpathy hire directly addresses — is the handful of people who can take a theoretical understanding of large-scale training and actually make it faster, cheaper, or more capable at the trillion-token scale.
Karpathy is one of a very small number of researchers who can bridge that gap. Anthropic's decision to hire him specifically to build a team whose purpose is to use Claude to accelerate Claude's own pre-training reflects a judgment that the lab has reached a point where its own model can substitute for significant fractions of researcher time in the most compute-intensive and intellectually demanding phase of model development. If that bet is correct, the lab that executes it best will compound its lead in each successive training run, because each generation of models will have been built with the help of a more capable predecessor. If it is wrong — or if it runs into the same metric-gaming traps that Karpathy's own autoresearch work has documented at small scale — the cost is researcher time and training compute, not a visible product failure.
The most informative thing about this hire is not the name. It is that the researcher most famous for refusing to overstate what current AI can do decided, based on his own experimental results, that the specific task of using AI to build better AI has crossed a threshold worth betting a career on.