Andrej Karpathy Joins Anthropic to Use Claude to Train the Next Claude

Here's a sentence that would have sounded strange three years ago: one of the most respected AI researchers alive is now paid to use an AI chatbot to improve the science of training that same chatbot. That's Andrej Karpathy's new job.

Karpathy announced on May 19 that he's joining Anthropic. He starts immediately on the pre-training team, reporting to Nick Joseph, Anthropic's head of pre-training. His specific mandate, according to multiple reports: build a new group focused on using Claude to accelerate the research that produces the next version of Claude.

It's recursive in the most literal sense. And for the broader AI industry, it marks a shift from an era where raw compute was the defining variable to one where the ability to automate the research process itself may matter just as much.

How He Got Here

Karpathy is one of the few researchers in AI whose career reads like a genuine highlight reel rather than a résumé curated for LinkedIn. He was born in Slovakia, grew up in Canada, and got interested in machine learning during his undergraduate years at the University of Toronto, which at the time was one of the few places in the world where Geoffrey Hinton was actively working on neural networks.

He went on to get his PhD at Stanford under Fei-Fei Li, focusing on image and language understanding with neural networks, before co-founding OpenAI with Sam Altman, Greg Brockman, Ilya Sutskever, and others in 2015. That founding cohort was a concentration of technical talent that hasn't been matched since. Karpathy spent two years at OpenAI working on deep reinforcement learning and computer vision before leaving for Tesla in 2017.

At Tesla, he ran the Full Self-Driving and Autopilot programs as director of AI. This wasn't a research role. It was an engineering and systems role at enormous scale: building a real-time perception pipeline for millions of vehicles, processing data from eight cameras simultaneously, making architecture decisions that had to work reliably in adverse conditions without a second chance to debug them in the field. The experience gave him something rare in AI research: genuine understanding of what it takes to make a system work outside of benchmarks.

He left Tesla in 2022, returned to OpenAI for one year, then announced in July 2024 that he was starting Eureka Labs. The pitch was ambitious and earnest: an AI-native school that would use AI teaching assistants to let anyone learn anything. The first product was LLM101n, a course that would walk students through training a language model from scratch. Eureka Labs raised roughly $20 million in seed funding from investors including Conviction and Sam Altman. The idea was clearly personal. Karpathy had been doing AI education informally for years through YouTube tutorials and X threads that became standard references. Hundreds of thousands of people learned neural network fundamentals from his lectures.

But Karpathy isn't at Eureka Labs anymore. In his announcement, he was clear about what changed: "I think the next few years at the frontier of LLMs will be especially formative. I am very excited to join the team here and get back to R&D." He added that he remains "deeply passionate about education" and plans to return to that work in time. The framing suggests the move isn't a repudiation of the education thesis but rather a sense of urgency about this particular moment in AI development. He wants to be at the frontier, not adjacent to it.

Before his Anthropic announcement, Karpathy had built one of the largest audiences of any active AI researcher through open-source work and educational content. His nanoGPT project, a minimal clean implementation of a GPT model in about 300 lines of Python, became a standard reference for anyone wanting to understand language model internals without wading through enterprise codebases. His YouTube lectures on backpropagation and transformer architectures have drawn millions of views from people who are now themselves working in AI. He didn't just study the frontier; he made it accessible.

That combination of deep technical credibility and broad public reach matters for Anthropic in ways beyond the research work itself. Karpathy's hiring signals something to other researchers thinking about where to go next. His nanoGPT repository is probably in the reading list of a majority of AI researchers under 35. His willingness to spend his most productive years at Anthropic rather than at OpenAI or Google, or at his own company, is a non-trivial signal about where he thinks the best work is being done. That kind of recruitment signal is hard to manufacture.

There's also a practical dimension to his background that doesn't show up in his publication list. Karpathy has spent years communicating complex ideas to large audiences. That skill applies directly to research leadership. A team leader who can explain what they're building clearly, who can write and present ideas precisely, who has experience teaching the underlying concepts to beginners, is unusually well-positioned to run a research group that needs to write, iterate, and communicate rapidly. Technical depth and communication clarity are not always the same person. In this case they are.

What He's Actually Building at Anthropic

The pre-training team at Anthropic is responsible for the large-scale training runs that give Claude its core knowledge and capabilities. Pre-training is the phase where a model learns from massive amounts of text data, developing its understanding of language, facts, reasoning patterns, and world knowledge. It's the most capital-intensive phase of building a frontier model, involving compute jobs that can cost tens of millions of dollars per run, and thousands of decisions about architecture, data quality, training schedules, and optimization.

Pre-training research is also among the most labor-intensive parts of AI development. Running an experiment requires writing code, setting up a training pipeline, waiting hours or days for results, interpreting what the numbers mean, and deciding what to try next. A large pre-training team might run dozens of experiments per week. The bottleneck isn't always compute. Often it's the human bandwidth to design, run, and analyze experiments quickly enough to actually make progress before the compute window closes.

Karpathy's role goes beyond simply joining the team. He's building a new group within it, focused on using Claude to speed up the research process. That's a different goal from buying more GPUs. The idea is to use Claude as an autonomous research assistant, running experiments, analyzing results, and iterating on improvements with minimal human intervention in the loop. The human researchers set the direction, evaluate the significant findings, and make the judgment calls. The AI handles the high-volume experimental work in between.

This isn't speculative. In March 2026, Karpathy released a 630-line open-source Python script called AutoResearch that demonstrated exactly this approach at a smaller scale. The script gives a coding agent a training script and a five-minute compute budget per experiment. The agent proposes code changes, runs tests against a validation metric, and keeps only improvements that pass. In a two-day run, the system executed roughly 700 experiments, found approximately 20 stacking improvements (including a bug the human researchers had missed), and achieved an 11 percent training speedup. The GitHub repository accumulated more than 8,000 stars in its first week and more than 80,000 stars by early April 2026 as developers at other companies adapted the pattern for their own workflows.

AutoResearch was notable partly for what it found and partly for the pattern it demonstrated. The AI wasn't just running experiments faster. It was doing the full research cycle autonomously: form a hypothesis, modify the code, run the experiment, evaluate the result, iterate. That pattern, if it scales to frontier pre-training with more capable models on larger training runs, changes the economics of AI research fundamentally. Research velocity could increase by an order of magnitude without proportionally increasing researcher headcount.

At Anthropic, Karpathy will be applying that approach at a much larger scale, using multiple Claude agents running in parallel on pre-training experiments. The goal is to use Claude to help produce better versions of Claude. It's the kind of recursive feedback loop that's been theorized in AI research for years. Now there's a team dedicated to making it concrete and reliable.

This context makes it worth revisiting something Karpathy said publicly last year. In an October 2025 interview on the Dwarkesh Patel podcast, he described much of the agentic output produced by frontier models as "slop." He argued the industry should expect a decade of progress on agents rather than a year, and estimated AGI was at least five to ten years away, far more conservative than the estimates circulating in San Francisco at the time. He also called reinforcement learning "terrible" for producing genuine reasoning, describing it as "sucking supervision through a straw."

Those are pointed critiques. They're also consistent with why he'd join a team focused on pre-training rather than on building agent products. If current models produce slop, the fix isn't more scaffolding around weak models. It's better base models. Pre-training is where that happens. The "slop" comment and the Anthropic hire are the same worldview applied at different levels: dissatisfied with the current output, convinced the fix is in the foundation.

The Multiplier Anthropic Is Betting On

Anthropic's decision to structure this hire around using AI to improve AI research reflects a broader shift in how frontier labs think about research efficiency. Google, OpenAI, and Anthropic are all exploring ways to use their own models to accelerate parts of the research and engineering pipeline. The concept has moved from aspirational to active in the past year. But placing Karpathy in charge of the specific problem of pre-training, with Claude as the primary tool, is a more aggressive and public version of that bet than any competitor has announced.

For Anthropic, the competitive dynamics are clear. The company is competing with OpenAI and Google on model capability. Both organizations have substantially more resources. OpenAI has its deep partnership with Microsoft and is generating significant revenue from ChatGPT and enterprise APIs. Google has effectively unlimited compute through its internal infrastructure and chip program. Anthropic has raised several rounds of capital and has growing Claude revenue, but it can't match either competitor dollar for dollar on raw training spend.

What Anthropic can try to do is find a multiplier. If using Claude to automate parts of the pre-training research cycle means Anthropic gets materially more research throughput from the same team size, that partially offsets the compute disadvantage. It's the kind of asymmetric bet a well-capitalized challenger makes against a larger incumbent when it can't win the resource war directly.

This hire also needs to be read alongside Anthropic's other recent moves. The company recently acquired Stainless, the SDK generation startup used by OpenAI, Google, and Cloudflare to build their developer-facing libraries, signaling intent to control more of the developer tooling layer. The Karpathy hire is a different kind of move, aimed at the model quality layer rather than the distribution layer. Together they sketch an Anthropic that's trying to reduce dependence on external infrastructure at both the model-building and model-delivery ends of the stack.

For researchers and developers watching the field, the deeper implication is about what happens when a frontier AI company's most critical internal research workflow starts running substantially on its own model. If Claude can materially speed up the pre-training research that produces the next Claude, that's not just an efficiency gain. It's a shift in the economics of frontier AI development. Research cost per unit of model improvement goes down. The advantage compounds across training runs.

The feedback loop is what makes this hire unusual relative to most senior research appointments. Most of those are about adding human expertise. Karpathy's role is explicitly about building a system where AI expertise augments human expertise, and then using that combination to produce better AI. If the experiment works at scale, the model that comes out is better. The research that produced it was partly automated. And the next round of research uses that better model to automate itself further.

This is not the science fiction version of recursive self-improvement. It's slower, more constrained, requires careful human oversight, and will produce incremental gains rather than sudden capability jumps. But it's real, it's already demonstrated to work in Karpathy's own AutoResearch experiments, and now it has a dedicated team at one of the three most important AI labs in the world.

The reaction in the research community has been striking. Shortly after the announcement, the Hacker News thread on Karpathy's post drew thousands of comments. Most recognized the hire as a significant signal. Some raised pointed questions, including concerns about Karpathy's association with Tesla's FSD program and its safety record, and skepticism about whether AutoResearch's approach is genuinely novel compared to existing methods like AlphaEvolve. Those critiques matter because they reflect real tensions in how the field thinks about automated research and its limits. Running 700 experiments autonomously on a small training setup is a proof of concept. Running 700,000 experiments on a frontier training run is a different engineering problem with different failure modes.

Karpathy is clearly aware of the gap. His public communications since leaving Eureka Labs have emphasized caution about AI agent quality and skepticism about near-term AGI timelines. He's not arriving at Anthropic with the belief that AutoResearch is already the answer. He's arriving with a proven approach, a clear sense of the research question, and a team mandate to find out how far it can scale. The difference matters. A researcher who thinks the problem is solved is dangerous. A researcher who thinks the problem is tractable, has a methodology, and is realistic about what's unknown, is exactly what this kind of work requires.

For anyone tracking where model capability is going over the next eighteen months, Karpathy's move to Anthropic is a better leading indicator than any benchmark result or press release. He's worked at enough of the major labs to have a calibrated view of where the real progress is happening. He left a funded startup he believed in. He's betting his near-term career on the thesis that the pre-training frontier, and specifically the ability to automate pre-training research, is where the most important work in AI is right now. That judgment is worth taking seriously.

The recruitment also sharpens a question that's been circulating in the field for a while: what happens to AI labs that don't invest early in AI-accelerated research? If Anthropic succeeds in using Claude to meaningfully speed up pre-training iterations, and if that advantage compounds across training runs, the labs that stick to purely human-driven research workflows will fall behind on research velocity regardless of how much they spend on compute. The gap might not show up in the next model release. It could show up in the one after that, when the speed differential in the research pipeline starts producing larger capability differences than the hardware difference explains.

If you're evaluating which Claude model or competing LLM fits your team's current needs while this next generation of research plays out, the Claude vs GPT vs Gemini comparison on AIntelligenceHub breaks down where each model currently stands on capability, cost, and use-case fit.

TechCrunch broke the story on May 19, with additional details in their full reporting on Karpathy's role and background at Anthropic.