Two OpenAI stints, a Tesla AI leadership role, and one of AI’s loudest technical megaphones are now moving inside Anthropic.
Karpathy joins Anthropic’s pre-training group under Nick Joseph
Andrej Karpathy, the AI researcher who co-founded OpenAI and later led AI work at Tesla, has joined Anthropic to work on pre-training, according to TechCrunch. The move puts one of the most recognizable researchers in modern deep learning directly inside the machinery that shapes future Claude models.
Karpathy announced the move Tuesday on X.
“I’ve joined Anthropic. 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.”
TechCrunch reported that Karpathy started this week and is working under Nick Joseph, Anthropic’s pre-training team lead. The company describes pre-training as the large-scale training work that gives Claude its core knowledge and capabilities. It is also one of the most compute-heavy and expensive stages in building a frontier model.
That makes the hire more than a résumé headline. Anthropic is putting Karpathy into one of the highest-leverage parts of model development, not a public-facing evangelist role or a research-adjacent advisory seat.
An Anthropic spokesperson told TechCrunch that Karpathy will start a team focused on using Claude to accelerate pre-training research. That detail matters. It suggests Anthropic wants its own model to help improve the research process behind future models.
Analysis: The signal here is direct: Anthropic is not just buying experience. It is trying to shorten the feedback loop in frontier-model development, where better research tooling can matter alongside raw compute.
A rare résumé: OpenAI founder, Tesla AI lead, technical educator
Karpathy brings a blend of model research, computer vision, engineering leadership, and public technical teaching that few AI researchers can match.
At OpenAI, he worked on deep learning and computer vision before leaving in 2017 for Tesla. At Tesla, he led the company’s Full Self-Driving (FSD) and Autopilot programs before departing in 2022. He then returned to OpenAI for one year before leaving again in 2024 to start Eureka Labs, a startup focused on applying AI assistants to education.
His public profile also extends well beyond corporate AI labs. Karpathy has taught an online course called Neural Networks: Zero to Hero, aimed at helping students build neural networks from scratch in code. He also posts lectures on LLMs and AI on YouTube.
That teaching track record is not incidental. Anthropic’s stated plan for Karpathy — building a team that uses Claude to speed up pre-training research — demands someone who can connect theory, implementation, and developer workflows.
| Person | New Anthropic role | Prior experience cited in source | Strategic relevance |
|---|---|---|---|
| Andrej Karpathy | Pre-training, under Nick Joseph | OpenAI co-founder and researcher; Tesla FSD and Autopilot lead; Eureka Labs founder | Frontier model training and AI-assisted research |
| Chris Rohlf | Frontier red team | More than 20 years in cybersecurity; Yahoo’s “The Paranoids”; Meta; Georgetown CSET CyberAI project | Stress-testing advanced AI models against severe threats |
Anthropic also brought on Chris Rohlf to its frontier red team, TechCrunch reported. Rohlf said on X:
“We have a real opportunity in front of us to dramatically improve cyber security with AI. I can’t think of a better company or team to join at this critical moment in time.”
The two hires point at different pressure points: one on making frontier models stronger, the other on testing them against severe risks.
Anthropic’s bet: use Claude to improve Claude research
The most important line in the hiring news is not Karpathy’s title. It is Anthropic’s plan for him to start a team focused on using Claude to accelerate pre-training research.
Pre-training is where large language models absorb the broad statistical structure that later appears as knowledge, reasoning, coding ability, and general capability. The source material does not detail Anthropic’s exact methods, datasets, architectures, or timelines. But it does say this phase is central to Claude’s core capabilities and expensive to run.
Analysis: If Anthropic can use Claude to make pre-training research faster or more efficient, the payoff could show up in better experiment design, faster debugging, sharper evaluation loops, or more productive researchers. The announcement does not prove any of that has happened. It does show where Anthropic wants to apply AI internally: not only in customer products, but inside the research process itself.
That is where Karpathy’s background fits. He has worked across deep learning, computer vision, autonomous-driving AI systems, and LLM education. TechCrunch describes him as one of the few researchers who can bridge LLM theory and large-scale training practice.
For readers tracking the developer side of the same AI race, this hire sits near the same pressure zone as coding agents and model-assisted software work. MLXIO has covered that developer shift through OpenAI Codex Stops Making iPhone Users Babysit Tasks and the risk profile around Anthropic’s tools in Claude Code Exposes the New Coding Risk: Blind Trust.
The connection is not that Karpathy is joining a coding-product team. He is not, based on the source material. The connection is deeper: frontier labs increasingly need models that can assist technical work, including the work of building better models.
No new Claude model yet — the evidence will be in future releases
Karpathy’s arrival does not mean Anthropic is announcing a new Claude model, benchmark result, or product release. TechCrunch says he has joined the pre-training team and will form a group focused on using Claude to accelerate pre-training research. That is the confirmed news.
What remains unclear is how this affects Eureka Labs. TechCrunch reported that Karpathy has not shared many updates about the startup since launch and that it is not clear whether he will continue with it. Karpathy said he remains committed to education.
“I remain deeply passionate about education and plan to resume my work on it in time,” Karpathy said.
The practical read is simple: judge the hire by model behavior, not the announcement. Future Claude releases may reveal whether Anthropic’s pre-training work improves capability, reliability, coding performance, reasoning, or research velocity. But those outcomes are not yet in evidence.
The next signals to track are specific: whether Anthropic discloses more about Karpathy’s team, whether it describes Claude-assisted research workflows in more detail, and whether future Claude launches show capability gains that Anthropic links back to pre-training changes.
For now, Anthropic has added a founder-level OpenAI veteran to the part of the company that shapes Claude’s core capabilities. In frontier AI, that is about as close as a hiring move gets to product strategy.
The Bottom Line
- Anthropic is adding a high-profile researcher to one of the most important stages of frontier model development.
- Karpathy’s work on using Claude to accelerate pre-training research could shorten Anthropic’s model improvement cycle.
- The hire signals intensifying competition among AI labs for talent that can improve both research quality and compute efficiency.
