Anthropic hires OpenAI co-founder Andrej Karpathy

Andrej Karpathy, an AI researcher who co-founded OpenAI and later led Tesla’s computer vision team, said on May 19 that he is joining Anthropic’s pretraining team.

Anthropic said Karpathy starts this week and will build a team focused on using Claude to speed up pretraining research, which gives its models core knowledge and capabilities.

Karpathy left Tesla in 2022, briefly returned to OpenAI, and then worked on AI education startup Eureka Labs until this move.

The hire adds to Anthropic’s recent recruitment push, which also brought in former xAI founding member Ross Nordeen earlier in May.

🔗 Source: CNBC

🧠 Food for thought

Implications, context, and why it matters.

Karpathy joins a team central to Anthropic’s mission

Andrej Karpathy is joining Anthropic’s pretraining team, which builds the next generation of the company’s large language models, including Claude 1.
The team follows a “big science” approach. Anthropic puts its effort into a small number of large research programs instead of many smaller ones 1.
Its work sits at the center of the company’s goal to build “reliable, interpretable, and steerable AI systems” 1.
That importance shows up in pay. Annual compensation for research engineers or research scientists in pretraining ranges from US$350,000 to US$850,000 1.

The hire signals a strategic bet on AI-assisted AI research

Karpathy’s work on using Claude to speed up research would make formal a practice already in place. Engineers and researchers use Claude most often to fix code errors and learn the codebase 2.
The strategy sets up a feedback loop where a stronger model speeds up the research needed to build the next one.
That matches how teams use Claude Code, Anthropic’s coding assistant, to handle routine data engineering work and troubleshoot infrastructure issues 3.
The likely aim is to build focused tools, much like Claude Code, which Anthropic describes as a simple “Unix utility,” a basic command-line software tool that fits existing developer workflows rather than replacing them 4.