2026-05-22 — views

Andrej Karpathy joins Anthropic — to use Claude to accelerate Claude's pretraining

Read this because The mandate is the story, not the hire. "Use Claude to accelerate Claude's pretraining" is recursive self-improvement as a job description — the loop Recursive Superintelligence raised $650M to chase, now staffed inside a frontier lab.

Andrej Karpathy joined Anthropic's pretraining team under Nick Joseph (May 19). His mandate: use Claude to accelerate Claude's own pretraining R&D.

Andrej Karpathy — OpenAI founding member, ex-director of Tesla Autopilot AI, founder of Eureka Labs — started this week at Anthropic (announced May 19), joining the pretraining team under lead Nick Joseph.

The mandate is the headline

Per Anthropic, Karpathy will start a team focused on using Claude to accelerate pretraining research. Pretraining is the large-scale training that gives Claude its core knowledge and capabilities.

Read that mandate literally: use Claude to make the next Claude’s pretraining better and faster. That’s recursive self-improvement stated as a job description — the AI improving the process that builds the AI.

It’s the same loop the Recursive Superintelligence lab raised $650M to chase. The difference: that’s a 30-person startup betting on it; this is a frontier lab staffing it with one of the field’s most recognized researchers, inside an existing $350B+ pretraining operation.

Why Anthropic, why now

Talent magnet signal. Karpathy choosing Anthropic over staying independent (Eureka Labs) or returning to OpenAI is a coup in the elite-talent war — and another data point that Anthropic is where senior researchers want to be post-Claude-Opus-4.7.
He framed it as “getting back to R&D.” In his X announcement: “I am very excited to join the team here and get back to R&D” — while noting he’ll “resume my work on education in time.” A research re-entry, not a management role.
Pretraining is the moat layer. Post-training (RLHF, agents) gets the product headlines, but pretraining is where raw capability is set. Putting a marquee researcher there signals Anthropic believes the next capability jump comes from pretraining efficiency, not just post-training polish.

Why it matters

“AI accelerates AI research” is now a staffed function, not a thesis. When a frontier lab assigns a named team to “use our model to improve our model’s training,” the recursive-improvement timeline stops being speculative and starts being a quarterly OKR.
The hire validates the test-time-compute/AutoTTS direction — the broader theme of AI optimizing its own pipeline, from inference orchestration up to the pretraining run itself.

Practitioner note

Watch for pretraining-efficiency papers from Anthropic in 2-3 quarters. If Karpathy’s team ships, the signal will be a step-change in Claude’s training compute efficiency or capability-per-FLOP — the kind of thing that shows up in a model card, not a press release.
Talent flow is a leading indicator. Where senior researchers concentrate predicts which lab leads in 18-24 months. Karpathy → Anthropic is one more vector in the same direction as the post-4.7 momentum.
Don’t over-index on one hire. A single researcher, however prominent, doesn’t change a model roadmap overnight. The signal is directional (Anthropic’s pull + the recursive-improvement framing), not an immediate capability claim.

The under-considered angle: the recursive-self-improvement story has quietly moved from manifestos to org charts. A year ago “AI improving AI” was a thesis VCs funded (SSI-style); now it’s a named team with a named lead inside a top-3 lab. The question is no longer whether labs will try to close the loop — it’s which one ships a measurable pretraining-efficiency gain from it first.