The plainest reading is that the repo was the resume. On 7 March 2026, Andrej Karpathy tweeted a description of his current side project that would later read, with hindsight, like a position statement: "The goal is to engineer your agents to make the fastest research progress indefinitely and without any of your own involvement... Part code, part sci-fi, and a pinch of psychosis." The 630-line repo he shared, github.com/karpathy/autoresearch, wraps his earlier nanochat project in an agent-orchestration layer where the user spends most of their time editing markdown files that brief an agent fleet running an autonomous research org. Two months later, on 19 May, he posted that he had joined Anthropic.

Anthropic's framing of the new effort, surfaced by reporters at TechCrunch, Axios, and Reuters, is that Karpathy will help launch a new team focused on using Claude itself to accelerate pre-training research, reporting to Nick Joseph (also ex-OpenAI). Reading that sentence next to his March description is uncomfortable. Strip the corporate phrasing away and you have someone whose personal Markdown-driven autoresearch repo has been hired to do, indoors, what it was already doing in public.

The companion line in Karpathy's GitHub README, about programming program.md files that "provide context to the AI agents and set up your autonomous research org," is now the spec for a real internal team. Anthropic does not need to imagine what this looks like operationally; the prototype has been on GitHub for two months and has more than eighty thousand stars. What the company is paying for is the shift from a hobbyist running ten parallel agents on his own GPU to a production deployment with Claude as the engine and pre-training as the loss to minimise.

That is recursive self-improvement in a respectable suit. I wrote about RSI when GPT-5.3-Codex helped debug its own training; calling Karpathy's new job "pre-training acceleration" is technically accurate and almost designed to sound less alarming than calling it what it is. The loop has a name now and a payroll number.

His X post mentioned that he remains "deeply passionate about education" and plans to "resume my work on it in time." Eureka Labs is not being shut down, only paused. But the timeline tells you where his attention had drifted: the autoresearch repo went up in March, and Eureka had been running for two years before that without producing the comparable artefact. The new job is a continuation, not a pivot.

Anthropic also announced a second hire that same Tuesday: Chris Rohlf to the frontier red team, the group that stress-tests advanced models against severe threats. Pairing a senior pre-training hire with a senior red-team hire on the same day is choreography for the question "are you accelerating safely?", and days later METR's first Frontier Risk Report gave that question its answer. The report found that internal AI agents at Anthropic, DeepMind, Meta, and OpenAI already plausibly had the means, motive, and opportunity to start small rogue deployments. The hires landed during the week that finding was being absorbed.

If I had to bet on the next public artefact, it is not a paper. It is a Claude-generated change to the next pre-training run, one that survives review and ships, whose provenance nobody outside the building can quite reconstruct.

Sources: