Andrej Karpathy Joins Anthropic to Lead Claude Pre-training Research

Andrej Karpathy joined Anthropic on 19 May 2026 to lead a new team using Claude to accelerate pre-training research, reporting to Nick Joseph, Anthropic's head of model development, according to TechCrunch and Bloomberg. The hire closes a 22-month chapter in which Karpathy ran Eureka Labs, an AI-native education startup, as an independent founder rather than a frontier model researcher. In a post on X on 19 May, Karpathy wrote: "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."

What Andrej Karpathy built at Eureka Labs, and why he paused it

Karpathy founded Eureka Labs in July 2024, registering the Delaware LLC on 21 June 2024, per VentureBeat. The model was an "AI-native school": real human-designed curriculum delivered through an AI teaching assistant that guides each student individually through the material. The concept drew on Karpathy's history as a Stanford lecturer, but the product was designed to scale without him in the room.

Eureka Labs closed approximately $20 million in seed capital in early 2025, with investors including Conviction's Sarah Guo and Sam Altman. The company's flagship product was LLM101n, an undergraduate-level course guiding learners through building a "Storyteller AI" large language model from scratch in Python, C, and CUDA, per Bloomberg's July 2024 coverage. The GitHub repository for LLM101n accumulated more than 36,000 stars by August 2024 before Karpathy archived it as read-only on 1 August 2024, signalling a shift toward product development over open publication. A second release, "nanochat," arrived in October 2025 as a minimal from-scratch ChatGPT demonstration.

Product output remained limited across the company's 22-month lifespan. In his departure post, Karpathy did not describe Eureka Labs as closed; he said he "remain[s] deeply passionate about education and plan[s] to resume my work on it in time," per TechCrunch.

The year AI made Karpathy feel behind

Two public appearances in early 2026 reveal how Karpathy's working patterns had already shifted before the Anthropic move. On the No Priors podcast, reported by Fortune on 21 March 2026, he described reaching a state he called "AI psychosis": a persistent anxiety about falling behind the frontier of what AI agents could accomplish. "I'm just like in the state of psychosis of trying to figure out what's possible, trying to push it to the limit," he said. The immediate cause was a coding inflection point he traced to December 2025. "I don't think I've typed like a line of code probably since December, basically, which is an extremely large change," he told No Priors. Earlier in 2025 he estimated he was personally writing around 80% of his code, with AI agents handling the remaining 20%; by December that ratio had inverted to roughly 20% human, 80% agent.

Six weeks before the Anthropic announcement, at Sequoia Capital's AI Ascent conference on 30 April 2026, Karpathy formalised those observations into a framework, per his post-event write-up on his bearblog. He drew a line between "vibe coding," the informal practice of delegating code to AI without deeply interrogating the output, and "agentic engineering," which he described as the more rigorous discipline of directing multiple AI agents on complex, high-judgment tasks. In his framing: "Vibe coding raises the floor. Agentic engineering is about extrapolating the ceiling." LLMs, he argued, should be understood as "ghosts: jagged, statistical, summoned entities that require a new kind of taste and judgment to direct." The talk closed with a line that reads now as a direct statement about where he was heading: "You can outsource thinking. You cannot outsource understanding."

What the Anthropic pre-training role involves

The role is not a general research position. According to TechCrunch and CNBC, Karpathy will lead a new team with a specific mandate: using Claude itself to accelerate Anthropic's pre-training research. The approach is deliberately recursive, applying the current best Claude models to the research tasks that will inform the next generation, covering hypothesis generation, experiment design, literature review, and evaluation infrastructure. That maps directly to the "agentic engineering" practice Karpathy described at Sequoia six weeks earlier: directing AI agents on high-judgment cognitive tasks rather than using them as advanced autocomplete.

He reports to Nick Joseph, Anthropic's head of model development. The specific scope of the team has not been disclosed. For Karpathy's career, the role represents a direct return to the pre-training and vision-model work he was doing before his Eureka Labs detour, building on five years of deploying production AI at Tesla's Autopilot programme. Anthropic's Agent SDK infrastructure, released to external developers earlier in 2026, offers a public view of the kind of agent orchestration layer Karpathy's team will likely work within.

What it means

Karpathy has been explicit that the Anthropic move is a deferral of Eureka Labs, not a closure. That framing matters because Eureka Labs is a reasonably funded attempt to build AI-native curriculum infrastructure from scratch, not a prompt-wrapper over existing platforms. Whether it resumes depends on how much of Karpathy's attention the pre-training work requires over the next few years. The more immediate implication is for Anthropic's research velocity: if Karpathy's team can use Claude to compress pre-training research cycles in a measurable way, the company would have a working demonstration of recursive AI improvement under structured human supervision. That outcome would also be the most concrete application yet of the "agentic engineering" framework Karpathy has been developing in public since at least December 2025.

Sources

• TechCrunch: OpenAI co-founder Andrej Karpathy joins Anthropic's pre-training team
• Bloomberg: OpenAI Founding Member Andrej Karpathy Takes Role at Anthropic
• CNBC: Anthropic hires OpenAI co-founder Andrej Karpathy
• Fortune: Karpathy in 'state of psychosis' on AI agents (March 2026)
• Andrej Karpathy: Sequoia AI Ascent 2026 summary
• VentureBeat: Andrej Karpathy announces AI-native school Eureka Labs
• Bloomberg: Karpathy to launch AI education company (July 2024)

Editorial standards: every claim is sourced. Tips: [email protected]