Summary
On May 19, 2026, Andrej Karpathy announced via X that he had joined Anthropic's pre-training team. Karpathy co-founded OpenAI in 2015, later directed Tesla's Full Self-Driving and Autopilot programs, and most recently ran Eureka Labs, an AI education startup. At Anthropic, he will lead a new team focused on using Claude to accelerate pre-training research, working under Nick Joseph who leads pre-training. The announcement followed by one day Anthropic's acquisition of Stainless.
What Happened
Andrej Karpathy's career trajectory prior to joining Anthropic spans the core institutions and applications of the deep learning era. He was among the seven co-founders of OpenAI at its December 2015 founding, working on deep learning and computer vision research. He left in 2017 to join Tesla, where he built and led the Autopilot and Full Self-Driving teams until 2022. He rejoined OpenAI briefly in 2023, then departed again in early 2024 to found Eureka Labs, focused on applying AI to education. He also became widely known for his YouTube teaching channel on neural networks and for popularizing the term "vibe coding" to describe AI-assisted software development.
Karpathy announced the move on X on May 19: "Personal update: 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. I remain deeply passionate about education and plan to resume my work on it in time." The post drew approximately three million views within one hour.
An Anthropic spokesperson told TechCrunch that Karpathy will start a team specifically focused on using Claude to accelerate Anthropic's own pre-training research — a meta application in which the model assists in improving the processes that train its successors. He works within Nick Joseph's pre-training organization.
Why It Matters
Pre-training is the foundational phase of large language model development — the large-scale training runs on broad data that determine a model's core knowledge, reasoning patterns, and capabilities before any fine-tuning or alignment work. Control of pre-training quality is widely understood within AI labs as the primary determinant of long-run model capability, and it has historically been among the least publicly described aspects of frontier lab operations.
Karpathy's hire is notable partly for his specific career context. He co-founded the company that is Anthropic's closest competitor, worked on the applied AI application (autonomous driving) that required the most industrial-scale machine learning outside language models, and spent time running an education startup before joining a frontier lab's core research function. His stated reason for the move — that "the next few years at the frontier of LLMs will be especially formative" — is a public signal from a technically credible observer about where he believes high-leverage research will occur.
His stated mandate — using Claude to accelerate Anthropic's pre-training research — represents a specific research direction: applying current-generation models to improve the pipelines that produce next-generation ones. This direction has been pursued internally by multiple frontier labs without detailed public disclosure; Karpathy's hire and stated focus makes it visible.
What is not known: the scope or timeline of any team he builds, what specific pre-training bottlenecks he is addressing, or how his education work at Eureka Labs intersects with his Anthropic mandate. His prior pattern of leaving organizations makes longevity at any given employer an open question.
§ How to read the metadata
- Landmark
- Fundamentally alters the trajectory; 2–5 per year.
- Major
- Meaningfully shifts the landscape; 2–4 per month.
- Notable
- Worth documenting; significance can be upgraded later.
- Confidence
- High = primary sources corroborate. Medium = credible secondary only. Low = provisional. Disputed = credible sources disagree.
- Contestation
- Uncontested = no formal challenge. Contested = at least one challenge open. Superseded = replaced by a later entry. Unresolved = dispute still open.