On May 19, 2026, Andrej Karpathy posted on X that he had joined Anthropic. Short message. No fanfare. “I think the next few years at the frontier of LLMs will be especially formative,” he wrote. “I am very excited to join the team here and get back to R&D.”
Within the hour, the AI world had erupted with opinions. That kind of reaction does not happen for most hires. It happened for this one.
Who is Andrej Karpathy?
Andrej Karpathy was part of the founding team at OpenAI back in 2015, when it was still a strange, idealistic experiment rather than the most recognised AI brand in the world. He left in 2017 to join Tesla, where he ran the Autopilot and Full Self-Driving programme for five years.
That Tesla chapter is worth pausing on. This was not academic work. It was not benchmarks in a controlled lab. It was getting neural networks to work reliably, on real roads, inside real cars, at a scale most researchers never touch. He did that for five years and left in 2022.
He returned to OpenAI briefly, then left again in 2024 to start Eureka Labs — a project built around one idea he has held for years: that AI could fundamentally change how people learn. His YouTube series Neural Networks: Zero to Hero became a go-to resource for a generation of engineers trying to understand how these models actually work, not just what they produce.
Now Eureka Labs sits in uncertain territory. He has not said much about its future since the announcement.
What He Will Actually Do at Anthropic
Karpathy has joined Anthropic’s pre-training team, working under team lead Nick Joseph. He also plans to help build a new group with a specific focus: using Claude itself to speed up pre-training research.
That second part deserves more attention than it has received.
Pre-training is where a model’s character is actually formed. It is the phase where massive training jobs run on enormous datasets, the process through which a model learns language, reasoning patterns, and the general structure of human knowledge. It is extraordinarily expensive in compute and time. Small decisions made here ripple outward into every product and application built on top of that model.
The idea Anthropic is pursuing, using an existing model to assist in its own pre-training research, is genuinely ambitious. AI-assisted AI research. It sounds circular until you consider what a capable model can actually do when pointed at a hard research problem. Whether it works at meaningful scale is an open question. But the fact that Anthropic is making this bet, and putting someone with Karpathy’s background at the front of it, says something real about the company’s direction.
What This Costs OpenAI
The financial cost of losing Karpathy to a competitor is easy to calculate. The symbolic cost is harder.
He was there at the founding. He was in the room when OpenAI’s culture was still being shaped, when the mission of building powerful AI safely was still a genuine aspiration and not a press release. The lab he helped create has since become the most recognised name in consumer AI. ChatGPT. The company that arguably set off the current era.
And now one of its founding members has walked across to its primary rival.
OpenAI is not in crisis over this. The lab has enormous resources and product dominance that one hire cannot dent. But in a field where reputation and perceived momentum matter, where talent tends to follow talent, this signals something. Researchers notice where people like Karpathy choose to go.
The Bigger Picture
Karpathy’s arrival did not come alone. On the same day, Anthropic also announced that Chris Rohlf, over 20 years in cybersecurity, including stints at Yahoo and six years at Meta, had joined its frontier red team, the group responsible for stress-testing models against serious threats.
Two hires. One on the capability side. One on safety and adversarial testing. That combination maps cleanly onto how Anthropic has always described itself: a company that wants to build powerful models and take seriously what those models can do when things go wrong.
Why It Matters Beyond Silicon Valley
The models being built at Anthropic, OpenAI, and Google DeepMind are already inside tools used by developers, students, and professionals every day. Decisions made in San Francisco research labs move fast.
Karpathy’s Neural Networks: Zero to Hero series has a substantial following among engineering students who have no other access to this quality of instruction. His eventual return to education work, which he has said he plans, will matter in ways that may not register at industry conferences but will be felt by real people learning real things.
For now, he is at Anthropic. Working on pre-training. Trying to use AI to make AI better. It is either a very clever idea or a very difficult one. Probably both.
The next year or two will show which.
