“Personal update: I’ve joined Anthropic,” Andrej Karpathy wrote on X on Tuesday, in a post that drew nearly 3 million views within one hour. He said the next few years at the frontier of LLMs would be “especially formative” and that he was eager to get back to research. He started this week on the pretraining team.

The decision suggests that Karpathy, whose writing on AI is followed by nearly 2 million people on X, has decided to put his stake in the AI race. He was first a founding member of OpenAI in 2015, left to run AI at Tesla, came back in 2023, and left only a year later to start his own education company, Eureka Labs.

Karpathy has been well-known in AI for a decade, but the thing—or phrase—that etched him into AI legend was a tweet from last year. In February 2025, he posted that there was a “new kind of coding” that he called vibe coding: Describe what you want plainly and let the model do the work. 

The phrase escaped the industry and infected the business world, which turned its back against software companies and raced to develop its own bespoke agents, touching off the much-debated “SaaSpocalypse” in its wake. (What wasn’t debated is that tens of billions of dollars in stock valuations evaporated as firms tried to vibe code their own solutions.) Collins Dictionary named it Word of the Year. The model he cited in that original tweet was Anthropic’s.

As an Anthropic employee, Karpathy will be building on the work from another viral post. In March, he wired up an AI coding agent, handed it a single small language model, and let it run unsupervised for two days, testing and tweaking the training code on its own. After 700 experiments and 20 self-found optimizations, he said the same tweaks applied to a larger model cut training time by 11%. This, he said, was called autoresearch: “Part code, part sci-fi, and a pinch of psychosis” he wrote with a smiley face. The method would come to be known as “the Karpathy Loop.”

Teaching that method looks, more or less, like his new job. According to Anthropic, Karpathy will be starting a team focused on using Claude to accelerate pretraining research, the large-scale training runs that give Claude its core knowledge and capabilities. He sits on a team led by Nick Joseph. 

Long before any of this, though, Karpathy was famous for something else: his Rubik’s Cube skills. He ran a YouTube channel called “badmephisto,” where a generation of competitive cubers learned to “speedcube” by seeing the cube as 26 individual “cubies” rather than 54 colored stickers. By sticking to the small structure, he could move the whole thing. He solved a Rubik’s Cube in about 17 seconds.

Certainly the puzzles for Karpathy got harder—neural nets and then language models—but the method never really changed. Get a small enough system fully under control, and you can move something much bigger.