On May 19 at 1:30 PM, Andrej Karpathy posted on X: "Personal update: I have joined Anthropic." Within an hour, the post garnered nearly 3 million views—impressive for someone with only 2 million followers.
Not just any defector: an OpenAI founding member
Karpathy was a founding member of OpenAI in 2015. He later served as Tesla's AI director, leading Autopilot development. He returned to OpenAI in 2023 for over a year before leaving to start his own AI education company, Eureka Labs. Despite his comfortable position, he said in the post that "the frontier of large models will be especially important in the coming years" and he was "eager to get back to research," putting education on hold.
His role is more significant than the title suggests
Karpathy is not joining as an advisor. He will work on Anthropic's pre-training team, reporting to Nick Joseph. Pre-training is the most capital-intensive and uncertain phase of large model development, determining the model's ceiling. Anthropic's direction, per the official statement, is to "use Claude to accelerate pre-training research." In plain terms: using AI to help tune how AI trains. This approach had a prototype in March, when an internal Anthropic AI agent ran 700 experiments in two days, discovered 20 self-optimizations, and cut training time by 11%. The industry nicknamed it the "Karpathy Loop." At the time, Karpathy was still at Eureka Labs; now the namesake has arrived.
The inventor of 'vibe coding' joins the company that wants to bury the term
In February 2025, Karpathy coined the term "vibe coding"—using natural language to let AI write code. Collins Dictionary named it 2025 Word of the Year. But on Tuesday (May 20) at London's Code with Claude event, Claude Code head Boris Cherny publicly criticized the term, saying it "increasingly annoys me" because it "implies sloppiness and lack of rigor," while Anthropic's work is more systematic. He proposed "agentic engineering" instead.
"The default is not 'I'll prompt Claude' but 'let Claude prompt itself.'" — Boris Cherny
The inventor of vibe coding is now helping the company that wants to move beyond it, working on pre-training. A dramatic twist.
What this means for OpenAI
OpenAI has seen many departures, but few have gone directly to a rival for core work. Karpathy is not a strategic advisor; he is embedded in core research. If his direction—"let Claude help train Claude"—succeeds, it could be a key differentiator for Anthropic against OpenAI. Where and what he works on speaks louder than any corporate announcement. Whether this path works will be clear in 6 to 12 months.
Sources: OpenAI co-founder Andrej Karpathy joins Anthropic's pre-training team (TechCrunch); CocoLoop; Andrej Karpathy, OpenAI founding member and inventor of 'vibe coding,' defects to Anthropic (Fortune); Anthropic hires OpenAI co-founder Andrej Karpathy, former Tesla AI leader (CNBC); OpenAI co-founder Andrej Karpathy announces he's joining Anthropic (VentureBeat)