Andrej Karpathy, co-founder of OpenAI and former director of AI at Tesla, announced Tuesday that he has joined Anthropic. He will work on the company’s pre-training team under lead Nick Joseph, building a group focused on using Claude itself to accelerate pre-training research.
“I think the next few years at the frontier of LLMs will be especially formative,” Karpathy wrote on X. “I am very excited to join the team here and get back to R&D.”
Why Pre-Training Matters for Agents
Pre-training is the large-scale foundational training phase that gives Claude its core knowledge and reasoning capabilities, according to Anthropic. It is also one of the most expensive and compute-intensive stages of building a frontier model. For agent builders, this is the layer that sets the ceiling on autonomous reasoning, tool use, and multi-step planning. Improvements here propagate downstream to every Claude-powered agent in production.
Karpathy’s specific mandate, building a team that uses Claude to speed up its own pre-training research, points to a recursive improvement loop: better models helping train even better models. An Anthropic spokesperson told TechCrunch that Karpathy is “one of the few researchers who can bridge the gap between LLM theory and large-scale training practice.”
A Month of Aggressive Talent Moves
Karpathy is the latest in a string of high-profile hires. Ross Nordeen, a founding member of Elon Musk’s xAI and former Tesla employee, announced earlier this month that he had joined Anthropic, according to CNBC. That announcement came the same day Anthropic struck a deal with SpaceX to lease the entire Colossus 1 data center for Claude inference capacity.
Separately, Chris Rohlf joined Anthropic’s frontier red team this week. Rohlf brings over 20 years of cybersecurity experience from Yahoo’s Paranoids security team and Meta, plus a fellowship at Georgetown’s Center for Security and Emerging Technology. “We have a real opportunity in front of us to dramatically improve cyber security with AI,” Rohlf wrote on X.
The pattern is clear: top researcher talent (Karpathy), compute infrastructure (Colossus 1), cybersecurity expertise (Rohlf), and enterprise customer revenue ($300 million in Anthropic token spend from Salesforce alone). Anthropic is building vertically across every layer needed to deploy reliable autonomous agents at scale.
Karpathy’s Path to Anthropic
Karpathy helped found OpenAI before leaving in 2017 to lead Tesla’s Autopilot computer vision team. Elon Musk personally recruited him, describing Karpathy in an email as “arguably the #2 guy in the world in computer vision” behind Ilya Sutskever, according to trial exhibits reported by CNBC. That email surfaced during the Musk v. Altman trial, which concluded Monday with the jury ruling in OpenAI CEO Sam Altman’s favor.
Karpathy left Tesla in 2022, returned briefly to OpenAI, then founded Eureka Labs, an AI education startup. He also maintains a widely followed YouTube channel and the “Neural Networks: Zero to Hero” course. He holds a PhD in computer science from Stanford.
“I remain deeply passionate about education and plan to resume my work on it in time,” Karpathy said in his announcement.
The Compute Bet vs. The Talent Bet
Anthropic’s hiring spree reflects a strategic thesis: at the frontier, the quality of researchers matters as much as the quantity of GPUs. OpenAI has Microsoft’s compute backing. Google has its own TPU infrastructure. Anthropic is betting that assembling the right people, and giving them tools that accelerate their own work, is how a smaller company stays competitive in the race to build models that can reliably power autonomous agents.