On May 19, Andrej Karpathy, co-founder of OpenAI and former Tesla AI chief, announced on X that he has joined Anthropic. Starting this week, he will lead the team to conduct Claude pre-training research. Anthropic confirmed that pre-training is the foundational training phase that gives Claude core language capabilities, and it is the part of the entire AI model-building process with the highest computational costs.
Pre-Training Responsibilities Confirmed: The Most Expensive Stage for Giving Claude Core Abilities
An Anthropic official statement confirmed that Karpathy is responsible for Claude’s pre-training research—using large-scale corpora to conduct foundational training on the model, which determines its core capabilities in language understanding and reasoning. Karpathy previously combined AI coding agents with small language models in 2024, completing 700 experiments in 2 days and successfully reducing large-model training time by 11%. The industry calls this automated research approach the “Karpathy Loop.”
Karpathy’s Career Path: The Route from OpenAI to Anthropic
Karpathy was one of the co-founders of OpenAI in 2015. In 2017, while Elon Musk was serving simultaneously as a Tesla and OpenAI board member, he recruited Karpathy to Tesla to serve as AI chief and lead the computer vision team for assisted driving. Court documents show that Musk at the time described Karpathy as the world’s second seat in computer vision and said, “The people at OpenAI would want to kill him, but it’s something that has to be done.”
Karpathy left Tesla in 2022, briefly returned to OpenAI, and then founded an AI education startup called Eureka Labs. He is also the proposer of “Vibe Coding”—users give instructions to an AI coding assistant via prompts, without directly writing code. This concept quickly became a common term in the AI development community.
Anthropic’s Personnel Update Confirmed This Month
Ross Nordeen (an early member of xAI) joined Anthropic earlier than Karpathy’s announcement this month. Anthropic also confirmed that Chris Rohlf—who previously worked at Yahoo and Meta and has 20 years of cybersecurity experience—has joined the frontier red team to participate in stress testing. On the same day Nordeen announced he was joining, Anthropic reached an agreement with SpaceX to lease xAI’s compute capacity. Anthropic was co-founded in 2021 by seven former OpenAI employees, including the current CEO Dario Amodei and the current president Daniela Amodei.
FAQ
What role does pre-training play in AI model development, and why is it so costly?
Pre-training is the process of conducting foundational training on an AI model using large-scale corpora. It determines the model’s language understanding, reasoning, and knowledge accumulation. The training process requires massive GPU hashrate to run continuously for weeks to months, making it the most compute-intensive and most expensive part of the entire AI model development pipeline. It is also the prerequisite foundation for subsequent fine-tuning (Fine-tuning) and safety alignment (Alignment) training.
What is the specific concept behind Vibe Coding, and how does Karpathy define it?
Vibe Coding (atmosphere coding) is an AI-assisted development approach proposed by Karpathy—users are mainly responsible for conceptualizing ideas and logic, and instruct an AI coding assistant to generate code through natural-language prompt commands, without writing code themselves. After this concept was introduced, it quickly became a widely adopted term in the AI development community, representing an AI-centered human-machine collaborative programming pattern.
What specific founding connections does Anthropic have with OpenAI?
Anthropic was co-founded in 2021 by seven former OpenAI employees, including Dario Amodei (current CEO) and Daniela Amodei (current president). Since then, additional former OpenAI employees have joined, including John Schulman, who moved to Anthropic in 2024, and this time Karpathy’s joining.
