Andrej Karpathy Joins Anthropic to Lead Pretraining Research

Andrej Karpathy, one of the most recognisable names in deep learning, has joined Anthropic to work on pretraining, the company confirmed last week. He started this week reporting to pretraining lead Nick Joseph, and will set up a new team focused on using Claude to accelerate the company's pretraining research, according to TechCrunch.

Karpathy is a co-founder of OpenAI, where he led deep learning and computer vision research from 2015 to 2017. He left to become director of artificial intelligence at Tesla, where he ran the Autopilot and Full Self-Driving programmes for five years, then returned briefly to OpenAI before founding the education startup Eureka Labs in 2024. CNBC and Axios describe the move as one of the most significant individual transfers between frontier labs this year.

Why pretraining

Pretraining is the part of model development that turns raw data and compute into the broad knowledge and capabilities that show up in every later interaction with a model. It is also the most expensive and least glamorous phase, with months of training runs on tens of thousands of chips and results that are difficult to predict in advance. Anthropic told reporters Karpathy's team will look for ways Claude itself can shorten that cycle, whether by generating training data, helping interpret runs, or proposing experiments.

That last point matters. Several labs have begun describing a feedback loop in which current models are used to design the next generation, and the speed of that loop is becoming a competitive variable in its own right. Anthropic's co-founder Jack Clark has spoken publicly about expecting Nobel-grade research within a year, a forecast that depends on exactly the kind of acceleration Karpathy's team is being set up to chase.

Talent gravity

The hire follows a string of senior researcher moves toward Anthropic over the past eighteen months. With the company's funding round at a $900 billion-plus valuation set to close imminently, observers in the industry expect more departures from OpenAI and Google DeepMind to follow. Karpathy himself remarked on his personal channels that he sees pretraining research as the most leveraged thing he can spend his time on right now.

Coverage from The New Stack and Proactive Investors notes the move will be closely watched as a signal of where elite research talent expects the next gains to come from.