Anthropic hires Andrew Tulloch to optimize its models

Andrew Tulloch, an AI researcher with experience at Meta, OpenAI, and Thinking Machines Lab, will join Anthropic next week to optimize the training and inference of its models. This recruitment comes as Anthropic prepares for its IPO, faces strong competition, and has recently seen the resignation of researcher Jacob Coxon.

Anthropic Prepares for Its IPO Amidst Tension

Anthropic is gearing up to go public while facing intense competition to attract clients and managing growing concerns about the risks associated with AI. This week, researcher Jacob Coxon announced his resignation from Anthropic, stating that some individuals involved in building AI believe it could pose a major danger by the end of the decade. Since the release of ChatGPT, Anthropic, Meta, OpenAI, Google, and other tech companies have been competing to recruit AI researchers. Andrew Tulloch's departure from Meta comes shortly after the launch of Muse, an AI agent designed to perform tasks autonomously for users.

Andrew Tulloch's Journey: Meta, OpenAI, Thinking Machines Lab

Andrew Tulloch worked at Meta for 11 years before his most recent stint, then joined OpenAI in 2023, where he contributed to the training of the GPT-4o, GPT-4.5, and o3 models. After his time at OpenAI, he co-founded Thinking Machines Lab in 2025, having previously worked within that startup. According to the Wall Street Journal, while still at Thinking Machines Lab last year, he declined a salary offer of $1.5 billion from Mark Zuckerberg. In October, he made a brief return to Meta for less than a year. Neither Andrew Tulloch nor Meta immediately responded to requests for comment. Semafor was the first to report his departure from Meta.

Tulloch's Role and Arrival at Anthropic

According to a company spokesperson, Andrew Tulloch is set to leave Meta to join Anthropic next week. He will join the team focused on inference and performance, where he will concentrate his efforts on optimizing the training of Anthropic's AI models as well as enhancing their ability to respond to users.