← All people

Brief profile

Watercolor portrait of Andrew Tulloch

Andrew Tulloch

Former researcher, TBD Lab, Meta

Training-systems engineer who helped train GPT-4o and o3 at OpenAI, then co-founded Thinking Machines Lab.

Andrew Tulloch grew up in Perth, Australia, and studied mathematics at the University of Sydney and the University of Cambridge before spending about 11 years at Facebook, later Meta, working on machine learning systems. Tulloch's name appears on "Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour" (June 2017), the Facebook AI Research paper that trained a ResNet-50 image classifier on 256 GPUs in an hour by scaling the learning rate with the batch size, and on Meta's 2021 paper on the ZionEX system for training recommendation models across many machines.

A move to OpenAI followed in 2023, where by Tulloch's own account the work included training GPT-4o, GPT-4.5 and o3. In February 2025 Tulloch was among the co-founders of Mira Murati's Thinking Machines Lab. After Murati turned down an offer from Mark Zuckerberg to buy the company, the Wall Street Journal reported in August 2025, Meta approached individual staff, and the package offered to Tulloch could have reached 1.5 billion dollars over at least six years with bonuses and stock gains; Meta spokesperson Andy Stone called that description "inaccurate and ridiculous." The offer was turned down at first. In October 2025 Tulloch told Thinking Machines staff of a departure, which a company spokesperson explained as a decision "to pursue a different path for personal reasons," and moved to TBD Lab, the research group inside Meta Superintelligence Labs led by Alexandr Wang, describing the work as "working on superintelligence."

On 9 September 2026 Semafor reported that Tulloch was leaving Meta, having waited until the day after Meta launched its Muse models and assistant. The next day the Wall Street Journal reported that Tulloch would join Anthropic's inference and performance team. Neither company had confirmed the move when this profile was checked.

Known for

Training ImageNet in one hour (2017)

Co-author of the Facebook paper showing that a linear learning-rate scaling rule and warmup let ResNet-50 train with 8,192-image minibatches on 256 GPUs without losing accuracy.

Read the original

Training systems for OpenAI's 2024 and 2025 models

Worked on the training of GPT-4o, GPT-4.5 and the o3 reasoning model before leaving to co-found Thinking Machines Lab.

Read the original

Career

  1. until 2023

    Meta

    Engineer on machine learning systems

  2. 2023-2024

    OpenAI

    Member of technical staff; worked on training GPT-4o, GPT-4.5 and o3

  3. 2025

  4. 2025-2026

    Meta

    Researcher, TBD Lab in Meta Superintelligence Labs

Sources

Last verified 2026-09-25.

  1. Andrew Tulloch, personal website (checked September 2026)
  2. Goyal et al., Accurate, Large Minibatch SGD, Training ImageNet in 1 Hour (arXiv, June 2017)
  3. Mudigere et al., Software-Hardware Co-design for Fast and Scalable Training of Deep Learning Recommendation Models (arXiv, April 2021)
  4. Andrew Tulloch leaves Thinking Machines Lab for Meta (Dataconomy, October 2025)
  5. AI researcher Andrew Tulloch is leaving Meta (Semafor, September 2026)
  6. Meta's star AI hire Andrew Tulloch is leaving for Anthropic (The Next Web, September 2026, archived)
  7. Andrew Tulloch, poster child for out-of-control AI spending, reportedly quits Meta (Gizmodo, September 2026)

AI-generated watercolor interpretation based on a reference photograph. Photo reference: IQ.wiki.

Brief profiles cover a person's professional record: roles, work and public statements, each from a source listed here. Full profiles, with positions across the debates, are kept for the people the site follows most closely.