Andrew Tulloch is leaving Meta less than a year after rejoining the company and is expected to join Anthropic’s inference and performance team next week, according to the source material. There, he will work on optimizing the training process for Claude Code and improving how efficiently models respond to user instructions. Specific team details have not been disclosed.
The timing stands out. Meta launched its personal AI assistant Muse on Sept. 8, and Tulloch, who had been involved in Muse-related development work inside Meta’s TBD Lab, is now set to exit just two days after the product went live.
A second stint at Meta that lasted about 11 months
Tulloch was one of the most expensive and most debated targets in Meta’s AI recruiting campaign last year. The article says Meta CEO Mark Zuckerberg and Alexandr Wang repeatedly messaged him in August 2025 and offered a compensation package valued at as much as $1.5 billion, or about 10.8 billion yuan, to bring him back. Tulloch turned that offer down at the time.
Two months later, he changed course. In October 2025, he left Thinking Machines Lab, the startup he had helped found, and returned to Meta. He joined the TBD Lab led by Wang, where he worked on superintelligence research and on development related to Muse.
From late October 2025 to reports of his departure in September 2026, Tulloch’s second term at Meta lasted roughly 11 months.
Meta described Muse on its website as a personal AI assistant that can answer questions, break down long tasks, coordinate time and resources, fill out forms, send emails, book travel and complete shopping tasks. The company said Muse is powered by Muse Spark, which it called its strongest model so far for real-world agent tasks.
For Zuckerberg, the product was presented as a key step in Meta’s “personal superintelligence” strategy. Tulloch’s decision to leave came only after that product had shipped.
Another departure after a major Meta AI release
The article links Tulloch’s exit with another recent departure inside Meta’s AI efforts. Jiahui Yu, also recruited to Meta with a rich compensation package, left eight days after the release of Muse Spark 1.2. Less than a month later, Tulloch is now leaving two days after the launch of Muse.
The similarities are notable. Meta brought Yu over from OpenAI in June 2025 with annual compensation of 700 million yuan. He joined Zuckerberg, Wang and others in building the TBD Lab and took charge of multimodal work. In August 2026, a little more than a week after Muse Spark 1.2 was released, Yu announced he was leaving to start a company of his own, ending a stint that lasted about one year and a little over one month.
Yu stayed through the release of Muse Spark 1.2. Tulloch stayed through the launch of the full Muse product. In both cases, the product moved ahead and a key figure left soon after.
Tulloch’s path from Meta to OpenAI and Thinking Machines Lab
Tulloch is from Australia and studied at the University of Sydney and the University of Cambridge. He worked at Meta from 2012 to 2023, a period of about 11 years, and rose from machine learning engineer to Distinguished Engineer.
On his personal page, he summarized that first Meta stretch in a single line: 「Worked on machine learning systems at Meta.」
Behind that short description sits a long list of infrastructure work. The article says he was an honorary maintainer for PyTorch CUDA, indicating involvement in low-level framework engineering. It also says he contributed to the well-known effort that cut ResNet-50 training time on ImageNet to one hour, using 256 GPUs and reaching about 90% scaling efficiency.
He also worked on FBGEMM, a high-performance operator library used in training and inference, especially for recommendation systems, quantized computing and generative AI. Beyond that, the article says he contributed to deep learning inference in Facebook data centers, on-device inference, model quantization and AITemplate.
The piece argues that Tulloch’s scarcity value was not in designing a chat product. It was in making very large models trainable, faster to run and cheaper to deploy.
In 2023, he left Meta for OpenAI, where he worked on training GPT-4o, GPT-4.5 and o3. He later followed former OpenAI CTO Mira Murati to become a co-founder of Thinking Machines Lab.
Meta targeted Thinking Machines Lab in its AI hiring push
According to the article, Zuckerberg took a more direct role in Meta’s AI strategy after Llama 4 fell short of expectations in 2025. Meta then invested $14.3 billion in Scale AI and brought founder Alexandr Wang into the company. After that, it assembled Meta Superintelligence Labs and launched a broad recruiting push aimed at OpenAI, Google, Anthropic and several prominent startups.
Thinking Machines Lab became one of the main targets. The article cites The Wall Street Journal as saying Zuckerberg approached Mira Murati about acquiring the newly formed AI company. After she declined, Meta contacted more than 10 employees at Thinking Machines Lab in an effort to hire them directly.
Tulloch was described as the top target. Zuckerberg and Wang kept messaging him to return. Meta also put forward a package reportedly in the $1 billion to $1.5 billion range to persuade him to leave a company he had helped build.
At first, that effort failed. The article says Tulloch rejected the initial approach, and no one at Thinking Machines Lab accepted Meta’s offers at that time. In October 2025, he left the startup for personal reasons and returned to Meta’s TBD Lab.
There is still no authoritative public answer on how much he ultimately received from Meta. The article notes that figures of $1 billion and $1.5 billion have both circulated. Either way, it says, the amount would place him among the highest-paid non-founder employees in the tech industry.
The talent war extends beyond Meta
The piece also stresses that constant movement among top AI researchers is not unique to Meta. It points to other examples, including Pang Ruoming, who worked on Apple’s foundation models team before moving to Meta and OpenAI, and Schulman, who moved from OpenAI to Anthropic.
For researchers who already have financial freedom and hold invitations from multiple labs, money may no longer be the main reason to stay. Axios summarized the current talent war this way: winning top AI talent is already difficult, and retaining that talent may be just as hard even with huge pay packages, prestigious titles and abundant computing power.
The article adds that, for these researchers, choosing a lab may no longer be only about choosing a job. It may also mean choosing the version of the AI future they believe is most likely to be realized.
How long the bidding war lasts may depend on scarcity
The final section asks how long this high-cost competition for AI talent can continue. Its answer is tied to how long elite researchers remain scarce. As long as companies feel they need to buy access to rare, hard-won experience inside the heads of star researchers, the market for that talent is unlikely to cool.
The article argues that the environment could change only when companies stop relying on one-off expertise from celebrity researchers and instead need engineering capability that can be reproduced systematically across an organization.
Until then, top AI researchers may continue to resemble free agents in professional sports, choosing among large tech companies and startups based on pay, compute, technical direction, teammates and stage.
For Meta, the back-to-back departures of Jiahui Yu and Andrew Tulloch offer another clear example of the problem: a company may be able to buy talent at a high price, but not necessarily keep it.