Even a staggering sky-high compensation package of 10.8 billion yuan failed to retain him, Meta has once again lost its top AI executive who received remuneration at the 100-million-yuan level.

Shortly after Muse was released, some people have started to leave Meta's TBD Lab, the team that built it.

The latest leaver is Andrew Tulloch, the candidate with the highest quoted price and the most controversial reputation during Meta's frantic AI talent poaching spree last year.

Before he took the sky-high offer to return to his former employer, he had worked at Meta (formerly Facebook) for more than 11 years in total. After that, he participated in the training of GPT-4o, GPT-4.5 and o3 at OpenAI, and co-founded Thinking Machines Lab with Mira Murati, former CTO of OpenAI.

△

Last August, Mark Zuckerberg and Alexandr Wang of Meta sent him messages one after another, and offered him a compensation package with a total value of 1.5 billion US dollars (about 10.8 billion RMB) to invite Tulloch to return.

But dramatically, Tulloch rejected Meta's sky-high return offer at that time;

It was not until two months later that Tulloch changed his mind, left Thinking Machines Lab which he co-founded, returned to Meta, and joined the TBD Lab led by Alexandr Wang to research superintelligence and be responsible for the R&D work related to Muse.

From returning at the end of October 2025 to the news of his departure in September 2026, Tulloch's second tenure at Meta only lasted about 11 months in total.

And just now, sources from Anthropic revealed that Andrew Tulloch will join Anthropic's Reasoning and Performance Team next week, dedicated to optimizing the training process of Claude Code and improving the model's response efficiency to user instructions. (Specific team details are to be announced.)

△

This can't help but remind people of another "100-million-level AI talent" at Meta — Jiahui Yu.

Also a top talent poached by Meta with huge sums of money, Jiahui Yu announced his departure only eight days after the release of Muse Spark 1.2.

Now less than a month later, Tulloch also chose to leave just two days after Muse went online.

For Meta, this may once again show that talent can be bought with huge sums of money, but may not be retained.

Shortly after Muse was released, core product leaders left one after another

The timing of Tulloch's departure is more thought-provoking than the specific reasons.

On September 8, Meta officially launched its personal AI assistant Muse, which can not only answer questions, but also break down long tasks on its own, coordinate time and resources, help people fill out forms, send emails, book itineraries and complete shopping.

According to Meta's official website, Muse is built on Muse Spark.

Meta proudly calls it the most capable model it has ever built for real agent tasks.

For Mark Zuckerberg, this is not just another AI assistant, but also a key product implementation of his "personal superintelligence" strategy.

And Tulloch didn't decide to officially leave until this product was delivered.

This can't help but remind people of Jiahui Yu, who left Meta to start his own business more than half a month ago. The scenarios of the two joining and leaving the company are surprisingly similar:

In June last year, Meta poached Jiahui Yu from OpenAI with an annual salary of 700 million RMB, and he worked with Mark Zuckerberg, Alexandr Wang and others to build TBD Lab and be responsible for the multimodal direction.

As a result, in August this year, just over a week after the release of Muse Spark 1.2, Jiahui Yu announced his departure to start a business, with a tenure of only one year and more than one month.

Only Jiahui Yu waited for Muse Spark 1.2, while Tulloch waited for the complete Muse.

Now, the product has taken a step forward, but the two core figures have left one after another.

△

The story of Tulloch and Meta dates back to 2012.

He was the most expensive name in Meta's frantic talent poaching spree last year

This is not the first time Tulloch has left Meta.

Andrew Tulloch comes from Australia and graduated from The University of Sydney and University of Cambridge;

From 2012 to 2023, Tulloch worked at Meta for about 11 years, and grew from a machine learning engineer to a Distinguished Engineer.

On Tulloch's personal homepage, he only used one sentence to summarize his early work experience at Meta:

Worked on machine learning systems at Meta, focusing on related R&D work.

But behind this simple sentence, he has been involved in three projects respectively: PyTorch, ResNet-50 and FBGEMM:

• He is an honorary maintainer of PyTorch in the CUDA direction, which shows that he has participated in the work of the underlying framework;
• He participated in the classic work that compressed the training time of ResNet-50 on ImageNet to one hour, achieving about 90% scaling efficiency with 256 GPUs;
• He participated in the construction of FBGEMM, a high-performance underlying operator library for model training and inference, which is especially used for recommendation systems, quantitative computing and generative AI.

In addition, he has participated in projects such as deep learning inference in Facebook data centers, edge-side inference, model quantization and AITemplate.

In other words, Tulloch's most valuable advantage is not designing a chat product, but enabling large-scale models to be trained, run faster and be deployed at lower costs.

In 2023, Tulloch left Meta to join OpenAI, and participated in the training of models such as GPT-4o, GPT-4.5 and o3.

Later, he followed Mira Murati, former CTO of OpenAI, and became the co-founder of Thinking Machines Lab.

In 2025, the performance of Llama 4 failed to meet expectations, and Mark Zuckerberg began to take over Meta's AI strategy in person.

He first invested 14.3 billion US dollars in Scale AI, brought founder Alexandr Wang into Meta, and then established Meta Superintelligence Labs to launch large-scale talent competition against OpenAI, Google, Anthropic and several star start-ups.

Thinking Machines Lab naturally became one of Meta's key targets.

According to reports from The Wall Street Journal, Mark Zuckerberg once contacted Mira Murati, hoping to acquire this newly established AI company;

After being rejected, Meta then contacted more than ten employees of Thinking Machines, trying to poach talents directly to split the team.

Among them, Tulloch was the top target.

Mark Zuckerberg and Alexandr Wang kept sending him messages to invite him to return to Meta.

In addition, in order to get him to leave the company he co-founded, Meta also offered a compensation package worth 1 to 1.5 billion US dollars.

But to Silicon Valley's surprise, not only did Tulloch reject the offer at that time, but none of the employees of the entire Thinking Machines accepted Meta's quotation.

However, this situation ushered in a turning point two months later.

In October 2025, Tulloch announced that he would leave Thinking Machines for personal reasons, return to Meta, and join the TBD Lab led by Alexandr Wang.

Now less than a year later, he has left Meta again.

There is still no authoritative answer to how much compensation Andrew Tulloch has received at Meta, with figures ranging from 1 billion US dollars to 1.5 billion US dollars.

But no matter which number it is, it is enough to make Andrew Tulloch one of the highest-paid non-founding employees in the technology industry.

OpenAI and Anthropic are also poaching each other, with an average tenure of only five months

However, the frequent job changes of top talents are not a unique problem for Meta.

From Roger Ruo Ming Pang, who was once in charge of Apple's foundational model team and later jumped to Meta and OpenAI, to Schulman who left OpenAI for Anthropic...

For these researchers who have already achieved financial freedom and received offers from many labs at the same time, money no longer seems to be the biggest motivation to keep them staying.

△

Axios's summary of this round of talent wars is: