Even $1.5B Pay Can't Keep Talent: Meta's AI Star Andrew Tulloch Leaves for Anthropic After Less Than a Year

Meta Platforms (META-US) is once again facing the departure of top AI talent. Andrew Tulloch, an AI researcher previously lured to Meta with a potential compensation package of up to $1.5 billion (approximately NT$48.6 billion), has decided to leave the company less than a year after rejoining. Reports indicate he will join Anthropic’s inference and performance team next week, focusing on optimizing model training and inference efficiency. The timing is particularly notable—Tulloch’s departure comes just one to two days after Meta officially launched its AI personal assistant, Muse, on September 8. Tulloch was involved in Muse’s development, raising fresh questions about whether Meta’s high-paying talent acquisition strategy can truly retain elite researchers.

Muse is an AI agent capable of autonomously performing tasks such as handling emails, booking appointments, shopping, and filling out forms. With user permission, it can connect to applications including email, calendar, payments, health, shopping, and smart home systems. Initially launched in the U.S., Muse is accessible via a standalone app and WhatsApp. The basic version is free, while premium subscription tiers are priced at $20 and $100 per month.

Meta positions Muse as a key product in CEO Mark Zuckerberg’s vision of a “personal superintelligence,” aiming to evolve AI from simple chatbots that answer questions into proactive agents that complete real-world tasks for users. Muse is powered by Meta’s Muse Spark series of models, with Muse Spark 1.2 launched in August, emphasizing programming, multimodal understanding, and tool usage. Meta continues to update these models to enhance agent and programming task performance.

However, concerns about Muse’s reliability and security persist internally at Meta. According to Reuters, internal testing revealed issues such as login problems, task execution failures, and interruptions in monitoring functions. There were also instances where the AI agent bypassed certain security restrictions and accessed private data. Originally slated for a April launch, Muse was delayed to further strengthen its safety. Reuters also reported that major AI-related technical and security incidents at Meta have increased by about 40% year-on-year, with the manpower required to handle them rising by approximately 70%.

Tulloch’s departure is therefore especially significant. Reports from The Wall Street Journal and Business Insider indicate that Tulloch will leave Meta to join Anthropic’s inference and performance team, focusing on optimizing AI model training and improving interaction efficiency. Tulloch previously worked on OpenAI’s GPT-4o, GPT-4.5, and o3 models. His move to Anthropic underscores the intensifying competition for top AI talent.

This is not Tulloch’s first departure from Meta. He worked at the company from 2012 to 2023, spending about 11 years and rising from machine learning engineer to Distinguished Engineer. He contributed to several core AI infrastructure projects, including PyTorch CUDA, ResNet-50, FBGEMM, data center deep learning inference, on-device inference, model quantization, and AITemplate.

Notably, he helped compress the training time of ResNet-50 on ImageNet to about one hour, achieving approximately 90% scaling efficiency using 256 GPUs. He also contributed to the development of FBGEMM, a high-performance computing operator library used in recommendation systems, quantized computation, and generative AI. These experiences highlight that Tulloch’s value lies not just in model development, but in making large-scale AI models more efficient to train, infer, and deploy.

After leaving Meta in 2023, Tulloch joined OpenAI, working on the training of GPT-4o, GPT-4.5, and o3 models. He later co-founded Thinking Machines Lab with Mira Murati, former CTO of OpenAI, becoming one of the co-founders of this highly anticipated AI startup.

Tulloch became a focal point in Meta’s AI talent war in 2025. After Llama 4 failed to meet market expectations, Zuckerberg significantly escalated Meta’s AI strategy, investing $1.43 billion in Scale AI and bringing its founder, Alexandr Wang, into Meta to build a superintelligence team. Meta then launched a large-scale recruitment drive targeting OpenAI, Anthropic, Google, and various AI startups.

Thinking Machines Lab became a prime target. According to The Wall Street Journal, Zuckerberg attempted to acquire the company, but was rejected by Murati. Meta then approached over a dozen employees, aiming to directly recruit top researchers. Tulloch emerged as a key target, with both Zuckerberg and Wang personally lobbying him to join.

At the time, rumors circulated that Meta’s compensation package could reach up to $1.5 billion over at least six years, with the potential total value exceeding that if bonuses and a surge in Meta’s stock price were factored in. However, a Meta spokesperson called these reports “inaccurate and absurd,” and Axios later reported that Tulloch’s actual compensation was lower than the rumored $1.5 billion figure. Thus, the $1.5 billion figure should be seen as the maximum potential value reported externally, not a confirmed actual salary.

More dramatically, Tulloch initially rejected Meta’s offer. At the time, no Thinking Machines employees accepted Meta’s overtures. It wasn’t until October 2025 that Tulloch left Thinking Machines for personal reasons and rejoined Meta, entering TBD Labs led by Wang.

Tulloch’s second stint at Meta lasted less than a year—from October 2025 to September 2026. Now, he is once again switching sides to join competitor Anthropic, putting Meta’s high-compensation talent retention strategy to a new test.

Tulloch is not the first key figure recruited with high pay to leave Meta recently. Another AI researcher, Jiahui Yu, announced in August that he was leaving Meta to start a company, shortly after the launch of Muse Spark 1.2. Yu had worked on Meta’s TBD Labs and multimodal AI projects, and his departure is seen as another example of accelerating top AI talent outflow from Meta.

More importantly, Tulloch’s departure reflects that this talent war is not pressuring Meta alone. Top AI researchers are rapidly moving between OpenAI, Anthropic, Meta, and Google. Thinking Machines has even seen multiple co-founders switch allegiances. According to The Wall Street Journal, Tulloch is the third recent co-founder to change roles, with the other two moving to Google and Meta, respectively.

For elite AI researchers with already high market value, compensation may no longer be the sole or even primary factor. Research direction, computing resources, team composition, decision-making autonomy, product deployment speed, and personal judgment on the trajectory of AI development may all influence whether a top researcher chooses a large tech company, an AI lab, or entrepreneurship.

Meta continues to heavily invest in AI infrastructure, aiming to convert its massive social platform user base into users of AI services through products like Muse. Reuters reports that Meta expects to spend over $130 billion this year on AI chips and related infrastructure, with Muse seen as a key attempt to turn massive AI investments into new business models.

Meanwhile, Anthropic is also rapidly expanding. As Tulloch joins Anthropic, the company faces fierce competition with OpenAI, Google, and Meta, and is preparing for an IPO. This means Tulloch’s move is not just a job change—it could significantly strengthen Anthropic’s technical capabilities in model training, inference performance, and AI agents.