Meta Loses 8 AI Researchers as Superintelligence Lab Stumbles

💡 TL;DR - The 30 Seconds Version

👉 Eight researchers and engineers left Meta Superintelligence Labs within two months of launch, including several who returned to OpenAI after under 30 days.

💰 Meta offered nine-figure compensation packages up to $100 million to lure top AI talent from competitors like OpenAI and Google.

🏭 Longtime Meta veterans are also departing, including Bert Maher (12 years, built PyTorch) who joined Anthropic and Tony Liu (8 years, GPU systems).

🔄 Meta has repeatedly reorganized its AI division this year, with employees describing managers changing several times amid constant structural shifts.

🌍 Geographic constraints hurt retention as Meta concentrates AI work in Menlo Park while competitors embrace distributed research models.

🚀 The pattern shows AI leadership depends more on stable teams and organizational clarity than signing bonuses in the race for superintelligence.

A wave of early departures is testing Mark Zuckerberg’s “personal superintelligence” push.

Meta’s splashy new Superintelligence Labs promised momentum; within weeks, key hires walked. At least eight researchers and engineers have exited in two months, including several who boomeranged back to OpenAI, according to Business Insider’s exclusive on the exits.

What’s actually new

The attrition isn’t just a headline. Two recruits, Avi Verma and Ethan Knight, reversed course in under a month and rejoined OpenAI. Another researcher, Rishabh Agarwal, said he was leaving after five months, praising Meta’s “talent and compute density” while signaling a desire to take a different risk. That’s a fast unwind.

Chaya Nayak, a near-nine-year Meta veteran who directed product for generative AI, is also heading to OpenAI. The pattern undercuts Meta’s recruiting blitz and gives its fiercest rival an unplanned tailwind. It’s not the story investors were sold.

Who left—and where they went

The exits reach beyond new hires. Longtime builders who helped shape Meta’s AI stack are moving on. Bert Maher, central to PyTorch and Triton, left for Anthropic. Tony Liu, who led PyTorch GPU systems, stepped away after eight years. Chi-Hao Wu departed to become chief AI officer at Memories.ai. Others, including Aram Markosyan and Afroz Mohiuddin, are out as well. That’s depth walking out the door.

Some of the shortest tenures were in the new lab itself. Verma and Knight had OpenAI on their résumés and went back. The “re-hire” dynamic matters because it suggests expectations inside Meta’s superintelligence effort didn’t match the recruiting pitch. People voted with their feet. Quickly.

The $100M pitch meets friction

Meta’s sales job was bold: nine-figure packages for top talent, per Wired reporting. The goal was obvious—compress the timetable to catch OpenAI and Google by concentrating people and compute in a single, hard-charging unit. Money buys attention.

But compensation can’t paper over operating reality. If teams churn, mandates shift, or decision rights stay murky, high earners and veterans alike will look elsewhere. Culture and clarity set the floor. Pay can’t be the ceiling.

Structure keeps shifting

Meta has repeatedly reworked its AI org chart this year, most recently splitting staff into four groups, according to the Wall Street Journal. Employees describe managers changing several times. Constant reshuffles drain momentum and complicate career ladders. Stability is a feature, not a luxury.

Meta counters that “some attrition is normal,” noting many departures were long-tenured employees. That’s true as far as it goes. It doesn’t resolve whether the new lab’s structure is fit for purpose. Governance matters.

Geography is a tax

Agarwal worked from Montreal. Meta’s AI leadership remains concentrated in Menlo Park. The industry has tilted toward distributed research, hybrid labs, and remote-first infrastructure. OpenAI and Anthropic have leaned into flexibility to widen their hiring funnel. Location policy can be culture by another name.

If proximity is mandatory, Meta may narrow its access to specialized talent clusters in Canada, Europe, and Israel. That choice can slow hiring and complicate retention. Friction compounds.

The boomerang favors OpenAI

Sam Altman called Meta’s poaching “distasteful” in an internal memo earlier this summer, per Wired. Now, OpenAI gains seasoned contributors without paying switching costs. It’s a reputational win and an execution boost. Momentum is a moat.

Meta still has scale advantages: vast platforms, massive datasets, and a willingness to spend on frontier compute. The company is also striking external deals, including research collaborations to accelerate multimodal work. Those levers remain real. They are not yet enough.

What this says about the race

Frontier AI is not only a parameter race. It’s an organizational race: who can maintain stable, mission-clear teams over multiple training cycles while managing safety, shipping cadence, and cost. That’s harder than it sounds. And it always takes longer than board decks assume.

History rhymes. Google’s early AI run depended on low churn around core groups. OpenAI’s ChatGPT moment rode team continuity through several scaling decisions. In both cases, talent stayed long enough to compound. That’s the lesson.

Limits and caveats

Early turbulence doesn’t doom Meta’s bid. The lab is weeks old; headcount is still high; the market for elite researchers is liquid and cyclical. Some exits reflect personal moves, not structural failure. Meta says attrition is normal. Both things can be true. Time will tell.

Why this matters:

AI leadership depends on retention and stable teams more than signing bonuses; churn erodes compounding learning and shipping cadence.
Organizational clarity and geographic flexibility are now competitive levers; missteps there can hand rivals momentum even without a new model release.

❓ Frequently Asked Questions

Q: What exactly is Meta Superintelligence Labs and when did it launch?

A: MSL is Meta's dedicated division for developing "personal superintelligence," announced by CEO Mark Zuckerberg in June 2025. The lab aims to create advanced AI systems that rival OpenAI's capabilities through concentrated talent and compute resources in a single unit.

Q: How much did Meta actually spend on these hiring packages?

A: Meta offered individual compensation packages reaching up to $100 million to lure top AI researchers. The company hired over 50 people before the exodus began, including 13 from Google, 3 from Apple, and 3 from xAI, suggesting total spending in the billions.

Q: What is PyTorch and why does Bert Maher leaving matter?

A: PyTorch is an open-source software framework that thousands of AI researchers worldwide use to build and train AI models. Maher spent 12 years developing it at Meta, making his departure to Anthropic a significant loss of institutional knowledge for fundamental AI infrastructure.

Q: Is this level of turnover normal for tech companies?

A: No. Losing eight senior researchers in two months from a flagship division is unusually high, especially when several return to competitors after under 30 days. Normal tech turnover involves gradual departures over longer periods, not concentrated exits of key personnel.

Q: How does this compare to OpenAI's team stability?

A: OpenAI maintained core team continuity through ChatGPT's development, with key researchers staying multiple years. The company now benefits from "boomerang" talent returning from Meta while avoiding integration costs, strengthening its position in the AI leadership race without major recruitment expenses.