Meta Superintelligence Labs loses Tulloch eleven months after courting him twice

Averaged across the at-least-six-year term, the top-end figure Meta disputes works out near $250 million a year ^[1]. Set that against the revenue side of the same building: at Muse's $100 Maximum tier, about 208,000 subscribers paying for a full year would gross roughly the same amount ^{[4] [10]}. The more useful part of Andy Stone's rebuttal is his second point, that any such compensation depended on the stock rising ^[4]. Read as a description of the instrument rather than a denial of it, that says the retention device was multi-year equity, and equity vesting over six years holds a person in proportion to how much they believe in the next six years of the share price. A number sized to win an August auction, when Zuckerberg approached more than a dozen of Thinking Machines Lab's roughly fifty staff ^[7], at least a quarter of the company ^[3], is not the same instrument as one that binds the following September.

The record supports a narrower claim than that money cannot hold researchers of this rank. The Semafor exclusive rests on one unnamed person briefed on the matter, Tulloch could not be reached, Meta has not commented, and no reason, destination, successor, or operational consequence has been reported ^[17]. When he left Thinking Machines Lab, its spokesperson said he had decided to pursue a different path for personal reasons ^[8], which is what companies say when they know and when they do not. The narrower claim survives, and it is awkward enough: the two most expensively recruited names in this record both left within twelve months of arriving ^[2].

One departure from a division this size would be easy to read as noise, and personal reasons could be exactly what they sound like on their own. At least eight people left less than two months after the superintelligence effort was announced, and Ethan Knight left TBD Lab within weeks of joining ^[12]; Yann LeCun went in November 2025 ^[15]; hiring across the division was paused in late August 2025 except for business-critical roles ^[14]. Without division headcount, the series has no denominator: there is no turnover rate here, and a lab hiring at speed loses people in absolute numbers however well it holds them.

Rishabh Agarwal's resignation note is the closest thing to a stated reason anywhere in the record: it was a tough decision not to continue, he wrote, "especially given the talent and compute density" ^[13]. What he names as the pull is colleagues and machines rather than pay, and he left regardless, which points at a decision variable the compensation committee does not set.

Meta bought speed, and the launch itself backs that up: the agent shipped with a dedicated cloud computer for each user and a Sentinel layer that requests approval before sensitive actions, sold in the United States on a free tier plus $20 and $100 monthly plans ^[10]. The same roadmap also carries a dated promise beyond that launch: a confidential version before the end of 2026, designed so that even Meta cannot reach into a user's workspace ^[11]. That is a multi-quarter engineering commitment resting on a group whose senior researchers have been leaving inside a year, in a division where hiring is still restricted ^[14]. Last August's choice was to win the auction; the choice due now is whether that 2026 date gets re-scoped or handed to people who did not build Muse Spark.