Three current Applied AI employees say the unit, formed in March to support Meta Superintelligence Labs, spends its days generating evaluation puzzles for the company's central AI research unit.
Earlier this week, a livestreamed, employee-only Meta presentation was interrupted by an expletive-filled outburst in which the speaker complained about "being the company's bitch" and urged colleagues to message a specific Meta AI executive to tell him "he's a piece of shit." One of the two presenters covered their face with their hands; meeting leaders resumed the technical talk after asking attendees to mute. The call, open to thousands of Meta employees, had started, by all accounts, "spicy."
According to three current Meta employees who spoke to WIRED, that moment was not an anomaly. It was the visible crack in a structure Meta built deliberately: a two-tier AI organization in which a roughly 6,500-person support unit called Applied AI does the drudgework that the glamour demos of Meta Superintelligence Labs rest on.
The architecture is straightforward once you see it. Meta Superintelligence Labs is the company's central AI research unit, the org whose model releases and benchmark headlines drive the public story of Meta's AI. Applied AI was formed in March 2025 to support that lab and now houses roughly 6,500 engineers and product managers, according to WIRED's reporting. The relationship is one of service: Applied AI exists to make Superintelligence Labs' work possible, WIRED reported. The 6,500 people inside it are not, by and large, the researchers whose names appear on model cards. They are the staff generating puzzle tasks that evaluate the models, per WIRED's reporting.
Three current Applied AI employees described the work to WIRED on condition of anonymity, because they were not authorized to speak to the press. They said the day-to-day consists of puzzle-generation tasks designed to improve AI models, work several of them described in blunt terms. One called the environment "literally the gulag." Another said the work was "soul-crushing." A third described the puzzle tasks as the kind of assignment that makes the job feel less like engineering and more like piecework in service of someone else's product.
The complaints are not about a single bad day. All three told WIRED there is widespread dissatisfaction with how the unit was assembled, and the recent employee outburst read, to them, as the predictable result of an org chart that concentrates prestige in one place and routinized labor in another. The interrupter's identity and motives have not been established; WIRED reported that the speaker could not be reached for comment. Treating the outburst as evidence of broader dysfunction therefore rests on the convergence of three independent accounts from inside the unit, plus the recording, rather than on confirmed causation from any one source.
The structural frame matters because it changes what the story is about. Read as a personality story, the outburst is an isolated incident on an internal call, the kind of thing a company apologizes for and moves past. Read as a labor-structure story, it is a window into how a major AI lab actually runs: a small number of researchers at the top of the org chart, supported by a much larger pool of workers whose job is to feed the evaluation pipeline that those researchers' models are measured against. The 6,500-person Applied AI unit is the size of a mid-sized company's entire engineering workforce, and it is, according to the three employees, largely invisible in the public story of Meta's AI.
That invisibility is the through-line. The "gulag" quote, the "soul-crushing" quote, and the puzzle-task description are three people's words about three different facets of the same arrangement. The outburst on the livestream made the arrangement audible for a moment. What remains, after the stream is muted and the meeting moves on, is the structure: a glamour research org, and a 6,500-person support unit underneath it, doing the work that the glamour org's outputs depend on, WIRED reported.
The unresolved piece is whether Meta treats the structural question as a real one. The company has not, in the material available, addressed the Applied AI / Superintelligence Labs split on the record, and the anonymous sourcing means the sharpest claims, about morale, about the nature of the work, about the "gulag" framing, belong to the three employees who made them. A Meta statement, additional on-the-record voices, or follow-up reporting on how the unit's headcount and mandate have changed since March could meaningfully shift the framing. For now, the picture is three anonymous accounts pointing at one conclusion: that "building AI" at Meta is not one job but a hierarchy of jobs, and that the workers at the bottom of that hierarchy are describing it in their own words.