About 6,500 engineers and product managers got pulled onto a new Meta team in March. One of them later called it "a gulag."
That's the word an employee used to describe the company's freshly minted Applied AI unit, according to internal reporting. It's the kind of word that travels fast inside a company already losing its grip on morale.
Now the man who helped build the team is admitting, in writing, that the rollout was a mess.
Meta's chief technology officer, Andrew Bosworth, told staff on Monday in an internal post that the company had done an "atrocious" job standing up its AI division, as WIRED reported after reviewing the memo. His promised fixes run from the structural (fewer manager reshuffles) to the almost comically small (better snacks). More on the snacks later.
What went wrong, in the CTO's own telling
Bosworth didn't dodge. He wrote that leadership had eroded employees' trust that their expertise mattered, that their careers would grow, that they could actually make a dent at the company. He pinned part of the blame on a hiring cycle that lurched between boom and bust, stranding whole teams when strategy shifted underneath them.
The AI reorg was the sharpest version of that problem. Meta built the Applied AI group to sharpen its generative models. To staff it, the company just moved people over, into roles they hadn't asked for. Workers called the assignments menial.
Replying to a comment on his own memo, Bosworth was blunt: the company had failed to explain the vision, failed to show people how their careers fit the shift, failed to sketch how any of it would evolve.
That's a lot of failures for one paragraph. Give him credit for the candor, at least.
He didn't fully retreat, though. He held that conscripting people onto the AI team for speed was the right call, and he warned that some assignments simply won't be fulfilling. There will be stretches, he told staff, where the work demands sacrifice. Tricky message to land: we botched this, and also we'd probably do the drafting part again.
The walk-back, and the new rules
The practical changes are where this gets interesting. Bosworth said Meta plans to cap managers at roughly 20 direct reports and cut down on how often workers get handed to a new boss during restructurings. Managers, going forward, are supposed to manage first and do their own independent work second. Employees also get access to optional "AI coaching" tools, whatever those turn out to be.
The more concrete concession came from Maher Saba, the vice president running Applied AI. In a separate post late last Friday, Saba told the people forced onto his team that they could now apply for other roles inside Meta, if they could land them. He framed the original draft as a play on Meta's edge: its scale and the depth of its workforce, things he argued rival AI labs can't match.
Going forward, though, Saba said the team is returning to normal and giving people the freedom to chase roles they actually want. He also rebranded the old company mantra, swapping "move fast and break things" for a tidier promise to move fast and fix forward. The group, first aimed at boosting the coding and agentic abilities of Meta's frontier models, could later branch into security, debugging, and product work. Traditional engineering roadmaps, Saba added, no longer apply, because the work changes too fast to plan in six-month blocks.
Read between the lines. The team that pulled people in against their will is now offering them the door. That's not a small reversal. It's what happens when internal anger reaches a level executives can no longer ignore.
The broader morale problem
The Applied AI unrest isn't happening in a vacuum. Meta has been through mass layoffs, complaints about worker surveillance, and a general slide in how employees feel about the place. In recent days a string of executives, Mark Zuckerberg included, have posted internal notes acknowledging the mood and pledging to do better.
Bosworth, long read as a close Zuckerberg ally, leaned into the reassurance. He said Meta doesn't think AI will wipe out its workforce wholesale. Then he hedged with an old line: AI won't take your job, but a person who knows AI might. Performance reviews, he said, will hinge not on whether employees use AI but on whether they produce real impact with it.
There's a resource crunch under all this, too. Bosworth admitted there will be "tough trade-offs" over how much computing power each team gets for its AI work, and he urged people to flag bottlenecks rather than sit on them. Compute is the scarce currency of every AI shop right now, and Meta is no exception. Telling employees to escalate problems is fine. Whether the escalations get answered is the part that actually matters.
About those snacks
Then there's the morale offensive. Bosworth said Meta will improve its microkitchens (the office nooks stocked with drinks and snacks) and bump up travel budgets and spending on in-person social events. The stated goal, by his account, is to make the place fun and enjoyable again and rekindle the culture people signed up for.
Easy to roll your eyes here. A team called its work a gulag, and one of the headline remedies is better office snacks. But the snack budget isn't really the point. It's a signal. When leadership reaches for the small, visible gestures, it usually means they know the trust they spent will take a long time to earn back, and they want something to show in the meantime.
Meta declined to comment on the reporting.
So what's worth watching? How many of those drafted engineers actually take Saba up on his offer to leave. If a chunk of the 6,500 walk toward other teams, Applied AI loses the scale advantage Saba just bragged about. If almost nobody moves, the agency was always more symbolic than real. Either way, you learn something the memos won't tell you.