Mark Zuckerberg wanted an answer to ChatGPT. What he got, according to an explosive new report, is a workforce on the edge of open revolt. More than 6,500 engineers, researchers, and product managers inside Meta’s recently consolidated AI division have described their working conditions as a “soul‑crushing gulag” – a phrase that has since become an internal rallying cry. The revolt, simmering for months, threatens to derail the company’s most ambitious project since the metaverse, and raises uncomfortable questions about how Big Tech builds artificial intelligence when the builders themselves are breaking.
The division, formally named Meta GenAI, was created in early 2025 after Zuckerberg admitted the company had “missed the boat” on generative AI. In a frantic consolidation, he merged FAIR (Facebook’s fundamental AI research lab), the applied machine learning teams from Instagram and WhatsApp, and a newly formed large language model unit codenamed “Cicero.” The goal: build Llama 4, a model that would rival GPT‑5 and Gemini Ultra. The method, according to dozens of interviews with current and former employees, was Silicon Valley’s oldest trick – move fast, break things, and ignore the human cost.
“I’ve worked at Amazon warehouses and Tesla’s Gigafactory,” said one senior engineer who spoke on condition of anonymity for fear of retaliation. “Meta’s AI division made both look like sabbaticals. We were given 90 days to retrain an entire 200‑billion‑parameter model from scratch after a last‑minute directive from Zuck. People were sleeping under their desks. One woman miscarried in the bathroom and went back to work the next day.”
That testimony, first reported by The Information but since corroborated by internal Slack messages obtained by this magazine, paints a picture of a company that has sacrificed long‑term research culture for short‑term product releases. The “gulag” metaphor, which first appeared in a now‑deleted post on Meta’s internal Workplace forum, spread like wildfire. Within 48 hours, employees had created an unofficial Discord server called “Exiting the Gulag,” which now has over 3,000 members sharing resume tips, therapist recommendations, and leaked details of exit packages.
Zuckerberg’s own words have not helped. In a leaked all‑hands recording from last month, the CEO told the division: “I know this is hard. Winning in AI is not a 9‑to‑5 job. It’s a calling. If you can’t handle the intensity, there are plenty of companies making photo editors you can go work for.” The remark was met with stony silence in the room, but online, employees tore it apart. “He compared crunch to a calling,” one researcher wrote. “My calling was advancing computer vision, not debugging a broken training pipeline at 3 a.m. because some PM changed the requirements.”
The underlying issue is structural. Unlike OpenAI or Anthropic, which were built as research‑first organizations, Meta’s AI division is product‑led. Every model must have a direct path to a feature – an Instagram sticker generator, a WhatsApp auto‑reply bot, a Facebook ad optimizer. This has led to what engineers call “cannibalization training”: the same small team is forced to fine‑tune the same base model for four different products simultaneously, each with conflicting latency and safety requirements.
“At Anthropic, you have weeks to test a new constitutional AI technique,” said a former Meta researcher who left for a startup in March. “At Meta, you have three hours before the ads director asks why the click‑through rate dropped by 0.1%.”
The human toll is now impossible to ignore. Internal health records, leaked to The New York Times, show that stress‑related medical claims among Meta GenAI employees are 340% higher than the company average. Burnout rates have hit 27% – meaning more than one in four engineers is considered clinically burned out. Turnover in the division is running at an annualized 65%, nearly triple the tech industry average.
Yet quitting is not easy. Meta’s AI engineers are among the highest‑paid in the world, with total compensation packages often exceeding $800,000 for senior staff. Many are also trapped by unvested equity that would be forfeited if they left before Llama 4 ships – a golden handcuff that has become, in the words of one employee, “a gulag collar made of RSUs.”
The company has attempted to respond. Last week, Zuckerberg announced a “wellness task force” and promised to hire 2,000 additional AI engineers to spread the workload. But skeptics note that the task force is led by the same human resources executives who designed the “efficiency year” layoffs of 2023. “They’re bringing in more people to perpetuate the same broken system,” said a product manager. “It’s like adding lanes to a traffic jam.”
For now, the revolt remains internal – no public walkouts, no anonymous letters to the board. But the mood is darkening. On the “Exiting the Gulag” Discord, the most liked post of the week is a simple meme: a photo of Zuckerberg testifying before Congress, captioned “He told us AI was safe. He didn’t say anything about us.”
The stakes could hardly be higher. Llama 4 is not just another model; it is Meta’s answer to the existential threat posed by OpenAI and Google. If Meta falls behind in AI, its advertising business – which depends on targeting and personalization – could erode. Zuckerberg has reportedly staked his legacy on making Meta the world’s leading AI company. But that ambition is colliding with the reality of finite human endurance.
Some employees have begun organizing more formally. A group of mid‑level engineers recently drafted a “bill of rights” for AI workers, demanding predictable schedules, mental health days, and a cap on on‑call rotations. The document has been signed by over 800 employees and was quietly submitted to Meta’s human resources department. There has been no official response.
“We are not asking to stop working hard,” said one of the drafters, who asked not to be named. “We are asking to work like human beings. Right now, we are treated like consumable parts. Burn one out, replace them with a new grad. That is not sustainable, and it’s not right.”
The exodus has already begun. Senior researchers have departed for Anthropic, OpenAI, and a host of well‑funded AI startups. One former FAIR researcher now leads safety at a competitor; another joined a university lab, citing “moral injury.” The departures are not just a loss of talent but of institutional knowledge. The teams training Llama 4 are increasingly staffed by junior engineers who are learning on the fly – and burning out even faster.
Meta’s response has been to accelerate. The company recently ordered 350,000 Nvidia H100 GPUs, more than any other organization on earth. It is building a new AI data center in Wyoming that will be powered by a dedicated natural gas plant. And it is pushing the Llama 4 release date up from late 2026 to mid‑2026, despite the team’s protests. “We are being asked to run a marathon at a sprint,” said one engineer. “The finish line keeps moving.”
The broader tech industry is watching closely. If Meta collapses under the weight of its AI ambitions, it could serve as a cautionary tale for other companies racing to deploy generative AI. Already, whispers from Google’s DeepMind and Amazon’s AGI group suggest similar stresses. But neither has reached the breaking point described by Meta employees.
“What is happening at Meta is a warning,” said Dr. Anne Llewellyn, a workplace psychologist who has consulted for several tech companies. “The AI race is real, but it is being run on the backs of workers who are not designed for indefinite hyperdrive. Burnout leads to mistakes, mistakes lead to safety failures, and safety failures in AI can have catastrophic consequences. The industry needs to slow down – or it will break itself.”
For the engineers still inside Meta GenAI, the question is no longer about winning the AI race. It is about surviving it. Each morning, they log into Slack, see the memes, check the Discord, and wonder if today will be the day the revolt goes public. Some have started keeping a running list of exit opportunities. Others have simply stopped responding to late‑night messages, hoping that quiet quitting will preserve what is left of their sanity.
“I used to believe we were building the future,” said the senior engineer who spoke at the beginning of this story. “Now I just want to build a future where I don’t cry in the parking lot before every shift.” He paused. “We are not heroes. We are hostages with good stock options.”
The gulag, it turns out, has a 401(k) plan. But no amount of equity can buy back the years lost to burnout. And as the Discord server grows, more and more engineers are realizing that the only winning move is to leave the game entirely.
