Why is Meta destroying its engineering organization?
TLDR
Meta is gutting its once-legendary engineering culture, forcing up to half of core teams into AI data-labeling work and tracking every keystroke to feed its Superintelligence Labs. The company that let engineers pick their own projects for two decades now treats its best coders like Hunger Games tributes, reassigned overnight with no opt-out.
Gergely Orosz, who's covered Meta's engineering org for years, has been documenting something that looks less like a strategy shift and more like a controlled demolition. For twenty years Meta ran on two cultures: the reckless "move fast and break things" era of the 2010s, and the more mature "move fast with stable infra" phase that followed. Both versions shared one thing — engineers had autonomy, picked their own teams during a six-week bootcamp, and were treated as the profit center that built Facebook, Instagram, and WhatsApp into the products they are. That era appears to have ended sometime around April.
The trigger is Meta's all-in bet on AI, cemented by the $14.8 billion deal for a 49% stake in Scale AI and the hiring of its CEO, Alexandr Wang, to run Meta's AI strategy after Llama 4 landed with a thud earlier this year. Wang's specialty is training data, labeling, and RLHF — the grunt work that makes large models good at coding and reasoning. The problem is where Meta is sourcing that grunt work: its own engineering workforce, whether they like it or not.
Two moves stand out. First, Meta enrolled engineers into a system that logs every keystroke and mouse click, with zero opt-out, ostensibly to generate AI training data. Reuters reported this month that after weeks of internal fury, Meta relented slightly, letting staff pause tracking for up to 30 minutes and request exemptions — a memo signed by VP Stephane Kasriel. Notably, the UK version of the rollout apparently never happened, likely because it would run headlong into local data protection law. Second, and arguably worse for morale, Meta forcibly reassigned 30 to 50 percent of engineers on product and infrastructure teams into a new Agent Data Optimization org, now around 6,500 people strong — bigger than the equivalent teams at OpenAI or Anthropic. That's roughly one in every five or six of Meta's 25,000 engineers now spending their days writing test cases, packaging Docker containers, and grading AI-generated code instead of building anything that ships to users.
What makes this sting is the contrast with how Meta used to operate. Transfers were engineer-initiated. Bootcamp let new hires shop around for teams. None of that survived contact with Wang's mandate. One infra engineer described it to Orosz as feeling like the Hunger Games — random selection, sudden removal from a familiar environment, dropped into repetitive labeling work that vendors elsewhere are paying $100-plus an hour for precisely because it's soul-crushing but valuable. A few engineers say they've found ways to make it interesting by varying their tooling. Most haven't. And the infrastructure and security teams that got hit hardest are exactly the ones Meta can least afford to hollow out while chasing frontier AI performance.
My take
I've written before that surveillance-driven productivity theater is a tell, not a strategy — and forcing thousands of skilled engineers into keystroke-logged data labeling to chase a Claude or GPT competitor from a standing start smells like panic dressed up as ambition. Zuckerberg spent two decades building a culture where engineers wanted to work at Meta; he's now spending a few months proving that culture was optional all along, and I doubt the best people stick around to find out what's left of it.
Read more about this at: TLDR
