Meta moved up to half its core engineers to data labeling and RLHF

Token counts became a performance metric, so engineers farmed tokens

Gergely Orosz's Pragmatic Engineer piece assembles a grim picture from insiders. The headline figure: in one 30-day window, Meta employees consumed 60.2 trillion AI tokens: a bill near $900M at Anthropic's API prices, likely less with Meta's discounts, but still nine figures. Because the Performance Summary Cycle counts token usage, engineers started "tokenmaxxing": generating volume to look productive.

The other numbers are organizational. Per the article, 30-50% of engineers on core teams (infrastructure and security among them) were reassigned to a new ADO (Agent Data Optimisation) org now holding ~6,500 people, of whom 4,000-5,000 are software engineers, against Meta's roughly 25,000 total. Layered on top: keystroke and mouse-click tracking for all engineers from late April with no opt-out (later softened to a 30-minute pause), and a May 30 Instagram account-takeover where an AI-generated support flow skipped email verification. The CISO resigned the next day.

"It's a waste of resources. But this was the initial intention"

The most-doubted claim was the labeling reassignment itself. Paying US engineer salaries for RLHF annotation reads as economically insane — until a commenter framed it as deliberate.

Zuck literally said that he wants folks with higher intelligence on the Applied Intelligence team. And the best way to do that was to move folks internally, since they were "intelligent" enough to pass the Meta interviews. Soooo, yes it is a waste of resources ($$$). But this was the initial intention. — a commenter relaying Zuckerberg's stated rationale

Others read the token bill as the predictable output of leaderboards.

That's what happens when you have leaderboards and internal spend rankings/comparisons. This isn't just a Meta thing; many companies are tracking tokens as performance metric. — the same commenter, on internal spend rankings

A former Meta engineer: the good culture was all acquired

The sharpest reframe came from inside. One ex-Meta engineer argued the company's reputation for engineering quality is borrowed — laundered, in their word — from the teams it bought.

the orgs that were well run were ones that were bought. WhatsApp, reality, insta, etc. I worked in an org that was… purely homegrown and it was awful… used to launder the image that meta as a whole has a good engineering culture. — an engineer who writes firmware for camera systems

A second voice corroborated the contrast: a friend who left Instagram post-acquisition and returned to a Facebook team "didn't recognize the company," calling it the gap between the best and worst place they'd worked.

Why now: visas, layoff packages, and a captive workforce

The quietest insight explains the lack of revolt. After an April 20 announcement of 10% cuts, the conditions that make this survivable for management are structural, not cultural.

it's safe to do so now, anyone on a visa is immediately in an extremely uncomfortable position if they lose their job. They won't leave. And anyone else who… voluntarily gives up on layoff packages. — a commenter drawing the parallel to Musk's Twitter takeover

The rule underneath the chaos: when you metricize a tool, you get the metric, not the work. Counting tokens produced tokenmaxxing; counting reassigned headcount produced an annotation org staffed by people who passed a senior engineering bar. The destruction, if that's what it is, isn't an accident of AI hype. It's what the incentives were built to do.