A recent all-hands for Meta's Applied AI team started not with technical updates, but with a colleague telling leadership to inform a specific AI executive that he's "a piece of shit." The outburst, captured on an internal recording obtained by WIRED, is symptomatic of a deeper engineering culture fracture inside the 6,500-person unit formed in March to support Meta's Superintelligence Labs.
Three current engineers, speaking anonymously, describe the work as data assembly line drudgery. Their primary task: generating coding puzzles and complex software problems to train and evaluate Meta's frontier models. The output feeds a relatively small group of AI scientists who then pipe it to clusters of GPUs. One engineer called it "the gulag"—isolating, repetitive, and a sharp departure from building consumer apps at scale.
From a systems design perspective, the bottleneck is clear. Applied AI acts as a human-in-the-loop data pipeline, but with a deliberately flat management structure—ratios hit 50:1 on some teams. Engineers are required to produce two tasks per week, a throughput that feels mechanical rather than creative. The result: a workforce of highly skilled engineers performing low-level data augmentation, with no opt-out beyond quitting.
The tension isn't confined to Applied AI. A company-wide petition with 1,600 signatures protests a new initiative to monitor employee clicks and keystrokes for AI training data, though Meta recently allowed 30-minute pauses and exemptions. Meanwhile, Instagram's chief product officer Chris Cox described the atmosphere as "running a marathon in a hailstorm," referencing layoffs of 8,000 employees last month that added strain across data center engineering and product teams.
CEO Mark Zuckerberg acknowledged in a Friday memo that "we've made mistakes" and promised no further mass layoffs this year, capped manager-to-report ratios, and pledged assigned desks by year-end. He framed Applied AI as a waypoint—"very talented people" contributing while the company creates other roles. For now, engineers describe themselves as "draftees," watching each new wave of reassigned colleagues arrive with the same sinking realization: they're building the training set, not the product.
Source: Wired