Meta's PSC machine is eating its engineering org from the inside

What happened

Gergely Orosz's June 16 Pragmatic Engineer piece, "Why is Meta destroying its engineering organization?", landed at 263 on Hacker News with an unusually long comment thread for an inside-baseball org post. The argument is blunt: Meta's Performance Summary Cycle (PSC), combined with the 2023-2025 'low performer' culls and Zuckerberg's public reframing of layoffs as quality control, has turned the company's engineering ladder into a survival game with no remaining connection to engineering output.

Orosz pulls from current and former Meta engineers across E5 through E7, plus several engineering managers. The pattern is consistent. The PSC, run twice a year, requires every IC to write a self-review, gather peer reviews, and have their manager defend their rating in a calibration meeting against a forced distribution. The bottom buckets — "Meets Most" and below — feed directly into the rolling performance-based attrition that replaced the old 'compassionate transition' model. Roughly 5% of engineers are now expected to be cut per cycle regardless of absolute performance.

The second-order effects are what make the piece sting. Managers report spending 30-40% of a half writing PSC packets. Senior engineers describe choosing projects based on "PSC legibility" — work that produces a clean narrative arc with a measurable launch — over harder, longer infrastructure work that won't resolve inside one cycle. Cross-team collaboration has collapsed because helping another team is help that doesn't show up in your own packet. One quoted E6 puts it directly: "I haven't done a code review I wasn't named on in eight months."

Why it matters

The HN comment thread is the real corroboration. Threads from ex-Meta engineers describe the same mechanic from different angles: the L7+ who realized his promotion case required him to manufacture a re-org, the manager who admits she rated a strong engineer "Meets Most" because calibration forced someone into the bucket, the infra IC who watched a multi-quarter migration get killed because no single half would produce a launchable artifact.

This isn't the standard "big tech is bureaucratic" complaint. The specific failure mode Orosz documents is a feedback loop: PSC rewards legible solo impact, which suppresses collaboration, which makes the remaining collaborative work harder, which makes engineers retreat further into solo legible work. Google's Perf has a version of this problem; Amazon's stack rank has a sharper version. But Meta is the first to combine forced distribution with explicit, repeated, public performance-based attrition at a scale where the bottom of the distribution is a target rather than a tail.

The counter-argument — and several HN commenters make it — is that Meta is still shipping. Threads launched. Llama 3 and 4 happened. Reality Labs, expensive as it is, produces hardware. But "still shipping" is the wrong benchmark for a system that's been running for two years; the right one is what stopped getting built. The piece's strongest evidence is negative: the infrastructure projects that used to come out of Meta — RocksDB, PyTorch, React, GraphQL, Cassandra's ancestors — have no 2024-2026 equivalents. The internal tools have gotten worse. Engineers describe Buck2 and the monorepo tooling as visibly under-maintained because nobody can write a PSC packet about keeping a build system fast.

Third, the talent flow matters for the rest of the industry. Meta has been one of the largest training grounds for distributed systems engineers in the world. The ex-Meta engineer who shows up at your startup in 2024 was shaped by the old culture — bootcamp, strong mentorship, hard infra problems, freedom to switch teams. The ex-Meta engineer who shows up at your startup in 2027 will have been shaped by two years of writing PSC packets, and that's a different professional. They'll be excellent at narrativizing impact and worse at the unglamorous work that doesn't narrativize.

What this means for your stack

Three practical implications. First, hiring. If you're screening senior engineers from Meta in the next 18 months, the useful interview signal isn't "tell me about your biggest impact" — they've all rehearsed that answer in PSC format. The signal is the inverse: ask them to walk through a system they didn't own but had to operate inside. Ask about a project they helped on without getting credit. PSC-shaped engineers will struggle to answer; the ones who kept doing real work despite the system will have stories ready.

Second, dependency risk. If your stack depends on Meta-maintained open source — PyTorch, React, Buck2, Zstandard, RocksDB — the maintenance posture is worth re-examining. The projects with strong external maintainer communities (PyTorch, React) are insulated; the ones that depend on internal Meta staffing (Buck2, parts of the Hermes JS engine, internal-first OSS) are more exposed than they were in 2022. This isn't a reason to migrate, but it is a reason to budget for slower upstream responsiveness and to value contributors outside Meta more highly when you evaluate the bus factor.

Third, the hiring market correction is real but uneven. Senior Meta ICs are leaving in numbers, but they're concentrated at L6/E6 — the level where PSC pressure is highest and the equity refresh math has gotten worst relative to the public market. That's the band where startups and mid-stage companies should be aggressive. The L7+ talent is harder to pry loose because their PSC outcomes are still mostly favorable and their RSU cliffs are still meaningful.

Looking ahead

The interesting question is whether the PSC machine is self-correcting. Orosz's piece is implicitly pessimistic: forced distribution plus performance-based attrition is a stable equilibrium because the people who'd reform it are also rated by it. The only forcing function that historically breaks these cycles is a competitor that ships something Meta can't match, which makes the AI talent war the actual variable to watch — not because Meta will lose engineers to OpenAI or Anthropic, but because the engineers it keeps will be the ones who optimized hardest for the PSC, and that's the cohort that will define what Meta builds next.