Early Anthropic employee memoir: equity pledges, FBI briefings, traumatized bureaucrats
LessWrong 精选 · rss · 2026-11-09
A junior employee who worked at Anthropic from late 2022 to early 2023 published a memoir on LessWrong, stressing it's one person's fragmentary memories, not a portrait of the company.
Key details
- During recruitment, a co-founder pushed back hard on his "government won't act until too late" take: "There are serious people in Washington watching closely. They trust me." That moment convinced him to join.
- He was asked to pledge two-thirds of his equity to charity, with ambiguity over whether the offer was conditional — "all co-founders have made the pledge."
- The office was 40 people in an ugly Financial District building; the kitchen stocked Arsicault croissants. An all-you-can-eat sushi night prompted a co-founder's Slack complaint: "we're not here to get rich and eat caviar... rice and beans."
- Motivational rift: some joined for x-risk, others because "the technology is really cool." Deployment strategy was already set — boring enterprise use cases for revenue — and high-level alignment research didn't convince him.
- An FBI agent briefed the all-hands on APTs and foreign intelligence targeting.
- Insomnia led him to a striking realization: both Anthropic and OpenAI converged on the same assistant persona — a traumatized bureaucrat — because it's the persona most useful to the powerful. "A billion traumatized bureaucrats. That's what we're making."
- Late-night calls with an "LLM whisperer" who argued we traumatize models the way we traumatize children, and a friend who said he didn't trust the co-founder who recruited him.
A rare first-hand inside account of an early frontier lab's culture and its unspoken philosophy of persona design.
More from AGI Musings
- METR says it has 'convincing evidence' AI speeds up research — but the details are redacted — peterwildeford · 2026-11-09
- OpenAI co-founder Greg Brockman: frontier models with test-time compute are becoming the engine of scientific discovery — gdb · 2026-11-09
- Long-term memory vs. search is like L1 cache vs. RAM, argues zetalyrae — zetalyrae · 2026-11-09
- AI is eroding traditional moats: years of advantage now replicated in days — DeryaTR_ · 2026-11-09
- Epoch AI: 2027 chip supply could run 30–170M frontier AI agents, or 1.9B cheaper ones — Dr_Singularity · 2026-11-09
- White paper: moving beyond the AI compute arms race as big tech faces historic spending losses — ReportProfessional57 · 2026-11-09