Early Anthropic Memoir Reveals Plan to Train Dangerous Bio, Cyber, Nuclear Models to Scare Governments
pmddomingos · x · 2026-10-28
Pedro Domingos shares Ivan Vendrov's Substack memoir of early Anthropic (Oct 2022–May 2023). Highlights include a co-founder arguing strategy must rely on markets and civil society since governments will act too late, and a leaked-in-memory anecdote: a senior colleague considered a special team training dangerous bio/cyber/nuclear models to show scary demos to governments and force them to take AI risk seriously — Vendrov politely declined. He stresses it's one junior employee's partial view, hoping other lab insiders fill in the picture.
Related event: Ex-Anthropic Employee's Memoir Reveals Early Internal Stories(2 posts)→
More from Companies & People
- NVIDIA is hiring DevRels for open-source inference framework Dynamo — blelbach · 2026-10-28
- Ex-Meta comms staffer mocks Oracle's 14-press-release blitz: comms needs a seat at the table — alexvoica · 2026-10-28
- NVIDIA's Healthcare AI Map: Pharma Ranks Literature Review Top Agent Use Case at 55% — HealthcareAIGuy · 2026-10-28
- Alibaba's Z-Image team reportedly disbanded after releasing $630K-trained 6B model — teortaxesTex · 2026-10-28
- Microsoft buyouts risk pushing top talent to rivals while leaving tech debt — DavidLinthicum · 2026-10-28
- User blocked by Eliezer Yudkowsky after mocking him; screenshot surfaces Altman's 'three endgames' report — ryunuck · 2026-10-28