Herbie Bradley: preventing model weight theft via SL5 security imposes huge costs
herbiebradley · x · 2026-11-20
Herbie Bradley argues the "Los Alamos" analogy for AI security is disanalogous: preventing nation-state model weight theft would require drastic talent cuts from clearance issues, slower research from siloing, and forfeiting frontier models' defensive benefits. He rejects the SL5 consensus that the slowdown is worth "burning the lead," arguing it doesn't clearly improve safety and only imposes huge costs on a self-commoditizing technology.
More from AGI Musings
- ludwigABAP digs up 20 old 'collapse' essays and cringes at his own AGI doom predictions — ludwigABAP · 2026-11-20
- 'A year can change an industry': Anthropic IPO and Jensen Huang's exit one year after Opus 4.5 — ColleenMBrady · 2026-11-20
- Ex-Anthropic/OpenAI researcher bets 80% of cognitive labor automated by end of 2028 — rohanpaul_ai · 2026-11-20
- Kevin Xu shares what zero-sum AI looks like, Wall Street edition — kevinsxu · 2026-11-20
- AI nears doing most intellectual work while vendors still pitch 15% productivity gains — VraserX · 2026-11-20
- Philanthropy must fund proactive biology data collection or AI won't cure disease, researcher argues — juanbenet · 2026-11-20