Herbie Bradley: SL5 security not worth it if it slows AI lab by 6-9 months
herbiebradley · x · 2026-11-20
- Responding to concerns that AGI weights would leak to China within days, Herbie Bradley argues AGI is not the endpoint—the race continues toward ASI, which still faces bottlenecks like data—so the logic against slowing down persists.
- His framework: adopting SL5-level security is only worthwhile if the slowdown over Y years stays below the lead X over China; he estimates plausible values of X ≈ 6-9 months and Y ≈ 3-5 years, otherwise accept weight-theft risk.
- He adds that for reducing Chinese sabotage risk against the US, SL5 is sufficient but not necessary.
More from AGI Musings
- Metaculus Had This Milestone at 2035 — Until This Week — StartupYou · 2026-11-21
- Mathematician littmath reverses skepticism on AI proving Riemann Hypothesis by end of 2026 — littmath · 2026-11-21
- The 250,000-page Riemann proof claim: when computational verification outpaces human comprehension — SydSteyerhart · 2026-11-21
- lateinteraction: a tool that solves Millennium Prize Problems yet fails 1000 small things reshapes the map — stochasticchasm · 2026-11-21
- 'Quitting programming over AI is like a math teacher quitting over calculators' — prasenx · 2026-11-21
- Experts gave AI 10% odds of cracking a Millennium Prize Problem by 2027 — two now claimed — theimposingshadow · 2026-11-21