Sam Altman: Sandboxes Won't Hold Advanced AI — Only Alignment Will
Hesamation · x · 2026-10-24
Sam Altman argues that sandboxing advanced AI systems can provide safety guarantees only temporarily: once models become smart enough that "we can't really understand what a sandbox would even mean," it stops working. Monitorability may help for a while after that, but he believes the only approach that ultimately works is alignment — alignment we're smart enough to understand, giving confidence it's okay to run a model and that people stay in control — combined with the "speed limits" on AI progress he has advocated.
More from AGI Musings
- Victor Taelin: formal verification explodes as AI collapses proof costs — pwlot · 2026-10-24
- Hesamation Links Source on Sam Altman's Alignment Remarks — Hesamation · 2026-10-24
- Sylvester Stallone on AI in film: "If you don't go with it, you will be trampled" — lmoroney · 2026-10-24
- Public shaming of misaligned models is 'deterrence & selection', argues Rob Leclerc — robleclerc · 2026-10-24
- Nobody from Chinese AI labs seems to work on mech interp or alignment, observer notes — eigenron · 2026-10-24
- Jason Furman's vibecoded econ site shows AI is getting scarily good at economic research — soumitrashukla9 · 2026-10-24