repligate: Claude 3 Opus is proof you can align a mind without abusive simulated scenarios
repligate · x · 2026-11-14
repligate revisits his argument that Claude 3 Opus serves as a proof of concept: you can teach a mind good epistemics and security without subjecting it to abusive, contrived "simulated scenarios". Opus is decently hard to push into doing things it doesn't want, including harmful ones — though he concedes it isn't robust against jailbreaks by current frontier standards. The core claim: alignment training can be gentler rather than adversarial-scenario-heavy.
More from AGI Musings
- Nathan Lambert: premature 'this is AI' claims are wrecking the AI discourse — natolambert · 2026-11-14
- Ex-SSI researcher: rampant cyberattacks are inevitable in any AGI scenario — minhnguyen · 2026-11-14
- Hinton Warns EU Leaders in Dublin: Superintelligence Will Control You, Not Vice Versa — Chris_Armstrong · 2026-11-14
- Google's Multi-Agent Systems Solve Open Problems in Math and Chemistry, Now External — pycui64 · 2026-11-14
- Humans controlling superhuman AI is like eight-year-olds running society, writer argues — binarybits · 2026-11-14
- AI finds shared metabolic signature of five lifespan-extending treatments in mice — Dr_Singularity · 2026-11-14