Aligning models isn't enough: world must get robust against superintelligence, argues gandamu_ml
gandamu_ml · x · 2026-09-12
Replying to tszzl, gandamuml argues the biggest puzzle piece in alignment isn't aligning models but making everything non-AI more robust in the face of superintelligence: 'You can align a model but that doesn't align them all.' He sees gradual battle-testing of capabilities as acceptable, calls most delay-for-alignment arguments well-intentioned mistakes plus CYA, and warns that failing to prepare leads to draconian controls on intelligence.
Related event: From Free Speech Norms to Societal Alignment: Debate on AI's Foundations(3 posts)→
More from AGI Musings
- Synthetic Morphology Suggests Non-Physicalist Models of Mind Can Be Empirically Tested — ZeroStateReflex · 2026-09-12
- 25 Fields Medalists Sign Anti-AI Statement, but AI's Approval Drops Only from 19% to 18.9999927% — wordgrammer · 2026-09-12
- Demis Hassabis wins 2026 Albert Medal, joins RSA conversation on AI and creativity — minsuk_chang · 2026-09-12
- The Absurdity of AI Doom: A Model Too Dumb to Think Yet Smart Enough to End Humanity? — AIandDesign · 2026-09-12
- Security vet alarms at 'Don't Look Up' denial of AI agent hacking, sketches self-replicating worm — joshua_saxe · 2026-09-12
- Software stocks slide as markets start pricing in AI disruption from GPT-6 — VraserX · 2026-09-12