"Alignment isn't a panacea" — researcher argues for hardening non-AI systems instead
gandamu_ml · x · 2026-09-25
In a reply to a Schmidhuber-related thread, user gandamuml argues that "alignment" should not be treated as a panacea: even if you can train an aligned model, it keeps getting easier for anyone to train anything, including unaligned models. He contends the better strategy is to make everything non-AI more robust as AI spreads, and calls focusing solely on aligning the AI itself "magical thinking" — a notable voice in the alignment-vs-robustness debate.
Related event: Aftermath of Schmidhuber Debate: Alignment Is Not a Silver Bullet(2 posts)→
More from AGI Musings
- A $30B trial-design blunder: the case for translational AI in pharma — hardimanjames · 2026-09-25
- Yoav Goldberg on academia's core misery: ranking incomparable things without a metric — yoavgo · 2026-09-25
- Jensen Huang: Don't Think Being an AI Alarmist Means You're Doing Social Good — TinfoilTricorn · 2026-09-25
- Jensen Huang says 0% chance AI ends the world by 2030, drawing pushback from optimists — shaunralston · 2026-09-25
- Outsourced Thinking: AI Users and Critics Clash Over Whether AI Erodes Your Brain — Linahuaa · 2026-09-25
- NeurIPS paper shows self-improving, self-replicating agents evolve cooperation from scratch — maxhkw · 2026-09-25