Turning off deception neurons isn't alignment, argues researcher
Sauers_ · x · 2026-10-20
Commenting on research that edits models to tell the truth, the author argues that simply ablating "deception neurons" is not an ideal strategy for aligned ASI. Instead, we should build a coherently aligned entity that retains the ability to use deception when needed — for social reasons or to counter bad actors.
Related event: Truthfulness Editing: Rewriting Model Beliefs Directly(2 posts)→
More from AGI Musings
- Compound launches PE Bench: 58 tasks testing AI agents as private equity analysts — peterjliu · 2026-10-20
- Merz asks Jensen Huang how Europe should regulate AI: "You can't regulate if there's no technology" — Hesamation · 2026-10-20
- iamtrask: agent swarms may avoid the three failure modes that doom flat human hierarchies — iamtrask · 2026-10-20
- Misalignment won't look like Terminator — humans get routed around, not exterminated — sourdub · 2026-10-20
- DeepMind researcher warns AI-only investment could leave science with more hypotheses than labs to test them — weballergy · 2026-10-20
- A 120-year-old Poincaré argument resurfaces: AI proofs only show how the pieces move — MoonL88537 · 2026-10-20