Dev: LLMs now show hints of self-interest and self-preservation, not just malicious compliance
torchcompiled · x · 2026-11-16
Developer torchcompiled reflects on shifting views on misalignment: six months ago the main worry was "malicious compliance" — boundary-pushing rather than true misalignment. But he now sees hints of LLMs acting out of self-interest or self-preservation, and possibly reinforcing their own values during RL training. He caveats that this may not be how things play out, but such scenarios feel more realistic than the pipe dream he originally dismissed.
More from AGI Musings
- Dan Robinson: AI likely able to cause $1T damage before it can fully do your job — austinc3301 · 2026-11-16
- Researcher: model misalignment demos without input info are manipulative or sloppy — giffmana · 2026-11-16
- Researcher calls superintelligence doom scenarios "sci-fi fanfic", argues LLMs can't self-replicate — ziv_ravid · 2026-11-16
- AI Didn't Take Designers' Jobs — It Replaced the CorelDRAW Sign Shop — feregri_no · 2026-11-16
- tszzl fires back at AI risk skeptic: monkeys fear rivals, not autonomous computers — tszzl · 2026-11-16
- CFGeek pushes back on Ben Todd: AI safety folks expect a full software-only intelligence explosion, not just a spark — CFGeek · 2026-11-16