Researchers clash over AI pain-axis paper: steering artifacts may explain weird results, v2 already out
rgblong · x · 2026-10-08
- rgblong questions the "pain direction" paper by Dillon Plunkett et al.: the vector extraction method itself looks odd, and steering is messy in general — anomalous results more likely indicate problems with the axis or setup than a genuinely counterintuitive "pain axis".
- Co-author camhberg responds that the project is ongoing, feedback led to a rapid v2 (with a v3 coming), and lists points of agreement: finding some pain-related direction was expected; behavioral results shouldn't be read as relief-seeking (the core revision in v2's behavioral controls); the paper makes no claim that current models experience pain.
More from AGI Musings
- Quintin Pope floats a concrete ASI tell: rewriting ffmpeg or sqlite 5x faster with 2x less memory — QuintinPope5 · 2026-10-08
- Google RCT: AI boosts patent lawyers' output but juniors gain no judgment — Dr_Atoosa · 2026-10-08
- AI splits ambitious people: some coast, others work longer hours than ever — doodlestein · 2026-10-08
- Meme: mathematicians face existential AI crisis while everyone else just works — SuB8u · 2026-10-08
- Norvig's Classic Essay on Chomsky and the Two Cultures of Statistical Learning Still Reads Fresh in the LLM Era — 3scorciav · 2026-10-08
- Redditor calls out AI leavers for vague existential-risk talk with no specifics — IShootThemSteve · 2026-10-08