Gary Marcus doubts AI "pain axis" study; critics note it's just steering vectors, reproducible by anyone
burny_tech · x · 2026-09-20
Gary Marcus pushed back on a study claiming an AI "pain axis," saying he doesn't buy it but would love to try the authors' scripts on an alternative set of cases. A critic counters that the paper itself states "we have not shown that our pain axis is consciously experienced": the work is an empirically constructed steering vector with experiments run on it. Steering vectors work for almost anything depending on construction quality, the equation is in the paper, and open steering-vector libraries exist, so anyone can reproduce the approach.
More from Research
- Interactive diagram makes Self-Attention vs Cross-Attention instantly clear — ProfTomYeh · 2026-09-20
- Frank Nielsen's duo Bregman divergence paper unifies KL across exponential families — FrnkNlsn · 2026-09-20
- VBVR-Pro: 300 visual reasoning tasks plus verifiable rewards for RL training — 机器之心 · 2026-09-20
- 45 expert scientists stress-test AI reviewers against Nature-family peer reviews — windx0303 · 2026-09-20
- World Modeling for Physics workshop with LeCun at Aspen opens CFP, deadline Oct 9 — ylecun · 2026-09-20
- Claude Factors RSA-896 Challenge Number, Researcher Claims — Singularitarian · 2026-09-20