jd_pressman: Old-School Alignment Researchers Self-Censored Understanding Deep Learning
jd_pressman · x · 2026-10-29
In a discussion with FioraStarlight, jdpressman revisits the MIRI-era split between 'alignment' and 'capabilities' work, arguing many alignment researchers deliberately avoided thinking about how AI works to avoid advancing capabilities.
He adds that this self-censorship 'mentally neutered' them as alignment researchers, leaving them without real understanding of deep learning. FioraStarlight pushes back with an analogy: humans can do harmful things but don't, thanks to value alignment — a different thing from simply lacking capability.
More from AGI Musings
- Decade-long engineer: AI adoption at work has become a token-burning leaderboard — _jaydeepkarale · 2026-10-29
- HF AI report spun as existential warning; Yacine counters software isn't solved — akbirthko · 2026-10-29
- AI's math crisis won't come for the social sciences, says one meme — seanmcdonaldxyz · 2026-10-29
- AI won't replace smart people — it'll just expose who isn't, argues designer — AIandDesign · 2026-10-29
- Philosopher warns: negativity bias is warping how we model AI's real risks — sebkrier · 2026-10-29
- Academia brain drain vs AI automation: which trend wins? — RexDouglass · 2026-10-29