Alignment debate: was corrigibility wrongly dismissed as a path easier than direct value learning?
jd_pressman · x · 2026-11-14
jdpressman argues that, starting from zero epistemic position in 2008, corrigibility genuinely looked easier and more robust than attacking values directly—and that this isn't a knock on Yudkowsky or MIRI. He also recalls informal discussions claiming it would be hubris to attempt progress because many had failed, which he says simply wasn't true, though he doesn't consider those views core canon.
Related event: Alignment Researchers Debate Why Value Alignment Was Long Neglected(3 posts)→
More from AGI Musings
- Paper proposes 1500-word briefs plus machine-verifiable appendices to fix AI-era peer review — WalterReade · 2026-11-14
- binarybits: A 150 IQ drives no better than 120 — intelligence has diminishing returns per task — binarybits · 2026-11-14
- Two kinds of knowledge: compressible reasoning vs high k-complexity memorization — jd_pressman · 2026-11-14
- JD Pressman: relativity could've been formulated much earlier — Einstein got lucky — jd_pressman · 2026-11-14
- Starship flights are becoming routine: second-ever flight, possible first double catch, barely news — tszzl · 2026-11-14
- AI Builder Pushes Back on AGI-2027: Jaggedness Persists, Models Still Ignore Simple Instructions — ChrisGPT · 2026-11-14