AI safety researcher: my biggest update is that AI capabilities are spikier and need less agency than I thought
EpistemicHope · x · 2026-11-22
Responding to a challenge about what updates AI risk researchers should make, EpistemicHope says he's been surprised repeatedly and has mostly become more humble. His biggest specific update: AI capabilities are "spikier" than he once imagined, and more is possible with less agency than he thought.
He pushes back asking which update his interlocutor means — whether the implication is that AI takeover risk should be revised downward.
A candid window into how some AI risk researchers are recalibrating as real model progress lands.
More from AGI Musings
- Why the AI-anxious can see their roles dying but not what comes next — RileyRalmuto · 2026-11-22
- We are the ones inside the Chinese room, argues dev as AI hands out proofs — sytelus · 2026-11-22
- RH proof debate sparks reflection: centaur cognition, not lone genius, drives discovery — round · 2026-11-22
- AI drug design races ahead while clinical trials become the bottleneck — MaxUnfried · 2026-11-22
- "Marx predicted generative AI": the labor theory of value debate goes viral — felpix_ · 2026-11-22
- Peter Wildeford puts 15-30% odds on AI proving the Riemann Hypothesis by 2035 — peterwildeford · 2026-11-22