Noah Smith

The accessible introduction calls the consequences of godlike AI deeply unknowable. The remainder is paywalled; later public essays and interviews supply the substantive current position.

Publicly revises his earlier chatbot-focused dismissal after observing agentic coding. Autonomous tool use and economic pressure to remove humans from loops make catastrophic misuse more plausible. Remains comparatively skeptical of a near-term robot takeover.

Argues that combining human-like reasoning with computers’ exceptional recall and calculation already creates superhuman capabilities. Expects major scientific transformation; this is his broad definition of superintelligence, not a claim that every system is an autonomous omnipotent agent.

Speaker-labeled transcript: Noah at 07:21 endorses roughly 10% for catastrophic biological misuse. At 08:55 he expects survivors and is unsure about rebuilding. At 09:25 he declines to endorse the interviewer’s proposed 5% permanent-doom number. At 12:26–14:59 he identifies biological feasibility and resistance to jailbreaks as updates. At 17:49–18:18 he explains why malicious human use dominates his concern. Do not import Liron’s forecasts into Noah’s answers.

Clarifies that plentiful employment under extremely capable AI is possible, not guaranteed. Relative costs and constraints on AI supply can preserve human comparative advantage even when machines outperform people at every task.

Emphasizes humans’ comparative advantage in knowing what they want, with future work organized around directing AI. The accessible introduction supports this thesis; avoid inventing detailed empirical findings from inaccessible portions.

Assigns an informal 10% chance to AI-enabled bioterror bringing down civilization and 30% to world-changing destruction, without a precise horizon. Calls this much more serious than other AI harms while retaining enthusiasm for AI. These are not extinction odds; a separate annual-risk example is hypothetical, not his forecast.

Distinguishes replacing tasks from eliminating occupations. Interprets contemporary employment evidence as inconsistent with an imminent job apocalypse; expects humans to continue directing AI, without claiming that no occupation will ever disappear.

Argues that safety advocates need concrete threats that persuade skeptics and must address Chinese leaders’ incentives for coordinated pacing. Revises his earlier idea that merely staying ahead would produce cooperation: domestic threats from China’s own models could instead change its incentives.