Benjamin Todd: superforecasters and experts all underestimated AI progress over 18 months
ben_j_todd · x · 2026-11-13
In a long-form review, Benjamin Todd argues the last 18 months proved almost everyone underestimated AI progress — the opposite of overhype. Superforecasters and the public were most wrong; AI experts did better but still too slow; only AI CEOs were too optimistic, and only slightly — assuming their "6 months" claims take 6-18 months would have served you well. Key numbers: AECI implies 100h time horizons at 50% reliability (33h at 80%) for leading models, and the trend of 8x/year growth in horizons has doubled to 50x/year — an acceleration that already happened 6 months ago. His four drivers (pretraining, RL, test-time compute, agent scaffolding) keep working; his claim that the next leap could bring beyond-human problem solving has arguably already landed in maths. His verdict on his own AGI-by-2030 piece: he underestimated the rate of progress.
More from AGI Musings
- Levels 9 and 10 of AI adoption: simulating strategies with competing agent organizations — danshipper · 2026-11-13
- zetalyrae's ASI preference order: universal ASI beats benevolent totalitarianism — esrtweet · 2026-11-13
- VC: 'Jensen ran Nvidia as the central bank of AI — that era is over' — nmasc_ · 2026-11-13
- Dev laments: a week of problem-solving is no longer worth an OSS project in the AI era — nptacek · 2026-11-13
- Richard Ngo: intelligence explosion believers will be 'directionally right but factually wrong' — NathanpmYoung · 2026-11-13
- Analyst: Everyone Underestimates Non-AI R&D Acceleration — It'll All Be Ripping Anyway — teortaxesTex · 2026-11-13