AI wasn't overhyped: Benjamin Todd scores 18 months of forecasts
ben_j_todd · x · 2026-10-29
Benjamin Todd revisits his "Will we have AGI by 2030?" and finds almost everyone — superforecasters most of all — underestimated 18 months of AI progress. AI researchers were too slow; only AI CEOs were slightly too optimistic, so taking their 6-month claims as 6-18 months worked well. His four drivers (pretraining, RL, test-time compute, agent scaffolding) held up, but his own 2026 capability numbers were too slow, and maths may already be beyond-human.
More from AGI Musings
- The Commanding Heights: year-long essay argues AI gains accrue outside San Francisco — johncoogan · 2026-10-29
- Why lions lose to humans: the classic argument that strategic intelligence decides everything — StefanoGogioso · 2026-10-29
- One cold email to a Finnish MP helped Finland co-launch an international AI safety coalition — GarrisonLovely · 2026-10-29
- Could one person run a trillion-dollar company? Reddit debates AI, jobs and UBI — Cultural_Shake4690 · 2026-10-29
- Stanford's After42: experts rate 45 of 135 AI-proposed research questions worth pursuing — aigclink · 2026-10-29
- Agentic AI flags frequently missed cancer immunotherapy adverse events in real-world care — EricTopol · 2026-10-29