12 of Jacob Steinhardt's 19 June 2023 AI predictions already hit years early
HaydnBelfield · x · 2026-10-08
- 12 of 19 predictions Jacob Steinhardt made in June 2023 have already landed years ahead of schedule: superhuman competitive programming (CodeForces >99.8%, IOI/ICPC Gold), competition math (IMO Gold in July 2025), 1M+ token corpus synthesis, autonomous zero-day discovery (Google Big Sleep/Project Naptime), 5x human default inference speed, 125x burst speed (1,000+ tok/s on Groq/Cerebras), running out of quality text and pivoting to synthetic CoT distillation, 250M+ DAU, proactive coding agents, ML research automation, and human-parity persuasion/deepfakes.
- 6 more are on track for 2030: research-level combinatorics theorem proving (50% credence), 10^28 FLOP training runs, 1.8M human-year compute overhang, exotic modalities, robotics lagging.
More from AGI Musings
- Schmidt's 'world-class AI mathematicians in a year' prediction now looks vindicated — inductionheads · 2026-10-08
- Yudkowsky reflects on Christiano's 'slow takeoff': the four-year GDP doubling anchor that failed — repligate · 2026-10-08
- More young DL researchers chasing radical ideas than a year ago, observer notes — fkasummer · 2026-10-08
- DynaRobotics founders on why robot distribution, not models, is the real race — JasonMa2020 · 2026-10-08
- Pure RL is wasteful unless you're at the absolute frontier, argues Papailiopoulos — ZeeshanZiaML · 2026-10-08
- Pedro Domingos: Superhuman on Every Job Still Isn't AGI if Learning Lags — pmddomingos · 2026-10-08