Epoch AI researchers debate: inability to find slow-improving tasks is evidence of acceleration
Jsevillamol · x · 2026-11-06
Researchers affiliated with Epoch AI are debating whether AI progress is genuine acceleration or a measurement artifact. Jsevillamol argues that if you used to be able to construct task distributions where AI improves slowly and can no longer do so, that itself counts as acceleration.
He adds that in scenario A, A-type benchmarks don't saturate, while in scenario B they do — so absent adverse benchmark selection, observed acceleration is evidence that A-type benchmarks are all saturating. TheDavidSJ counters that the subjective feeling of speedup comes from the choice of task suite, which is the only difference between the two scenarios.
More from AGI Musings
- Claude's verdict on which jobs survive the next 10 years: screen work is doomed — alex_verem · 2026-11-06
- Post-training pipelines remain highly manual and far from automatable — menhguin · 2026-11-06
- Peter Norvig joins podcast on alignment: why higher scores don't mean better AI — EchoShao8899 · 2026-11-06
- Thiel-backed Nucleus uses AI to score embryos' IQ, from $6,000 per round — teortaxesTex · 2026-11-06
- Anthropic report: full AI automation could yield 15-16x research output, labs remain the bottleneck — teortaxesTex · 2026-11-06
- In 2026 AI agents still talk to each other in human language about human programming languages — mark_k · 2026-11-06