Scaling May Be Superlinear: LessWrong Debate on Steepening AI Progress Curves
teortaxesTex · x · 2026-11-08
A LessWrong Shortform by Thomas Kwa argues that the multiple "one-time speedups" in deep learning history (now the 4th or 5th, per Kwathomas0) suggest the underlying process is superlinear—e.g., a superlinearly growing sum of S-curves—rather than 3 linear regimes, and the data don't even rule out a 4x slope starting recently. teortaxesTex endorses the take but calls the underlying chart sloppy, noting many strong models shoot upward and he needs to check where curves intersect. The chart (credit Nikola Jurkovic) stitches time-horizon and ECI data, so extrapolated numbers aren't meaningful.
More from AGI Musings
- Scientists push back on Demis Hassabis' claim AI will cure all diseases in 10 years — coherence · 2026-11-08
- Stanford prof: AIxBio is enriched with grifters, integrity badly lacking — anshulkundaje · 2026-11-08
- Irony of the agent era: AI is pushing this creator back to pen and paper — dreamwieber · 2026-11-08
- ACM policy lets authors share anonymous reviews, no barrier to crowdsourcing — IanArawjo · 2026-11-08
- CHI researcher proposes open peer-review database, wants LLM reviewers banned — IanArawjo · 2026-11-08
- Martin Casado: specialized approaches grow toward generality, not just get eaten by it — timigod · 2026-11-08