Scaling01: the .0 iterations always suck, the .5 versions are always the best
scaling01 · x · 2026-09-23
scaling01 doubles down on an old observation he calls a "cosmological constant": a model's .0 debut iteration always underperforms, while the .5 revision is consistently the best release — a widely recognized pattern in how AI labs pace their model iterations.
More from Fun
- Claude Opus 5.5 generates a 10v10 Halo-style multiplayer game from one prompt — Scobleizer · 2026-09-23
- Opus 5.5 Builds a Playable 3MB Single-File Game Inspired by the Antikythera Mechanism — edwinarbus · 2026-09-23
- Ilya Sutskever's weekly test: will Muse finally conquer the org? — ilyasu · 2026-09-23
- Replit CEO revisits 2025 spat with AI safety critic Liron over superintelligence bets — ctjlewis · 2026-09-23
- 'Opus 5.5 made me curse at Claude less' — an unofficial new benchmark — alliekmiller · 2026-09-23
- Clarke's new law: sufficiently advanced incompetence is indistinguishable from malice — burny_tech · 2026-09-23