Why RL lacks scaling laws: pretraining optimizes loss, open-ended maximization the lightcone
akbirthko · x · 2026-10-15
The author argues that RL "scaling laws" are hard to pin down because pretraining optimizes a well-defined scalar (loss), whereas open-ended maximization effectively targets consequences across the entire lightcone — an objective with no clear bound or metric. The accompanying Dyson-sphere-style artwork echoes the cosmic-scale metaphor.
More from AGI Musings
- PMs should build more themselves: how agents collapsed weeks of design into hours — thechrisperry · 2026-10-15
- Memia #2026.41: AI agents scan 3,956 articles to curate 323 frontier signals — ben_r · 2026-10-15
- Memia #2026.40: Gemini 4 leaks, protein watermarking, AI red phone and open USD — ben_r · 2026-10-15
- Memia #2026.39: 100 of 318 AI stories reduce to 8 narratives, says Ben Reid — ben_r · 2026-10-15
- Steve Hsu: AI doomerism discourse overlooks its real casualty — young people — StringChaos · 2026-10-15
- The math community is splitting over AI: science-first vs ego-first camps — haider1 · 2026-10-15