OpenAI Researcher: Downstream Capability Scales Roughly Linearly, Not Log
willdepue · x · 2026-11-08
Responding to a complaint that the original scaling laws paper's log-log plots made people wrongly assume a log relationship, OpenAI's will depue argues that while loss follows log-log scaling, downstream utility is not log-log at all — it is roughly linear (log-linear) in capability. This bears directly on debates over whether continued scaling is worth the compute.
More from Research
- New tool catches authors lying in their AI disclosure statements on arXiv — danish037 · 2026-11-08
- Fleuret vouches for Bigstate: 1.5B model jumps from 26.8% to 37.2% on GSM8K — francoisfleuret · 2026-11-08
- Robotics olfaction may spawn new continually-learning algorithms, argues Kordem France — AnnaCiaunica · 2026-11-08
- SnapJudge, an open-source AI detector for scientific text, checks 25K+ submissions in 3 weeks — danish037 · 2026-11-08
- Trained "oracle" model outputs weight edits that undo fine-tuning, closing 61% of the KL gap on Qwen3-8B — Sauers_ · 2026-11-08
- Trained 35 domain LoRAs on a MoE with MLX on one Mac Studio — the base model beat most of them — Holiday_Poetry_5133 · 2026-11-08