Two task families: counterexample suggests AI acceleration may be a scale artifact
TheDavidSJ · x · 2026-11-06
In the debate over whether AI acceleration is real, TheDavidSJ presents a technical counterexample (see image) with two task families:
- On family A, a model's log-odds of success are linear in skill.
- On family B, they're linear in a smooth convex function of skill — each extra unit of skill gains about 65% more log-odds on B than the previous unit.
- The two families just measure progress on differently curved scales; neither changes at any date, and neither is the "true" one.
The point: perceived acceleration may be an artifact of scale choice. Yafah Edelman conceded after reading closely that eci increases should indeed be comparable across the scale.
More from Research
- Lightwheel and Synopsys bring CAE-grade contact physics into robot simulation — lukas_m_ziegler · 2026-11-06
- Lecture derives R-GCN's 4.75M-parameter bill and a 14.7× basis-decomposition fix — JeremyCMorgan · 2026-11-06
- Random steering vectors turn LLMs into teapots, integers and 2019 movies — Sauers_ · 2026-11-06
- Epoch AI review: OSWorld tasks are simple, unstable, and barely test GUI use — gleech · 2026-11-06
- One random vector in 240 makes Qwen3-30B believe it's a cat via MoE steering — Sauers_ · 2026-11-06
- Steering a single MoE expert at strength +1.25 rewires Qwen3-30B's persona — Sauers_ · 2026-11-06