Contested claims that Claude wasn't trained on Pokémon: burden of proof cuts the other way
sebkrier · x · 2026-11-12
Herbie Bradley pushes back on claims that Claude wasn't trained on Pokémon or Board Games, arguing RL environment companies sell exactly those environments and the burden of proof should fall on anyone claiming a sudden capability wasn't trained for. He also separates task fuzziness from feedback-loop length: long loops can be easy to verify (does a company profit) while short ones can be hard (writing quality) — both remain capabilities problems.
More from Models
- Ricursive Intelligence hosts first hackathon around open-weight Ripple-65B model — annadgoldie · 2026-11-12
- DSH ships v0.2.2-rc2 with 166 commits as community awaits new DeepSeek checkpoint — teortaxesTex · 2026-11-12
- Redditor stunned as Claude agent nails a complex work task almost flawlessly — johnnypen2445 · 2026-11-12
- Open-weight defender: you don't need SOTA models 97% of the time — gnukeith · 2026-11-12
- Model casually estimates C(734,15) at 7×10^30 in its head — true value 6.4×10^30 — Singularitarian · 2026-11-12
- OpenAI falls to 3rd on Epoch Capabilities Index as benchmark growth accelerates 1.9x — teortaxesTex · 2026-11-12