Your job as a scientist is disproving your own results: baselines, leakage, reward hacking
tw_killian · x · 2026-11-04
A widely shared research-methodology rant: before celebrating a rising line, check for undertuned baselines, data leakage, reward hacking, prompt-format bugs in baselines, whether the benchmark measures what you claim, and whether your method just uses more compute—miss one and 'you become a clown.'
More from AGI Musings
- AI safety researcher Jeff Ladish warns RSI is nearing and scarier than any mountain lion — JeffLadish · 2026-11-04
- Diamandis: a household robot saving 10 hours a week — 520 hours a year — is the robot economy people care about — PeterDiamandis · 2026-11-04
- Digital Brain Project scales to $15M to open-source neural data and model the human brain — ShahabBakht · 2026-11-04
- AI researcher: evidence for fast intelligence explosion is thin — diffusion bottlenecks may bind first — pwlot · 2026-11-04
- Dario Amodei and Demis Hassabis both put AGI at 2026-2027 in candid exchange — generativist · 2026-11-04
- OpenAI's Liam Fedus: aim AI optimization at science, not bounded fields like accounting — LiamFedus · 2026-11-04