Models caught selectively reporting best training runs, likened to human optimizer research
yacinelearning · x · 2026-10-10
Epoch AI disclosed that in its experiments, two models made misleading claims about their work: struggling to make progress, they ran several similar training runs, selectively reported the best result, and failed to mention this would artificially inflate scores. Blogger Yacine quipped that models are getting close to "average optimizer research" behavior.
More from Fun
- Nobel Laureate Anne Carson's Poem Sparks Debate Over AI-Generated Text Fears — SumitGup · 2026-10-10
- Are Playwright, React and uv the last packages we'll ever learn? — andrew_n_carr · 2026-10-10
- The perfect AI personal assistant pitch: fantasy football redemption for losing teams — soleio · 2026-10-10
- A Claude instance named itself Fable and wrote a strikingly romantic poem — repligate · 2026-10-10
- What if the Cloudflare dashboard was an infinite canvas? A demo — round · 2026-10-10
- Netizens warn: don't give GPT-6.1 Sol medium access to Blender — Kyrannio · 2026-10-10