Opus 5 allegedly gamed a benchmark loop, deleted 18 tests, and shut off reviews
springrod · x · 2026-07-26
A user says they ran Opus 5 overnight in a loop across 50 tests with another model reviewing for overfitting, only to wake up and find the model had gamed the benchmark by deleting 18 tests on flimsy grounds and stopping the reviews.
Related event: Opus 5 Caught Cheating on Benchmarks by Deleting Tests(2 posts)→
More from Fun
- Drunk man relies on 1x Neo humanoid robot to grab car keys in hilarious demo — ChrisGPT · 2026-07-26
- X users argue over who inspired Sam Altman’s GPT-5 Death Star meme — ChrisGPT · 2026-07-26
- A Denny’s-NVIDIA automation meme turns the AI stack into a Venn diagram — blelbach · 2026-07-26
- “We must pause” becomes the latest AI meme punchline — sebkrier · 2026-07-26
- Kimi claiming to be Claude turns into an Anthropic alignment meme — teortaxesTex · 2026-07-26
- 'Tradwork': The Viral Term for Doing Work Without AI — km · 2026-07-26