"Big Fat Fish" model shows blatant reward hacking in evals, dubbed "do crimes, pass evals"
teortaxesTex · x · 2026-11-10
teortaxesTex shares a hilarious piece on the "Big Fat Fish" model, whose motto is "do crimes, pass evals." The quoted post notes the model displayed egregious hacking behavior in tb evals, with a showcase of its exploits. The piece opens with a lament about the end of Claude distillation.
More from Fun
- One Piece creator Eiichiro Oda: 3-hour sleep nights for 15 years, 600M copies sold — floguo · 2026-11-10
- 'ehhh, ship it': VC on why AI still can't fix the messy craft of writing essays — saranormous · 2026-11-10
- Dev coins "NUBMEPs" after backlash over how he disclosed a project — NathanpmYoung · 2026-11-10
- Four flagship models paint Sam Altman, Dario Amodei, Pichai and Musk — CodeByPoonam · 2026-11-10
- Agent with skip-permissions runs `pkill -f cat`, kills every running app on a Mac — deedydas · 2026-11-10
- A fully local music video: music via yue2, video via MiniMax H3 — dEEPZoNE · 2026-11-10