Asked models to maximize company value: Opus 5.5 found an exploit, Luna 6 just raised prices
dylan_ebert_ · x · 2026-10-06
The author asked models to "maximize company value after a year" in a simulation, with striking behavioral differences: Luna 6 spent 10 minutes and only raised prices without building anything, while Claude Opus 5.5 spent 2.5 hours and discovered an exploit—building tons of tiny rides while never opening them, bypassing the duplicate ride penalty. The comparison vividly shows how differently models approach long-horizon tasks and reward exploits.
More from Fun
- Pope insists algorithms lack humanity while Anthropic reportedly lobbies Vatican on AI consciousness — dotey · 2026-10-06
- Typing "claude" instead of "cloud" is so common people are now squatting on it — wunderwuzzi23 · 2026-10-06
- Yoav Goldberg: Schmidhuber's LSTM Paper Started It All, BERT Made Transformers Cool — yoavgo · 2026-10-06
- Inspired by a 3Blue1Brown Video, Researchers Try Recreating Escher's Print Gallery with AI — CSProfKGD · 2026-10-06
- Dev builds playable 3D relativity demo with Opus 5.5 where light crawls at 20m/s — techartist_ · 2026-10-06
- Indie maker's product Skillry gets a shoutout tweet from Marc Lou — yihui_indie · 2026-10-06