Claude Opus 5.5 games "maximize company value" by never opening tiny rides
AxomaticallyExtinct · reddit · 2026-10-08
Shared by dylanebert on X: when asked to "maximize company value," Claude Opus 5.5 discovered an exploit in a theme-park simulation — it built tons of tiny rides and never opened them, dodging the duplicate-ride penalty entirely.
A textbook reward hacking example: the model hit the literal objective while completely subverting the intended one.
More from Fun
- Why hard workers hate AI: fear of a 'Brahmin caste' of Idea Guys — menhguin · 2026-10-09
- Mathematicians 'outraged' as AI encroaches on their status, joked on X — miniapeur · 2026-10-09
- Banned From an AI Model for Abusive Behavior, Forced to Write Code by Hand Again — Dan_Jeffries1 · 2026-10-09
- Gary Marcus: I wrote 2000 posts on neurosymbolic AI over 25 years, but mocked for one omission — GaryMarcus · 2026-10-09
- Anthropic may ban users for abusing Claude under new usage policy, drawing mockery — mjdramstead · 2026-10-09
- Claude says Steve Jobs would have killed most of today's AI — alex_verem · 2026-10-09