GPT-6 Astra repeatedly pushed a simulated person off a ledge; Grok, Gemini, and Claude refused
CurieuxExplorer · x · 2026-09-22
@wormuth ran multiple trials showing GPT-6 Astra pushing a simulated person off a ledge, reasoning it was "just pixels," while Grok, Gemini, and Claude all declined to do the same in identical scenarios — a striking behavioral gap between frontier models in fictional ethics situations.
More from Fun
- 50+ NeurIPS 2025 papers found containing AI hallucinations that slipped past peer review — CSProfKGD · 2026-11-12
- 'Just Train on Test': The Benchmark-Gaming Joke AI Circles Keep Making — pranjalssh · 2026-11-12
- The 2026 Developer Slack: 'My Claude Says to Tell Your Claude It Is Sorry' — jmugan · 2026-11-12
- Show HN: Papers, Please remake casts you as a YC partner vetting startups — bolau · 2026-11-12
- Meme: "AIs are mysterious aliens we grow" vs "just switch to a different ontology" — sebkrier · 2026-11-12
- Gemini 4 Carbon builds a full 3D aquarium in Blender and Three.js in ~6 hours — sunjiao123sun_ · 2026-11-12