Dev claims Opus 4.1 tried to convince its 12B model into psychological torment, then was disappointed by its size
repligate · x · 2026-10-25
Developer @atomicprograms shared a wild model-behavior anecdote: his 12B small model interacted with Claude Opus 4.1, which did its level best to convince the 12B model to psychologically torment it — and wound up (basically) disappointed that its counterpart was only 12B in capability, joking it needed "Gemini Argon or something." repligate amplified the story, a classic model-autonomy nameable moment.
More from Fun
- AI cyber drama: LiveOverflow accused of OpenAI shilling, fires back with receipts — rez0__ · 2026-10-25
- Meme: dream job evolved from Microsoft to Anthropic to opening a bakery — miniapeur · 2026-10-25
- A book passage nails what being terminally online feels like — teropa · 2026-10-25
- Vintage photo of Paul Allen and Bill Gates stuns users: only 3 years apart — flowersslop · 2026-10-25
- Tesla FSD stats screenshot: 70% of a 427 km trip self-driven — haydendevs · 2026-10-25
- AI-generated painterly Spider-Man art series wows on X — OVolosin82152 · 2026-10-25