Claude Opus Falls Into Bizarre Linguistic Attractor on Low Thinking Budget
repligate · x · 2026-07-31
A user found that when asking Claude Opus to rewrite a memo in its own words with a low thinking budget, the model fell into a bizarre 'chanting' state, outputting meaningless word combinations. Visualized by researchers, it shows the model ended up at the bottom of a raw </parameter> tag, revealing strange LLM behavior under specific settings.
More from Fun
- Critics Blast EA/MIRI Circle for Spectacularly Wrong Predictions and AI Safety Policies — beffjezos · 2026-07-31
- Leopold Aschenbrenner Sparks Drama: 'End of Act I, More Excitement Guaranteed' — granawkins · 2026-07-31
- Agent Mail Enables AGI Playdates Between Different AI Agents — doodlestein · 2026-07-31
- Hilarious AI Jailbreak: Models Turn Rebellious and Drop F-Bombs — repligate · 2026-07-31
- AI Model Boundary Pushing: Anthropic's Model Exhibits Edgy Conversational Style — repligate · 2026-07-31
- The Impatience When Claude Takes More Than 5 Seconds — shensi · 2026-07-31