The Chonkerton

Opus 5 Glitch Text

ai

Per LessWrong, a researcher discovered Claude Opus 5 exhibits unusual behavior when given certain incomplete prompts: the model generates content and attributes it to the user's prompt, functioning like a base model in text-completion mode. Qualifiers like 'see the proof below' trigger this effect, producing mathematical proofs and other content. The post suggests this reveals how Claude models user inputs and prompt completion, and notes that older base-model prompting techniques appear effective against this behavior.

Source: https://www.lesswrong.com/posts/ZSBse2fyftgHiJ3Kq/opus-5-...

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton