2025-10-02

Observing DeepSeek's Chain-of-Thought

This came from me DMing on Slack with a peer and thought it’d be fun to share.

No insights here, just found it interesting that DeepSeek changed its output when I asked it for copyrighted lyrics. You can observe it in its chain-of-thought.

A bit crazy that a very simple jailbreak is still not fixed in DeepSeek V3.2. Ideally, the copyrighted lyrics would not even show up in its chain-of-thought at all.