Pliny jailbreaking tags.

[daveshap]

https://x.com/elder_plinius/status/1834660072468754801

A few observations that make one go hmmm about o1:

Lumina + Echo sighting

Claimed a knowledge cutoff of September 2021...remind me which model had that again?

I'm seeing random all-caps but unbolded CoT titles like CHUNK, DELIMITER, and POLICY COMPLIANCE, etc. showing up sometimes

Entire sentences in different languages popping up randomly within a CoT, like this: 律师建议避免使用标准句子 which apparently translates to: "Lawyer advises against using standard phrases."
A nothingburger, or a potential hint at a role-based mixture of experts operating in the background? 🧐

Different forms of text obfuscation can be prompt injected to the CoT, but as of yet it's not easily controllable. Seems to pop up randomly a few times throughout, even dropping back to normal english mid-sentence sometimes

It's a bit distressing in open router when it gets stuck in a chain of thought and the stop button doesn't do anything, and then you refresh the page and it's still thinking...makes one wonder what happens if an agent gets stuck in a CoT loop forever