I dislike the #AI #llm term "jailbreak" for the same reason many folks dislike "hallucination." Like it it not, they are...

I dislike the #AI #llm term "jailbreak" for the same reason many folks dislike "hallucination." Like it it not, they are both "terms of art" in the space, so we must deal with them.But it's helpful to know what they mean. Broadly, "hallucination" is a non factual response, and "jailbreak" is an undesirable (to the creator) behavior.But what both miss capturing, is that both are the system operating as designed. Possibly not as intended, but as designed.Next probable token. That's it.

Read Original

Related