Yeah.

I sorta misread your post, these bots can indeed be twisted, or “jailbroken” during conversation, to a pretty extreme extent. The error is assuming they are objective in the first place, I suppose.

Base models are extremely interesting to play with, as they haven’t been tuned for conversation or anything. They do only one thing: complete text blocks, thats it, and it is fascinating to see how totally “raw” LLMs tuned on a jumble of data (before any kind of alignment) guess how things should be completed.