Just a quick tip since I’ve seen a few posts recently with folks having really confrontational conversations with models and watching them spiral out.
At the heart of it all, LLMs are just token-prediction machines that will try to extend your conversation with the likely next words. They’re essentially theatre majors playing a very grand improv game.
Why is that important? Because the way you “treat” them matters enormously. Everything you say continues to modify the scene that’s playing out; every harsh criticism becomes self-fulfilling.
“You are useless at this, you keep making stupid mistakes!”
Everyone who’s used AI-assisted coding for more than a day or two will have felt the frustration. Maybe you’ve gone around in circles on a hard problem a few times (or even an easy one!). The temptation to say something harsh in frustration is strong. But when you do, you’ve changed the game.
These are not humans who will take the scolding and course-correct. Everything you say modifies the context and becomes ‘more true’. Telling the LLM it keeps making mistakes is like an instruction about its character.
It means the token prediction algorithm will now have that ‘fact’ in its context, and so it’ll respond in kind, ‘thinking’: “I am a useless AI that keeps making mistakes. Given that, what dumb thing would I be likely to say next?”
But it does make mistakes, what am I supposed to do about that?
Of course, you want to point out the errors so you can fix them and discourage repeat behaviour. My best advice: try to be nice about it. Treat your LLM like a (somewhat over-sensitive) junior dev, and encourage better behaviours instead of trying to punish unfavourable ones.
“You seem to have [done this incorrect thing]. Did you mean to? That will [cause these problems]. Let’s try to do better and not make that mistake again.”
Then the context contains the idea of it not doing the stupid thing, but striving to be better, and the prediction will continue along those lines. It’s a good way to avoid baking in the bad behaviour for the rest of your conversation.