A friend of mine prepared an arsenal of hooks and the like to address this and LLMs still disobey them at times.

I don't have high hopes that there exists a bulletproof solution to this.

A friend of mine prepared an arsenal of hooks and the like to address this and LLMs still disobey them at times.

It's a model of language, yes? Trained on a big corpus of text.

I have read a lot of stories and accounts in which people were told not to do something and inevitably they did it. Like, lots. Far more than stories and accounts in which people were told not to do something and they then didn't do it.

If I'm reading a story or account of something, and it's really hammered home that they've been told not to do something, it's kind of inevitable that they will then do that. I'm not even an LLM and I noticed that's the way these things usually go.

So is an LLM just doing what it's been trained to do? Sometimes in the stories and accounts, there's a whole lot of time and tension before the bad thing happens, but that's just part of the fun.

Personally, for me that represents job security. Having a human with a high level of domain knowledge in the loop seems pretty required to get any meaningful results.

solution is to always do what it has seen in training data and how it was RL. But ai companies dont tell you that. so you have to reverse engineer its training and stick to to that.

These are no general purpose machines. They are shipping a subset mindset not general intelligence like they want us to belive .