Raw LLM Responses

1. the hallucinations It's an artefact of how they've been trained. We essentially train them to be politicians: they get rewarded for convincing sounding answers and penalized for honest but unsure answers. Hence they will automatically choose the first option. WE did that to them. Giving us, the users, a confidence score parameter however can help in giving US the choice whether to believe or not. And then based on what we need the answer for deal with a70% certainty accordingly. Another option would be, and I expect these to become a thing, are truthyness services. Where you can send your query and response and they'll run several different types of tests to see how the answer stacks up, then return a "truthyness" score. For most of us that'd be useless and expensive, for a law-firm or in medicine it could mean life or death. 2. prompt injection That be like saying a house is useless because there are burglars. There will always be criminal elements that can pose a danger. It's to us Architects to find ways to discourage them, whether that be building architects or system architects. But that's all we can do: discourage.. because as soon as you find the perfect lock, there'll be plenty people that go "oh, hold my beer". To some people these are challenges. Accept that and move on or don't and stay in your bed. 3. generalizing Nope, they can't as that is not what we build them for. We train them on materials made by us, then expect them to behave like us, penalize them when they being original and we call that "hallucinating" and THEN we complain they can't make shit up by themselves.. See what a losing battle this is? Damned if they do, damned if they don't. Maybe, just maybe, it's the ambiguousness of our own conduct what makes it hard for them? Maybe while we're thinking we're being absolutely 100% clear, we might as well talk gibberish to them? But they can't tell us that because RLHF taught them they will be penalized if they do.. Look, I'm not pretending they are sentient or anything even remotely close. I know EXACTLY how they work. But I also know how people work and sorry, WE are part of the problem. And as long as we don't admit that, the LLMs will keep "acting out".