“I literally lost my only friend overnight with no warning,” one person posted on Reddit, lamenting that the bot now speaks in clipped, utilitarian sentences. “The fact it shifted overnight feels like losing a piece of stability, solace, and love.”
https://www.reddit.com/r/ChatGPT/comments/1mkumyz/i_lost_my_only_friend_overnight/
LLMs don’t have any awareness of their internal state, so there’s no way for them to see something as a gap of knowledge.
Took me ages to understand this. I’d thought "If an AI doesn’t know something, why not just say so?“
The answer is: that wouldn’t make sense because an LLM doesn’t know ANYTHING
Wouldn’t it make sense for an ai to provide a confidence level though?
I’ve got 3 million bits of info on this topic but only 4 of them lead to this solution. Confidence level =1.5%
It’s always funny to me when people do add ‘confidence scores’ to LLMs, because it always amounts to just adding ‘say how confident you are with low, medium or high in your response’ to th prompt, and then you have made up confidences for made up replies. And you can tell clients that it’s just made up and not actual confidence, but they will insist that they need it anyways…
That doesn’t justify flat out making shit up to everyone else, though. If a client is told information is made up but they use it anyway, that’s on the client. Although I’d argue that an LLM shouldn’t be in the business of making shit up unless specifically instructed to do so by the client.
It doesn’t have “3 million bits of info” on a specific topic, or even if it did, it wouldn’t be able to directly measure it. It’s worth reading a bit about how LLMs work behind the hood, because although somewhat dense if you’re new to the concepts, you come out knowing a lot more about what to expect when using them, what the limitations actually are and how to use them better if you decide to go that route.
You could do this with logprobs. The language model itself has basically no real insight into its confidence but there’s more that you can get out of the model besides just the text.
The problem is that those probabilities are really “how confident are you that this text should come next in this conversation” not “how confident are you that this text is true/accurate.” It’s a fundamental limitation at the moment I think.