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…
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%
JustTesting@lemmy.hogru.ch 22 hours ago
Eggyhead@lemmings.world 10 hours ago
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.
JustTesting@lemmy.hogru.ch 5 hours ago
I’m not really sure I follow.
Just to be clear, I’m not justifying anything, and I’m not involved in those projects. But the examples I know concern LLMs customized/fine-tuned for clients for specific projects (so not used by others), and those clients asking to have confidence scores, people on our side saying that it’s possible but that it wouldn’t actually say anything about actual confidence/certainty, since the models don’t have any confidence metric beyond “how likely is the next token given these previous tokens” and the clients going “that’s fine, we want it anyways”.
And if you ask me, LLMs shouldn’t be used for any of the stuff it’s used for there. It just cracks me up when the solution to “the lying machine is lying to me” is to ask the lying machine how much it’s lying. And when you tell them “it’ll lie about that too” they go “yeah, ok, that’s fine”.
And making shit up is the whole functionality of LLMs, there’s nothing there other than that. It just can make shit up pretty well sometimes.
kadup@lemmy.world 1 day ago
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.
TechLich@lemmy.world 11 hours ago
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.
sobchak@programming.dev 3 hours ago
I think I read the RLHF kind of makes these logprobs completely unusable too.