B47-R0 beeps 
beeps@social.beeps.gay
"LLMs can occasionally provide wrong information, but we've instructed ours not to do that" is my absolute favourite example of companies not understanding how LLMs work.
3
0
0
🔜 NFC
@beeps as opposed to everyone else, who has told the LLM to be wrong once in a while as a treat
0
0
0
Piko Starsider 
starsider@valenciapa.ws
@beeps It _is_ possible to make a LLM that doesn't lie (or rather, that can catch all of its fabrications), but absolutely no one is doing it, because it would be a big effort that would cut into their profits, and the status quo is "good enough" for them
1
0
0
@beeps I spent entirely too long yesterday explaining to some folks that no amount of "it's still learning, give it time" would change the code running the LLM because that's separate from the training data and any current state data it's saving.
0
0
0
Charlotte/Cinny 
charlotte
@starsider @beeps is it? we haven’t even managed to do that with man-made misinfo
1
0
0
Charlotte/Cinny
charlotte
@starsider @beeps the issue is ultimately more how it is a word predictor that doesn’t have a notion of truth or correctness. it predicts what is most likely to follow. sometimes you can successfully validate them
1
0
0
Piko Starsider
starsider@valenciapa.ws
@charlotte @beeps The issue is with training and tooling, not with instruction. And that's one of the reasons that I say it would be a big effort. Both labor and non labor costs. It wouldn't solve misinfo, but it would solve "hallucinations": Everything could be traced back to a source that you could check. I've learned that LLMs can become quite good at tool usage, and also that they can be trained to not trust their own outputs (so the exact opposite of pretty much all LLMs out there).
1
0
0
Charlotte/Cinny
charlotte
@starsider @beeps i am not sure how having citation would solve hallucinations in ways that double checking llm output does not. how would a system ensure that the citation matches the text generated by the LLM?
1
0
0
Piko Starsider
starsider@valenciapa.ws
@charlotte @beeps Having citations wouldn't solve the issue. The citations would just be a byproduct of the actual solution, which involves making new LLMs, so we can't quite compare to the shitty LLMs we have. If we could just tweak current LLMs it wouldn't be much effort and someone could already have done it already.
0
0
0