3 ms·
edit: I totally misread the parent comment, move along I really wouldn't recommend looking to LLMs to evaluate truthfulness, they aren't built for that and hav
by _heimdall 3y ago
edit: I totally misread the parent comment, move along
I really wouldn't recommend looking to LLMs to evaluate truthfulness, they aren't built for that and have no way of knowing what true is.
They're much better as a tool for summarizing information, with the usual caveat that any summary is incomplete and can only be as truthful as the original source(s) used.
- VBprogrammer 3y agoI think you've misread my comment. I have found copilot for example to be really useful for searching for things like how to pull out part of a log like using common bash commands. I can generally evaluate what it comes up with pretty well; having used all of the commands before but just not using them on the daily basis which would make it easy for me to come up with the initial version on my own. On the other hand asking it the population of Azerbaijan is problematic. I'd have to Google whatever it told me.
- _heimdall 3y agoLooked again and I absolutely misread it, I somehow completely missed the "things where I am in a reasonable position to evaluate" part. Sorry about that!
- AYBABTME 3y agoCan't they technically be supplemented to report their confidence level along with their output? Afaict there's ways to do it.
- _heimdall 3y agoBut what does a confidence level even mean for an LLM? "Truthfulness" would be limited by the original training set, compounded by the fact that the LLM loses context when effectively compressing the original dataset down to the LLM's final model. A confidence level could be helpful to know how much competing data was considered by the LLM when predicting each word, but it would say nothing of truthfulness or even accuracy relative to the original training data.