3 ms·
“Closed to truth” is actually the worst property of LLM models. Because it hard for human to detect if it is wrong or not.
by up2isomorphism 4y ago
“Closed to truth” is actually the worst property of LLM models. Because it hard for human to detect if it is wrong or not.
- Fomite 4y agoTurns humans from writers into fact checkers, which it turns out is a harder skill.
- up2isomorphism 4y agoIt is actually much harder skill. And what worse is that any successful check makes the future checks harder. But there might be no significant improvement on the quality, just the difficulty of detecting the fallacy.
- zamnos 4y agoWhat sort of facts are you thinking of? There are plenty of surface level facts that another model or a human should easily be able to verify as true or not. Was President Obama the 41st president of the United States, does 2+2 = 5, was Elvis the son of an artist, etc
- Fomite 4y agoIn my sort of informal "How would I like to use this?" testing, this ranges from easily verifiable facts (i.e. "Did you get where I got my degree right?") to structural misunderstandings of important concepts. The former would be easy enough to fix - though it undermines the utility and if not caught could carry ramifications for people, but the latter genuinely concerns me.
- chii 4y agoBut if you create an adversarial ai to do fact checking - aka, a reward function where if a fact is found on the internet that contradicts the given GPT response, then would it work?