LLMs are good at playing you
lemmy.ml
Large language models (LLMs) are eerily human-like: in casual conversations, they mimic humans with near-perfect fidelity. Their language capabilities hold promise for some fields — and spell trouble for others. But above all, the models’ apparent intellect makes us ponder the fate of humanity. I don’t know what the future holds, but I think it helps to understand how often the models simply mess with our heads.

TL;DR: LLMs are just mimicking natural language and conversation. Fact checking and healthy skepticism is not part of their model. For example they can be easily tricked into advocating conspiracy theories, like a fake moon landing. Google Bard is even stating arithmetic falsehoods like 5*6 != 30

  • Peanut@sopuli.xyzEnglish
    21·
    1 year ago

    not sure what you’re saying here. are you claiming it can’t do any sort of reasoning or open-ended problem solving?

    i think we’re fairly confident now that they can do structured reasoning to some degree. it is not flawless in that it might not give you real or accurate information every time, but we are also figuring out the contexts behind that. as for spreading misinformation, anything intentional prompted to be incorrect is irrelevant to gauging intelligence. unintentional results don’t necessarily mean it’s unintelligent either.

    there’s a really good document on this aspect as well.

    https://www.lesswrong.com/posts/D7PumeYTDPfBTp3i7/the-waluigi-effect-mega-post

    there are a lot of ethical and technical aspects of LLMs that are severely underdeveloped, but that shouldn’t be a surprise to anyone. i don’t think any of that would suggest that it’s reasonable to disregard the absurd pace of development this past decade, and last few years especially. good thing we have a sudden surge of attention towards developing these things.

    • smegforbrainsOPEnglish
      1·
      1 year ago

      not sure what you’re saying here. are you claiming it can’t do any sort of reasoning or open-ended problem solving?

      It’s right there in the title mate.