[–] 2 points 7 hours ago (1 child)

I always hate how people misinterpret the whole probability thing in quantum mechanics, or think that Schrödinger's Cat was meant to prove something rather than the fact that it was to show the ridiculousness of some of Heisenberg's work, among others.

The real issue is that the act of observing a quantum particle in any way that a human or their machines could possibly observe them necessarily alters the state of that particle in some way. It's like if the only way you could detect how far away a bullet sitting on the ground was from you was to fire another bullet at it and have it reflected back. Obviously whatever reading that were to give you is now no longer correct, the bullet has been moved by the detection bullet and possibly some other characteristics have changed. So you only do real detection at the moment you need to and realize that value is only valid in that moment and for everything else you use probabilities.

Of course an oversimplification, but pretending there's something surreal and supernatural about quantum mechanics is just silly.

  • source
  • [–] 4 points 3 days ago

    To test whether the software accurately reported what its sources said, the researchers broke the overviews down into 98,020 individual factual claims. They built an automated verification pipeline, powered by a separate large language model, that read the text of the cited websites.

    Uhh...

    Maybe you shouldn't have the hallucination machine verify the output of the hallucination machine?

  • source
  • [–] 9 points 4 days ago

    It doesn't 'know' anything, and getting it to, "triple check" itself is merely reinforcing those terms in the token space. These LLMs do not even begin to have the capacity to actually check their sources.

    The only way it does work is to add extra algorithms that check correlation between output and source, and even then, you're still back in the realm of word weight association.

    TL;DR - LLMs literally cannot do this.

  • source
  • parent
  • context
  • [–] 5 points 4 days ago (1 child)

    I'm surprised it's only 10%. I remember a couple of years ago I worked on a team trying to build an AI tool that, as a feature, could cite sources. I was skeptical then because gemini couldn't reliably cite correct sources on a google search, so what made our team so confident the OpenAI api version could do that? I mean, it still can't because it doesn't know where the information is coming from, and I'm sure getting it to double and triple check its work to get it right leads to it getting stuck in cycles.

    The point is you need to treat AI "facts" as if your trivia buddy told you a fun fact. It sounds right and you trust that they looked it up at some point, but they're confidently incorrect at least a couple of times a trivia night so you shouldn't put money on it without fact checking yourself.

  • source
  • view more: next ›