LLMs Will Always Hallucinate, and We Need to Live With This
arxiv.org
As Large Language Models become more ubiquitous across domains, it becomes important to examine their inherent limitations critically. This work argues that hallucinations in language models are not just occasional errors but an inevitable feature of these systems. We demonstrate that hallucinations stem from the fundamental mathematical and logical structure of LLMs. It is, therefore, impossible to eliminate them through architectural improvements, dataset enhancements, or fact-checking mechanisms. Our analysis draws on computational theory and Godel's First Incompleteness Theorem, which references the undecidability of problems like the Halting, Emptiness, and Acceptance Problems. We demonstrate that every stage of the LLM process-from training data compilation to fact retrieval, intent classification, and text generation-will have a non-zero probability of producing hallucinations. This work introduces the concept of Structural Hallucination as an intrinsic nature of these systems. By establishing the mathematical certainty of hallucinations, we challenge the prevailing notion that they can be fully mitigated.
  • aesthelete@lemmy.world
    41·
    4 days ago

    I think these things are maybe more useful (in very narrow cases) when people realize a little about how they work. They’re probabilistic. So they’re great at bullshit. They’re also great at bullshit adjacent things (quarterly goal documents for your employer for instance). Knowing they’re probabilistic makes me treat them differently. For instance, I wanted to know what questions I should ask about something, and so I used Google AI insights or whatever from their search engine to generate a large list by simply resubmitting the same question over and over.

    It’s great at that kind of (often extremely useless) junk. You can get lots of subtle permutations out of it. It also might be interesting to just continually regen images using the same prompt over and over and look at the slight differences.

    It would be more interesting to me instead of bullshit like Sora if they made something that just gave you the prompts in a feed and allowed you to sit there and regenerate the junk by hitting a button. People could see the same post and a slightly different video every time. Or image. Still stupid? Yes. Still not worth slurping up our lakes for? Yes. But hey at least it’d be a little more fun.

    The prompts are also, for the most part, the only creative thing involved in this garbage.

    Instead of the current knobgobblers that want to take a single permutation and try to make it more than worthless, or want to pretend these systems are anything close to right…or intelligent…or human or whatever it’d be much better if we started thinking about them for what they are.