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.
  • msage@programming.dev
    11·
    9 hours ago

    You dense motherfucker.

    No LLMs are being developed in the open.

    Even provided weights mean nothing.

    It’s not knowledge LLMs retain, just the ingressed text.

    LLMs should be skipped after confirming that they are indeed a dead end they always were. And the entire world should focus on anything else.

    • DigitalStefan@fosstodon.org
      1·
      8 hours ago

      @msage @yogthos I don’t know if I agree 100% with this, but I do like what you’re saying.

      It seems like all the AI companies are simply hoping AGI emerges from it and nobody is doing the actual research to make that happen.

      People were researching it when I was a child and I suspect they’ll still be researching it when I’m collecting my pension.

    • ☆ Yσɠƚԋσʂ ☆@lemmy.ml
      1·
      4 hours ago

      You’re such an angry little ignoramus. The GPT-NeoX repo on GitHub is the actual codebase they used to train these models. They also open-sourced the training data, checkpoints, and all the tools.

      However, even if you were right that the weights were worthless, which they’re obviously not, and there were no open projects which there are, the solution would be to develop models from scratch in the open instead of screeching at people and pretending this tech is just going to go away because it offends you personally.

      And nobody says LLMs are anything other than Markov chains at a fundamental level. However, just like Markov chains themselves, they have plenty of real world uses. Some very obvious ones include doing translations, generating subtitles, doing text to speech, and describing images for visually impaired. There are plenty of other uses for these tools.

      I love how you presumed to know better than the entire world what technology to focus on. The megalomania is absolutely hilarious. Like all these researchers can’t understand that this tech is a dead end, it takes the brilliant mind of some lemmy troll to figure it out. I’m sure your mommy tells you you’re very special every day.