- cross-posted to:
- [email protected]
146
- cross-posted to:
- [email protected]
LLMs Will Always Hallucinate, and We Need to Live With This
arxiv.orgAs 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.
Well, there are some theoretical improvements laid out in papers. Not for hallucinations or the Tech Bro ish AGI dreams, but more adaptation, functional use, things like that.
…But the incredible thing is that the AI houses with the money seem to be ignoring them.
American firms seem to only pay attention to in-house innovations, like they have egos the size of the moon. And I’m only speaking of the ones not peddling the “scale transformers up infinitely” garbage.
Chinese LLMs tend to be open weights and more “functionally” oriented, which is great. But (with a few exceptions) they’re still pretty conservative with architectural experimentation, and increasingly falling into traps of following/copying others now.
Europe started out strong with Mistral (and the first good MoE!) and some other startups/initiatives, yet seems to have just… gone out to lunch? While still taking money.
And regions countries like South Korea or the Saudis are still pretty small scale.
What I’m saying is you are right, but it’s largely from an incredible amount of footgunning all the firms are doing. Otherwise models can be quite functional tools in many fields.
The point of “AI” is not making useful, functional software. Those technologies have existed for a long time and hopefully will continue to be developed by reasonable people.
The new “AI” is about creating useful, functional rubes to take their money. It’s obvious just from the phony name “AI”. If these grifters are shooting themselves in the foot, it doesn’t seem to stop them from walking to the bank.