> The model has no concept of truth—only of plausibility.

This is such an important problem to solve, and it feels soluble. Perhaps a layer with heavily biased weights, trained on carefully curated definitional data. If we could train in a sense of truth - even a small one - many of the hallucinatory patterns disappear.

Hats off to the curl maintainers. You are the xkcd jenga block at the base.

▲

jcattle 3 days ago | parent | next [-]

I am assuming that millions of dollars have already been spent trying to get LLMs to hallucinate less.

Even if Problems feel soluble, they often aren't. You might have to invent an entirely new paradigm of text generation to solve the hallucination problem. Or it could be the Collatz Conjecture of LLMs, that it "feels" so possible, but you never really get there.

▲

big-and-small 3 days ago | parent [-]

Nuclear fusion was always 30 years away (c)

▲

quikoa 3 days ago | parent [-]

It would be nice if nuclear fusion had the AI budget.

	▲	Cthulhu_ 3 days ago \| parent [-]
		Fusion will at best have a few dozen sales once it's commercially viable and then take decades to realise, but you can sell AI stuff to millions of customers for $20 / month each and do it today.

▲

pjc50 3 days ago | parent | prev | next [-]

The "fact database" is the old AI solution, e.g. Cycorp; it doesn't quite work either. Knowing what is true is a really hard, unsolved problem in philosophy, see e.g. https://en.wikipedia.org/wiki/Gettier_problem . The secret to modern AI is just to skip that and replace unsolvable epistemology with "LGTM", then sell it to investors.

	▲	pksebben 3 days ago \| parent [-]
		There are some things that we can define as "definitely true as close as makes no difference" in the context of an LLM: - dictionary definitions - stable apis for specific versions of software - mathematical proofs - anything else that is true by definition rather than evidence-based (i realize that some of these are not actually as stable over time as they might seem, but they ought to do good enough with the pace that we train new models at). If you even just had an MOE component whose only job was verifying validity against this dataset in chain-of-thought I bet you'd get some mileage out of it.

▲

wongarsu 3 days ago | parent | prev [-]

Truth comes from being able to test your assertions. Without that they remain in the realm of plausibility. You can't get from plausibility to truth with better training data, you need to give LLMs better tools to test the truth of their plausible statements before spewing them to the user (and train the models to use them, obviously. But that's not the hard part).