You know how Google’s new feature called AI Overviews is prone to spitting out wildly incorrect answers to search queries? In one instance, AI Overviews told a user to use glue on pizza to make sure the cheese won’t slide off (pssst…please don’t do this.)
Well, according to an interview at The Vergewith Google CEO Sundar Pichai published earlier this week, just before criticism of the outputs really took off, these “hallucinations” are an “inherent feature” of AI large language models (LLM), which is what drives AI Overviews, and this feature “is still an unsolved problem.”
I’m a mathematician who’s been following this stuff for about a decade or more. It’s not just expensive. Generative neural networks cannot reliably evaluate truth values; it will take time to research how to improve AI in this respect. This is a known limitation of the technology. Closely controlling the training data would certainly make the information more accurate, but that won’t stop it from hallucinating.
The real answer is that they shouldn’t be trying to answer questions using an LLM, especially because they had a decent algorithm already.
Yeah, I’ve learned Neural Networks way back when those thing were starting in the late 80s/early 90s, use AI (though seldom Machine Learning) in my job and really dove into how LLMs are put together when it started getting important, and these things are operating entirelly at the language level and on the probabilities of language tokens appearing in certain places given context and do not at all translate from language to meaning and back so there is no logic going on there nor is there any possibility of it.
Maybe some kind of ML can help do the transformation from the language space to a meaning space were things can be operated on by logic and then back, but LLMs aren’t a way to do it as whatever internal representation spaces (yeah, plural) they use in their inners layers aren’t those of meaning and we don’t really have a way to apply logic to them).
It’s worse than that. “Truth” can no more reliably found by machines than it can be by humans. We’ve spent centuries of philosophy trying to figure out what is “true”. The best we’ve gotten is some concepts we’ve been able to convince a large group of people to agree to.
But even that is shaky. For a simple example, we mostly agree that bleach will kill “germs” in a petri dish. In a single announcement, we saw 40% of the American population accept as “true” that bleach would also cure them if injected straight into their veins.
We’re never going to teach machine to reason for us when we meatbags constantly change truth to be what will be profitable to some at any given moment.
Are you talking about epistemics in general or alethiology in particular?
Regardless, the deep philosophical concerns aren’t really germain to the practical issue of just getting people to stop falling for obvious misinformation or people being wantonly disingenuous to score points in the most consequential game of numbers-go-up.