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.”
They keep saying it’s impossible, when the truth is it’s just expensive.
That’s why they wont do it.
You could only train AI with good sources (scientific literature, not social media) and then pay experts to talk with the AI for long periods of time, giving feedback directly to the AI.
Essentially, if you want a smart AI you need to send it to college, not drop it off at the mall unsupervised for 22 years and hope for the best when you pick it back up.
I’m addition to the other comment, I’ll add that just because you train the AI on good and correct sources of information, it still doesn’t necessarily mean that it will give you a correct answer all the time. It’s more likely, but not ensured.
Yes, thank you! I think this should be written in capitals somewhere so that people could understand it quicker. The answers are not wrong or right on purpose. LLMs don’t have any way of distinguishing between the two.
That’s just not how LLMs work, bud. It doesn’t have understanding to improve, it just munges the most likely word next in line. It, as a technology, won’t advance past that level of accuracy until it’s a completely different approach.
They could also perform some additional iterations with other models on the result to verify it, or even to enrich it; but we come back to the issue of costs.
Why not solve it before training the AI?
Simply make it clear that this tech is experimental, then provide sources and context with every result. People can make their own assessment.
I think you’re right that with sufficient curation and highly structured monitoring and feedback, these problems could be much improved.
I just think that to prepare an AI, in such a way, to answer any question reliably and usefully would require more human resources than there are elementary particles in the universe. We would be better off connecting live college educated human operators to Google search to individually assist people.
So I don’t know how helpful it is to say “it’s just expensive” when the entire point of AI is to be lower cost than a battalion of humans.