Mustafa Suleyman

And what we're trying to do is figure out from a query, like given the thing that you're asking the model, should we put it into a more creative mode or a more like precise and specific mode that's more likely to be accurate?

And so sometimes you want it to be creative.

You're actually looking for it to...

find the connection between a zebra and a lemon and a New York taxi, right?

And if you ask it that question, it's going to come up with some creative, poetic connection that links those three things.

And that's kind of the hallucination challenge.

Well, narrow AI is the kind of conversational companions that I was describing.

But then sometimes you just want it to be super to the point.

and not wander off and talk all kinds of, you know, creative poetic nonsense.

Like, it's limited.

It doesn't improve on its own.

So, you know, you sort of train it within some boundaries, and then it gets good at a specific number of tasks.

You just wanted to give you the facts.

So that's a spectrum.

And what we're trying to do is figure out from a query, like given the thing that you're asking the model, should we put it into a more creative mode or a more like precise and specific mode that's more likely to be accurate?

Like, it's good at great conversation.

Maybe it can do document understanding.

Maybe it can even generate a little bit of code, but we know what its capabilities are to some extent.

And that's kind of the hallucination challenge.

That's a narrow version of AI when we use it for a specific purpose.