What we talk about when we talk about world models.

From American Public Media, this is Marketplace Tech.

I'm Megan McCarty Carino.

For a while now, large language models have been the hot thing in AI, but while they do some

remarkable things by predicting words in a sequence, they don't have a true internal understanding

of how the world works, the way humans learn grammar or the laws of physics. World models attempt to bridge that gap. AI pioneer

Fei-Fei Lee at Stanford has been working on them. So has Jan Lacoon, the former head of

meta's AI research. Google is also developing world models for robotics, but they're all pursuing

them in slightly different ways. To help explain, we asked

Gary Marcus, a cognitive scientist and author of the book Taming Silicon Valley. He argues we need

AI systems where the rules of the world are partially programmed into the algorithms by humans.

Inside a system could be inside of a human brain or inside of a machine, there's some

representation of the things outside, of the other people there, the objects there,

how things work in the world.

I'm talking to you, and I have a model of how you think, kind of psychology of you. I'm listening to you, trying to

understand your reaction. If we did this with cameras on, I would be looking at your expressions.

So I'm trying to understand what's around me, what the possibilities are. And I'm representing

things in terms of places and entities, sets of things, and so forth.

And how is that different than the kind of AI that's popular now, these probabilistic systems

like large language models chatbots?

So a simple world model would be the stuff in Wikipedia that says where people are born, what their birthdays are, and so forth.

And even that is something that a large language model can't reliably do.

...