Andy Halliday

But when you're working in a world model, the model is receiving a bunch of things through sensors and or simulated inputs that the sensors in the real world instantiation of a robotic embodied AI would get.

And it's going to reason about what it's perceiving and keep that reasoning alive.

understood as separate somewhat from its reasoning in the language model.

So it's going to output language, it's going to output action, and it's a vision language action model, but their pan world model does some separation that provides a technical advantage of actually being able to observe what its reaction is to the sensor input

that's either simulated or from real sensors in the real world environment.

And that helps structure the model so that it doesn't drift away from real world physics through imagination, let's call it, right?

The world of words and imagination and thinking

Isn't coupled so neatly to the regions of perception in our brains that you can cut off all your senses and you can still have imagination.

This model is being able to keep those two things separate in order to keep reasoning clear and connected.

There you go.

So that's a new element.

uh some interaction i will say before we go on i just want to say that um the the the commentators on uh you know from the muhammad bin zayed university of ai uh said the next step that they're going to take with this pan world model is to make its model imagine it's the model's imagination space or inner virtual visualization capabilities more rich and precise so

Just very interesting that imagination in advanced AI is a real thing once you jump over to world models because we don't want imagination so much in our text responses for a report that we're going to deliver to the board.

but we want unless it's about innovation but we definitely want imagination for practice like visualization think about this and i just saw a former professional skier so i you know i'm very interested in visualization as a practice for ski racing where you you you know you see the

Before the race, they spend time with their glove over their eyes and going back and forth, you know, visualizing the actual course.

And that has measurable and material improvements in the actual physical performance if they visualize the thing first.

So you can imagine this being applied to artificial intelligence in this way.

And that would allow the model to understand and render and act with finer detail and more fidelity in the real environment.

I do.

I want to talk about data centers and build out.