It’s basically like. Someone drawing a picture. Then watching the buttons you’re pressing on a controller. And then drawing a new picture. And based on the game that they think you’re playing in their head trying to guess what the next picture ought to look like. With no error correction and no conceptualization other than what the next picture should look like.

The… many limitations of this is the inability of image generators to rationalize 3 dimensional space. It can only approximate it based on what it thinks should appear on the screen. It lacks any ability to keep track of variable information. It really is more like a Doom-style hallucination than anything else. Some of the videos on that article are truly bizarre looking. I’d imagine after a few minutes every single one of them would devolve into an endless loop of being trapped in non-sensical geometry or killing the same enemy over and over again as the AI has no way of remembering the enemy existed to begin with, let alone that you killed it.

I’ll be honest I don’t think there is much use in this at all. It suffers from the same limits as any other model AI. Believability at a glance is not believability under scrutiny and if it’s only believable at a glance then there’s not much practical use in it. The advance in computational power and model sophistication required to stand up under scrutiny is massive.