LLMs are so 2025. World models are next

[ergophobe]

Moving beyond language will allow AI to interact in the world

"Nvidia backs startup Reactor as buzz grows for world models"
https://fortune.com/2026/10/05/nvidia-backs-startup-reactor-buzz-grows-world-models/

[littleman]

That's a really interesting idea for artificial thinking.  Us humans do a decent amount of thinking in vision too.  I think for a lot of us it is used as a verification of possible reality.  There is also the instinct level or reflexive type of thinking -- where our brain reacts faster than our comprehension.

Would a self-driving car be a SLM version of a visually thinking engine?

[ergophobe]

One of the insights from a couple years ago was that no amount of 2D images for training enabled models to generate good 2D images.

Generating 2D images requires a 3D model of the world, including the physics of light and gravity.

One of the implications was that the effort to build better image generators would lead to big advances in self-driving.