> Cool demos, but does anybody else find it troubling that companies keep claiming this tech can train robots? It’s so obviously not representative of the real world.

A proprietary version of this technology is already training Waymo's robot cars. Waymo blogged about their DeepMind-provided world model a year ago.

> It’s so obviously not representative of the real world.

That entirely depends on the use-case and level of granularity needed, doesn't it?

Subreddits where SolidGoldMagikarp was counting to a million in comments were used to train a machine that then writes web pages that more machines are trained on. The sheer volume of data with sufficient verisimilitude generated by these models may well work to train robots. It might not, but I think it might well work. The idea that things have to be high-fidelity high-quality to yield good results has clearly not held up in other domains.

Elon revealed to the world that you can just lie and people will give you money, so now they all lie.

This stands out to me too, it seems to be representative of physics the way a video game is representative of physics, likely less so given the non-determinism

Gemini Robotics is a thing: https://deepmind.google/models/gemini-robotics/

They clearly know how to train robots. Unless Genie is a misdirection...

I guess they want military contracts or something.

You can get useful levels of robot training out of Super Mario Bros 2; see John Carmack's talk from last year: https://www.youtube.com/watch?v=4epAfU1FCuQ

When your training simulation fails to reproduce certain features of the real-world environment, your control model is likely to have trouble with them, so more realistic simulations are helpful, but even fairly unrealistic training simulations are pretty useful.

In general, training neural networks to control robots is going to require simulations, ideally differentiable simulations. Even if they aren't differentiable, though, you can SGD your simulated training runs across a data center in a way that you just can't with training runs on Roombas.