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?
[–] sosodev 2 days ago
[–] overfeed 2 days ago
> Isn’t the Waymo version used for some small portion of the dataset to cover edge cases that they have no real data for?
The usecases in the blog post[1] paint a much more expansive picture of how Wayne used Genie 3: including enabling transfer learning to convert single-camera dashcam footage into simulated full-take sensor data (multi-camera and Lidar), which is extremely valuable.
> My basis is that I’m a real human and these videos look far more like bad video games than the real life I’ve experienced
1:1 fidelity isn't required for a lot of training tasks; for instance, computer vision can do a lot with 32x32 resolution images, which is trash quality to humans. I think the potential customers for this technology will constrain the model to fit their use cases.
1. https://waymo.com/blog/2026/02/the-waymo-world-model-a-new-f...
This blog is suggesting that we could train robots using only a world model.
> Do you have a basis for this idea, or it's a gut-feel?
My basis is that I’m a real human and these videos look far more like bad video games than the real life I’ve experienced.