Massive amount of "world" data - enough for Google to successfully train Veo3, a "video generation" model that is extremely good, and evidently includes a decent "world model" i.e. a model that can generate or, more accurately, predict the (short-term) "future" - like "what happens if you let go of an apple" (it falls down) etc.

This kind of "world models" that can understand physics enough to be able to predict short term future (like humans can) are crucial for any kind of real-world AI (i.e. robotics, including self-driving), because they constitute what could be termed as "physical common sense" (that humans have, but also animals, to some degree).

Is it enough for self-driving? No, you also need to understand road rules, communicate with humans (pedestrians and fellow drivers), etc. but it's a good, possibly necessary, step - it allows you to better handle many unpredictable (tail of distribution) situations:

https://www.reddit.com/r/SelfDrivingCars/comments/1g75ftb/wa...

https://www.youtube.com/watch?v=-tJH8hED11I

My guess is he means they have lots of car-cam videos like this:

https://m.youtube.com/watch?v=MejbOFk7H6c&pp=ygUJb2sgZ28gY2F...