Honestly I thinkyour self driving example is something this could be really cool for. If the generation can exceed real time (I.e. 20 secs of future image prediction can happen in under 20 secs) then you can preemptively react with the self driving model and cache the results.

If the compute costs can be managed maybe even run multiple models against each other to develop an array likely branch predictions (you know what I turned left)

Its even cooler that player input helps predict the next image.