> And I think the core argument is correct: If you don't have a neural net that can estimate depth from cameras as well as a human, you don't have a viable FSD system.

I agree that this statement is probably correct. But they're assuming the positive outcome here, and we have no proof that's correct.

In fact, it may well turn out that the current approaches to autonomous driving using neural networks won't ever yield a viable FSD system. In fact I'd take that side of the bet, to be honest.