That's an admission that neural networks are unknowable, but an assertion that they are better because, you know, neural networks.

I'd imagined that that statement was based on some collection of training and test data run through various simulation algorithms. It's hard to imagine where comma one would have gotten such a data set, but the problem he's commenting on is one universal to all of machine learning. What algorithm do you use to identify pedestrians from a point cloud? What is the estimated velocity of neighboring vehicles? Is that debris on the ground roadkill or nails?

At the end of the day you're mapping some noisy inputs to an abstracted output, and you can try fitting a bipedal model to a person, but that may error on statues near a roadway. Or you can plug the whole thing into a shiny set of general algorithms that integrate over space and time, let them work their magic, and pick the one with the highest false positive or false negative.

Not saying it's a right or wrong approach, and obviously any tests should include an appropriately large data set, along with added perturbations for all manner of lighting, noise, angles, added vehicles, etc. But it's a common question: do you trust the most accurate model or the one you understand the most? Ideally the former is the latter, but that can sometimes only come after years of analysis. I think the hip-young computer scientist answer to this would be make all data and analysis pipelines open, but something tells me comma wants the quick and easy solution.