gfdr

Not only that, but breadth-first makes it easy to: a) Iterate much faster on models; b) Keep better track of results and code and models, which is already quite difficult in ML; c) Enable other teams to work on other parts of the pipeline, in parallel.

> This is essentially equivalent of a driving test but would be a controlled environment that simulates real cities.

So, which is it? A controlled environment, or a real city? I think it is a bit naive to expect /every/ edge-case to be worked out prior to even a testing permit being given out. If you could pass a test that faithfully simulates ALL the intricacies of driving in a real city, you're well past testing phase anyway.

Further, this ignores that we may be able to incentivize this behavior (learn all the lessons you can in controlled environments before putting people at danger) in other ways, e.g. financial penalties.