I am so glad that I am not involved with self driving systems (except as a squishy object on the same road). The systems are so incredibly complicated and conv

I am so glad that I am not involved with self driving systems (except as a squishy object on the same road). The systems are so incredibly complicated and convoluted that I just cannot see how they could ever be assumed safe. I've worked on (relatively) large legacy hairball embedded systems where you don't have confidence in the behaviour determinism and it was deeply discomfiting even when the highest price of being wrong was broken hardware or ruining a production run and having the customer give a support engineer an earful. I can't imagine the stress of having to deal with a much larger, more complex system that ends up in an open, public, uncontrolled environment where the price of a mistake is smearing a once-living human into paste. I'm also intrigued by how these systems are tested. Presumably, as well as unit testing everything to death, you use some kind of integration test where you replay scenarios and check outcomes, and/or some stochastic generated scenarios. This must take a lot of time and resources. As a shower thought, wouldn't it be interesting if you could get a virtual car model of a public-road self driving car and be able to play scenarios at it at will to allow external testing.