It's fairly easy to train, and verify a system for driving in well-behaved traffic. Unfortunately, the problem space of not-well-behaved traffic is far wider - and is very hard to gather enough data to train a system well.
What you're going to get is self-driving cars which handle 99% of driving just fine - and when they end up in emergencies, find the human 'driver' to have dozed off at the wheel. (All-in-all, their safety record might end up better then the status quo - but that's not a certainty.)
The reactions of NN to unusual stimuli are likely to be counter-intuitive at best and unpredictable at worst.
Machines, on the other hand, are a different story.
It seems that the technology has already, or is very close to approaching human levels of proficiency on the road. If specific use cases (offroad, snowpacked road etc.) are problematic, they can be limited or prohibited in the mean time.
I have no inside knowledge, but I would be very surprised if Google's self-driving cars use neural nets as a significant component right now (which isn't to say there aren't people exploring its use).
What's better about an AI powered by neural nets is that you could train an AI to go offroading.
Get enough data and you've got a model for dealing with a given situation. Google's biggest strides with OCR, Voice recognition, Spam filters and other AI tech early on came from its ability to gather a huge corpus of data.
The real challenge is two fold. Gathering data, and feeding the AI with inputs with data actually matters. This is the secret sauce that Hotz refers to in the article as the information he's not willing to disclose. That information will become commoditized in due time (like low-latency optimization for HFT), but it will take plenty of institutional money & experience (Google, Apple, Tesla, Ford, etc) to get it there.