Doesn't surprise me, after reading http://www.bloomberg.com/features/2015-george-hotz-self-driv...: "Instead of the hundreds of thousands of lines of code found in other self-driving vehicles, Hotz’s software is based on about 2,000 lines."

In general I like the approach of using deep-learning instead of complex hardcoded rules (see the article above for more details), but even with deep-learning you're going to need a complex ruleset in order to deal with the edge-cases.

Getting a prototype of a self-driving car working is probably much less than 5% of the overall amount of work that's required. Getting the car respond correctly to the long tail of edge cases is the other 95% of the work.