The way I see it (as a software engineer myself), is that each wreck that the AP did not prevent gets added as a test case in the AP unit tests that all have to pass. Also, each time a human wrecks, it could/will send the data to Tesla to continue simulations and ensure AP would not wreck in the same conditions.
How many of the wrecks, even fatal ones, essentially a different driver making the same mistake over and over? Likely millions. Real fleet learning would simply solve that problem once, and prevent thousands of fatalities as a result. As more and more AP capable cars hit the road, this simply will exponentially increase. Thirty years is laughable seeing how much computing has improved in merely the past 20. Maybe 5 tops assuming the legal hurdles are actually surmountable.
http://www.theverge.com/2016/10/19/13341194/tesla-autopilot-...
They also have every single human driver wreck captured with AP capable hardware to make fully autonomous vehicles safer. The way they are doing it right now is brilliant.
Edit: As Elon has previously said, not releasing something that saves lives is actually unethical. If a single life is saved by AP software that would have died if the vehicle was driven by human meatware, not releasing it is the unethical thing.
It's just a question of more training(altough a recent NHTSA? regulation says the planning stage cannot use black box techniques like DNNs).
And once self-driving is a business, i suspect we'll find ways to get said training data for cheap enough and fast enough. For example Tesla/Uber drivers as trainers .
Somewhere in the middle there's probably a point where self driving is demonstrably safer than human, even before they're "ready" for everywhere.