Machine-learning of any kind has this uncanny ability to get you really far with very little work, which gives this illusion of rapid progress. I remember watching George Hotz' first demo of his self-driving thing, it's absolutely nuts how much he was able to do himself with so little. Sure, it drove like a drunk toddler, but it drove!

And that tricks you into thinking that the hard parts are done, and you just need to polish the thing, fill in the last few cases, and you're done!

Except, the work needed to go from 90% there to 91% there is astronomically higher than the work needed to go from 0% to 90%. And the work needed from 91% to 92% is even higher. Partly because the complexity of the corner cases increase exponentially, and partly because everyone involved doesn't actually know how the model works. It's been hilarious watching Tesla flail at this, because every new release that promises the moon always has these weird regressions in unrelated areas.

My favourite example of complexity is that drivers need to follow not only road signs and traffic lights, they also need to follow hand signals from certain people. Police officers, for example, can use hand signals to direct traffic, and it's illegal not to follow those. I can see a self-driving system recognizing hand signals and steering the car accordingly, but suddenly you get a much harder problem: How can the car know the difference between lawful hand signals, and some dude in a Halloween police uniform waving his hands?

You want to drive autonomously coast to coast? Cool, now the car needs to know how to correctly identify local police officers, highway patrol officers, state police officers, and county sheriffs, depending on the car's location.

Good luck little toaster!