Slashdot Mirror
iPhone Hacker Geohot Builds Self-Driving Car AI (bloomberg.com)
An anonymous reader writes: George Hotz, known for unlocking early iPhones and the PlayStation 3, has developed an autonomous driving system in his garage. "Hotz's approach isn't simply a low-cost knockoff of existing autonomous vehicle technology. He says he's come up with discoveries—most of which he refuses to disclose in detail—that improve how the AI software interprets data coming in from the cameras." The article has a video with Hotz demonstrating some basic autonomous driving similar to what Tesla rolled out earlier this year. He's clearly brimming with confidence about what the system can accomplish with more training.
We'll see. I do kind of hope that his youthful arrogance doesn't get him killed. It seems unlikely that one kid will be able to outdo the big-budget teams of researchers working on this problem -- but I don't think it's impossible.
If killing people on the highway hacks out fixes, leading to quality AI for cars just a year or two sooner than a measured approach does, you will have net saved several million lives over a handful of people.
So you're going to make sure to offer up yourself and your family to be the first people to be killed, right?
The problem is all the corner cases needed to actually make the system safe in real world driving. That is highly non-trivial.
Indeed and that is what seems to be quite interesting about his approach. Basically he is saying that he is developing a system that can generate all those rules and corner cases itself, without a human having to quantify each scenario and code the rules into the machine. He states in the video that the car has gotten to where it is now (basic highway driving) by teaching itself. If it has, and his approach is extendable, then this is quite an interesting solution to the problem, precisely because it may deal with the non-trivialities you describe in a, well, comparatively trivial way.
Unfortunately, based on the article and video, there isn't really any way to determine whether he will be able to extend his system to give better performance, or even whether his system is just one of those 'learning systems' that is actually so highly tuned to the problem domain that it is essentially just an obfuscated rule based program. I guess we will have to wait for it to either get better, or for him to release some more information.
The main thing that makes me suspicious is why he has gone to the media about this now. If he has gotten this far with a design that actually does use learning, then why not spend a bit longer and get it to the point where he can demo it in less predictable environments. That would get us all interested. As it stands his current system only works in very predictable situations, so without more information it is impossible to know if this is a scam or not.