neroden
Model S Owner and Frustrated Tesla Fan
Indeed.
I was trying to figure out where in this thread to interject this, and your post is a great one.
Adding to it, everyone should read about how AlphaZero became the most powerful chess playing entity on the planet in just 4 hours. What's incredible about this is that the researchers didn't do this by teaching their AI how to play chess. They did it by developing a general-purpose AI advanced enough, and then simply teaching it the rules of chess, and just letting it play.
No loading initial strategies, no showing it 1000's of games to model itself after. They let it learn by playing (itself), and 4 hours later it crushed every comer, including the previous best chess programs/AIs that had been painstakingly tuned for countless hours. It took 8 hours to master Go (considered significantly harder than chess) and obliterate everyone and everything there as well.
Interestingly, in both cases the AI developed some winning strategies that were counter-intuitive to it's creators and experts, but quite effective.
So that's not to say that we can load AlphaCarZero in to Tesla's FSD computer and let it loose and call it a day. But it does underscore @eloder 's point that if " advancement of the neural net could potentially be reaching exponential increases at this point" and if so "we're in for something pretty earth-shattering" when that happens.
It may not be on the 22nd, but at some point I suspect we are going to turn a corner and the issue of "good enough" will rapidly vanish as a question. It won't be perfect, but as Elon says, a 10X reduction in auto fatalities will be hard to argue against.
I will repeat that all of this Go and Chess history is totally irrelevant to "self driving" because *we don't actually know the rules for driving* -- or at least Tesla staff don't! As soon as a suitable set of rules is devised, the problem is essentially solved, yes. But devising that set of rules is very, very hard. I don't think we've ever tried running a neural net on a game with unspecified rules which the computer isn't told and which the humans argue about.
I mean, I think it's embarassing that they're still working on computer vision and object classification. They are deluding themselves if they think solving that solves self-driving; that's just a prerequisite for even being able to *start* work on self-driving.
Tesla does appear to be far ahead of everyone else, and if they manage to simply implement automatic emergency braking that ends all "run over pedestrian" crashes, they'll have an insurance goldmine on their hands.
But "sleep in your car" self-driving is a bridge too far for the near future; from what I can tell they still haven't really started on the corner cases.