FWIW I coded my own neural network from scratch and have been coding for a scary 39 years *cracks knuckles*.
Yes, teslas approach is not only the best, but IMHO the ONLY one that can possibly offer a truly universal solution, and the only one that can produce a really safe solution to FSD.
But hell... this stuffs hard. Neural networks become a lot more like art than science. Its a messy *fuzzy* business, and starts to feel more like voodoo than normal code. Despite all the memes, its not that hard to predict how far through a project you are when its normal object-oriented traditional C++, but the minute you have a massive neural net involved... things get way harder to predict.
FWIW my entirely subjective opinion is that they may still be another small rewrite away from FSD. I think the big gear change was karpathy. He understands that you need the code to be mostly neural net, with minimal procedural code. I think thats the only answer, and its a tough one to agree with unless you work on this stuff all day, because it feels SCARy as a coder to effectively cede control of decision making to a neural net nobody can understand.
From what I read, karpathys time at tesla has been one of progressively lobbying for the NN to take over more of the decision making. I think it will make up an overwhelming proportion of all the FSD code eventually.
I dont think we will get FSD until end of 2021 at best. But I am 100% certain nobody will beat tesla to it, or even get close. Persuading trad auto companies to let software control a car is very hard. persuading them to let software *nobody understands* to do it is almost impossible. Its only possible at tesla because elon is a coder with an interest in how brains work.
thanks cliffski
I agree traditional automakers are not at all likely to venture into this on their own.
if I have it correct that a neural network approach requires real world miles on something like the scale Tesla is using, than even the big tech companies like Google, Apple, Intel, Nvidia, who in one way or another have a foot in the FSD door, would need to both go the neural net route (not sure if any are yet) AND pair up with an automaker who agrees to have the hardware, software, and upload data/download updates model Tesla has feeding from the automaker partners cars to the tech giants FSD team.
Curious if you see this as a reasonably accurate read of the pathway these tech companies would need. If so, it implies, Tesla has a lead time that we can monitor and estimate based on when any such potential 'big tech/large automaker' partnership arises (not 20K Waymo/Ipace, but, scale like Tesla), and our estimate on how long it would take from that point to achieve FSD, depending on what we think of the quality of chips and software teams at such a large tech company are.