I wanted to explicitly thank you for posting this, it was very helpful in clarifying the process used to develop vision and neural nets. Going over it a bit in my head there are a few points that I think are worth noting:
- Implicitly, the neural net is a component of driving that is not just limited to "computer vision." In fact, from his presentation it wasn't even clear if such a distinction would be meaningful. This means that the uncertainty of neural nets is an essential component of Tesla AP/FSD and thus cannot be proven to be correct, only demonstrated to meet a probabilistic goal.
- People here make references to Tesla's huge data set, but he makes it clear that a huge data set is actually counterproductive (e.g., training for traffic lights if blue traffic lights make up <0.001% of the data will never truly be accounted for). The real issue is getting enough sample range for each edge case so that it can be trained for.
- While a traditional computing approach to computer vision was unworkable, the "software 2.0" approach is only better insofar as it is at least tractable. But that is small consolation when you realize that all you have done is shift the problem to one of getting sufficiently large, clean and accurate data set labeled to unleash your training optimizer on.
- It reinforces my belief that true FSD is not on the horizon, but that incremental gains that will benefit Tesla drivers will continue. His examples of problems were pretty trivial -- he wasn't having to reach at all to come up with edge cases that were difficult to handle. At the same time it is encouraging that they are actively trying to identify edge cases for labeling. As Tesla iterates by identifying more problems, using data collection campaigns on the fleet to generate relevant data sets, labeling the data, and then using that to train an improved neural net, AP/FSD will continue to improve via OTA.
- Tesla's homegrown interface for data labeling probably represents the first significant step in "real" neural net programming. I used the quotes because it isn't like neural nets are new, but -- no offense to prior work on language, vision or to AlphaGo and relations -- if he is correct in describing Tesla's interface as being equivalent to an IDE it's rather like the field is starting to grow up.
What does this mean for Tesla, its investors and $TSLA?
I think it is the clearest and most convincing demonstration of Tesla's lead in FSD. Dog and pony shows are easy. But having the internals to grow and support the development of something as significant and difficult as FSD -- that is real. Of course, it isn't flashy and doesn't get attention. I expect the upcoming to dog and pony show to provide some of that -- but I
don't expect it to be showy or flashy. No "this car drove from New York to get here" silliness (who would be convinced by showing an arrival anyway?).
I've seen comments by people who do not have a Tesla that they don't understand how AP can make driving easier when you are constantly having to monitor the car
and pay attention to traffic. I get that from someone who has not experienced it -- they think AP is some Rube Goldberg contraption that is a circus sideshow freak. Alternatively, there is a belief that AP is fundamentally flawed and unsafe, but drivers will be tricked into a false sense of security and lose situational awareness.
The upcoming FSD show-and-tell is, I believe, to get influential people the hands-on experience that will make them believers. Belief that this is a solid solution and that it is safe. They will then use their influence to spread the word about how Tesla's FSD
works.
If there is glitz, a fireworks display, a laserlight show, I will be surprised as those elements are not necessary for this. Audi needed that for the e-tron launch because what else was going to sell it?
What
is necessary is a compelling drive in a car that drives on its own. What will work best is to demonstrate that Tesla's approach is not constrained by the impressive caveats faced by Waymo, Cruise, et al, but even if it is a solid demonstration of something technically achievable by Waymo -- as long as it is solid that will improve Tesla's autonomy credibility.
And if Tesla gains some credibility in the autonomous driving arena that
should be reflected in stock price. Whether or not that happens, it
will drive uptake of AP and FSD and advertise for more sales of Tesla cars.
