Google took a completely different approach and relies heavily on pre-mapped environments. Tesla has several orders of magnitude more vehicles, more data, and a more diverse set of data given their cars are in many different countries. It's not implausible to think that Tesla could arrive at FSD faster than Google.
The idea that data will somehow magically translate into FSD is laughable. It relies in the delusion that we just need to train neural networks with the proper data and then we all can go to sleep.
There are many issues with Tesla's autopilot that are completely unrelated to the amount of data they have, and they will not be fixed with more data, and having more data will not make it easier to fix it. At this point, I would argue that the discussion about who owns more millions of miles of data is completely irrelevant.
It isn't laughable at all. The real problem of FSD is the ridiculous long-tail of scenarios in the real-world that you simply cannot account for or manage well. At this point, Tesla has a huge upper hand because every vehicle in their fleet can constantly collect and provide new semi-labelled training data every time there is a user disengagement or an unforeseen action taken by the driver.
Tesla has built out amazing infrastructure to capture extensive amounts of "hard" examples from their fleet, turn them around into labeled data for training very efficiently and then utilizing simulations to further broaden the distribution of such quirky long-tail events in their training-set. In the absence of AGI, this is a very effective "brute-force" approach and they have a huge upper hand over every other player in this space.
I say all of this even though I am very skeptical that anyone will achieve L5 self-driving with where the state of things are today. But Karpathy and team are very pragmatic and making lots of good decisions coupled with excellent engineering and infrastructure development.
I can't see how you can content Waymo has more data, they had around 600 cars on the road last year. Sure they have some more sensors but Waymo only operates in select cities and select routes, the diversity of the data is very limited. They won't have seen scenario's like a snow storm for example since they've only been testing in CA and AZ. Meanwhile Tesla has a robust framework for collecting clips of specific scenario's as needed by shipping a small model to roughly match the scenario.
Simulation is important for sure, and Tesla even talked about that as much. But you can only simulate the scenario's that you've thought of, it can't replace real world data, which again Tesla has several orders of magnitude more of than Waymo.
As for the comments about Dojo I think they're unfounded. Tesla has already proven that they can create their own custom chip and have shipped it to hundreds of thousands of cars. I don't see any reason to think they can't get the system running in the next year or so. But even if they can't do that in time they still have NVIDIA to fall back on. Elon even said in the talk that maybe it might not work out, and if that's the case they can always buy a solution.
