First, it seems that everyone but Musk is claiming so, so either everyone else is wrong, or Musk is wrong. I find it far more likely that this is just another "funding secured" type of claim, than an actual win. Case in point: see this evaluation [1] from only a year ago, where Tesla was ranked 19 out of 19 (Waymo was #1).
And the timing of the announcement, only two days before delivering really bad financial news... well, see the previous comment.
You provided the second reason yourself: Waymo has some incredible guys going after autonomy, and they've been doing that for a decade now. So whatever shortcut the Tesla people claim to have found, it's (a) either a shortcut nobody else has thought of before, or it's (b) they have thought of it before, and ruled it out. I believe (b) to be far more likely, hence I believe them to be on the wrong path.
Also, contrary to what you say, they do have something to show as of yet. They are operating hundreds of autonomous vehicles in Phoenix.
And yes, they are probably playing it super safe, but that's exactly what you should expect from someone accelerating 1+ tons of metal to lethal speeds, and giving it autonomy.
[1] https://www.teslarati.com/tesla-finishes-last-navigant-resea...
Talk to anyone with research experience who are building world scale ML systems that are used by real people, and it's unambiguous that Elon Musk is a con man.
It comes down to small details, like the nature of the training data, the nature of the customizations, the nature of the business problems, the nature of the regulatory problems, the difference between highway driving and city driving. The enormous amount of special casing, the nature of the QA infrastructure, the capital on hand, the routing and mapping problems...
What Tesla has is a decent lane finding system. What Google has is an end to end system. It won't take them long to expand to all the major cities; solving the geofencing problem is easy, they already made street view.
And they were doing L4 successfully years ago. They're keeping drivers in the cars right now mostly because something going wrong is a PR risk, not because they need them. And L5 isn't a particularly important goal.
My guess is I'd have to write up 20 pages of detailed analysis at minimum to convince folks. In general if you don't have 5 to 10 years in depth experience in this space, you probably shouldn't consider qualified to have an opinion.
With each of these breakthroughs, Musk presented a well-reasoned view that was highly contrarian against the authority consensus; and in each case, his view turned out to be not only correct, but so correct that it became the new authority consensus.
Maybe this time he is wrong and you are right. But if you are wondering why the market and the masses believe him, it's because he continues to provide reasonable explanations for what he's proposing, and he has a track record to back them up. If history predicts the future, then in a few years, we'll add "authorities said you needed LIDAR for self-driving" to the list above. This is already starting to play out in the SDC space, with Hotz of Comma.ai sharing Musk's views on vision, and Levandowski sharing his more generalized concern about LIDAR (the latter is more recent if you're not familiar):
> Levandowski contends that this “race” to deploy autonomous vehicles has yet to start in earnest largely due to shortcomings from “crutch technologies,” a descriptor he uses for hardware like LiDAR and HD maps. His position is that while these provide sensing and localization for the vehicle in the present moment, they have serious compromises and don’t produce the level of predictive ability required to commercially deploy autonomous vehicles.
https://techcrunch.com/2018/12/18/anthony-levandowski-pronto...
Karpathy was able to explain in very simple and understandable terms this week why Tesla's approach is advantageous. If the critics understand this domain so well, how come they can't provide a straightforward retort? If Feinman could do it with quantum physics, surely a machine learning expert can do it with neural networks. I would love to read/hear/watch such a retort if anybody is interested in presenting one.
Edit: fwiw I've actually taken a look at a number of waymos internal systems, and sat down with people who worked on autopilot.
However, Tesla is talking about making the tens of thousands cars they are shipping per quarter right now autonomous. We'll see how far they get with this. My guess is that the game for the next five years is going to be convincing regulators that these things are safe enough to allow on the road. Having a lot of cars on the road already will help their argument. Waymo has the same problem but all they can point too is some limited areas in the US where they are driving currently.
Tesla seems to be improving their current products continuously through software updates as well. This adds to their sales pitch: your car will gain capabilities and may even become fully autonomous one day. They've been riding this pitch for years and while they have under-delivered on the end goal, the in between milestones seem pretty interesting.
Two very different strategies.
Tesla doesn't spit out full self-driving cars in any number in any quarter, while Waymo is actually operating FSD cars in a limited, but public and commercial, role. So Tesla is still behind.
> However, Tesla is talking about making the tens of thousands cars they are shipping per quarter right now autonomous.
Yes, if one was looking for the industry leader in talking about delivering full self-driving, Tesla would be much more clearly competitive than they are in actually delivering. But talk is cheap.
Waymo last year ordered 20.000 cars from Jaguar Land Rover, then 62.000 cars from Fiat Chrysler, with deliveries apparently starting in 2019 [1].
[1] https://www.nytimes.com/2018/05/31/business/waymo-chrysler-m...
Emphasis is mine. What Tesla says, and what ultimately does are very rarely aligned.
Per my understanding, Tesla is tackling autonomy purely based on supervised learning. Waymo probably has something else, semi-supervised or unsupervised, where they collect data from a small geo-fenced area and can extrapolate that to the real world. I don't know.