That’s a very interesting neutral take that “Tesla’s strategy might be the winning horse in this race” and that the Waymo CEO may be defensive or scared. I personally don’t know enough about either company or their tech or how people react to self driving cars to call a winner when it comes to a something (in this case, fully self driving cars at scale) that literally don’t exist yet with enough certainty to predict winners or establish the emotions of either company’s leadership.

How else would you describe Waymo's message?

Trying to "educate" or talk down to a competitor with words like "this is not how this works" when Tesla has significant resources, expertise (actual cars on the road!) and knowledge in this area seems defensive to me.

To be fair, Waymo (back then Chauffeur) have solid experience with Tesla's approach. They started with the same strategy of gradual improvements. They found out (and shard) the same issues Tesla is running into in the past years: Users don't pay attention when monitoring a system doing a boring task - even when failures carry a high price. Waymo switched the approach from a gradual L2/3/4 improvements straight to developing an L4 system; Tesla is locked in to the gradual move due to it being part of their business model, and are now stuck with trying to make self-driving work with unqualified test drivers (i.e. customers) and outdated technical decisions (not using Lidar because they were too expensive back then).

So Waymo does have the experience to call out Tesla here.

I recommend watching Lex Friedman's interview with George Hotz - its has some good points that one might should consider around the different approaches on reaching L5.

The term you’re looking for is “Federated Machine Learning.” The more cars Tesla has, the faster they can train the neural networks edge cases.

They are collecting data on all cars yes, but are training a model in one place. That is not federated machine learning, it’s federated data collection at best.

I think the massive collection of data from more and more Teslas is going to eventually allow them to produce the best self-driving vehicle.

Its been a maxim of history that those with the most data generally win, whether that's military intelligence of troop movements from 2000 years ago, or LIDAR scans from a car today.

I think the best evidence for Tesla doing poorly here (or maybe that LIDAR really is necessary) is precisely that they have all this data and still aren't doing as well as others. Tesla has a seriously absurd advantage here: they can observe millions of drivers and see when the human drivers do things differently than their self driving software would have, and capture the data from those events. They've had this advantage for many years, and yet, they still don't seem to be performing that much better (if at all) than competitors.

I agree with this concern. They collect a lot of data, sure, but they are also the only one who thinks that video is enough. It's a bold bet; intuitively, one would expect that video-only is good enough for gimmicks and fares worse when it comes to covering the millions of edge cases that still need to be addressed.

This is a curious point: How much data does Tesla actually collect? I can imagine users not being too happy if their car clogs the upload for hours every time you park it in your garage?

More than you think. There are sensors everywhere and pretty much all of them are regularly uploading telemetry to the cloud. This goes much further than self driving cars, the entire CAN bus is constantly synced with the cloud, all sensors, vibration, tire pressure, etc. All in the name of predictive maintenance and whatnot.

Ah yeah you’re right. Distributed data collect.

Tesla does not use federated ML. Federated ML is specifically about training models on device, and is mainly a technique for preserving data privacy.

That’s not really what federated ml means, unless they are doing the gradient updates from the cars directly.

Thanks for sharing. I actually hadnt heard of "federated machine learning" before- need to read up on that.