One thing Tesla and Comma.ai overlooked in self-driving(deepdream.substack.com) |
One thing Tesla and Comma.ai overlooked in self-driving(deepdream.substack.com) |
That has all the costs of the self-driving hardware, plus the costs of the remote drivers, but it maintains the illusion that Tesla is a growth company.
[1] https://insideevs.com/news/760863/tesla-hiring-humans-to-con...
> Waymo uses what it calls a “fleet response agent,” a human assistant the vehicle can ping when it gets confused by a complex traffic scenario. These agents can view real-time exterior camera feeds, examine a 3D map of what the vehicle sees and even rewind the footage like a DVR to get better context. “As with the rest of our operations, a helpful human is no more than a touch of a button away,” Waymo said in a blog post.
> Tesla's setup appears to be similar. The robotaxis will do the driving, until they don’t. Then a remote human may quietly step in to lend a hand.
In San Francisco, Waymo has passed Lyft and is gaining on Uber.
[1] https://en.wikipedia.org/wiki/Taxis_of_the_United_States
That is not a credible data set (it's missing 5 of the top 10 cities by population, which just can't reflect reality).
Hell, the Wikipedia article isn't even internally consistent, several of the cities earning a full section have no entry in that table.
The last version of FSD to increment the major or minor version number was released last year. There haven't been any big upgrades since then, they've only incremented the patch version number. Before that it had improved extremely rapidly for many months. The June robotaxi launch represents at least 6 months of improvement over the current public version of FSD, not one month. It's clear the team has focused exclusively on the robotaxi launch, and given the incredible pace of improvement of the public version in the year prior to their change in focus, I could easily imagine a huge improvement over the current (pretty good!) release.
Not only that, but I think there's a pretty good chance that the robotaxi version of the Model Y will include updated compute hardware, which I expect will significantly improve the performance of FSD just by virtue of running larger models. The difference between the HW3 and HW4 versions of FSD is quite significant with the only difference being compute, so it seems likely that even more compute could improve things further.
Toyota Safety Sense 3.0 is a better experience and no one is taking about it.
Tesla is also basically a non-competitor.
Waymo’s only technological competitors are in China. Anyone closely watching the industry has known this for years.
I never thought this argument made much sense because that's not really how humans work. If you take someone who's an excellent driver in NYC and plop them in SF they'll need some time to learn the streets before they can be a really good driver in SF.
For sure they'll be able to navigate safely, but perhaps not well, and I don't want my robotaxi making a wrong turn onto a 1-way street or missing a turn.
Can a neural net be trained to drive well in NYC, SF, and elsewhere? Maybe, but why wait until it's perfect everywhere before starting somewhere?
Freeways are wide, generally well maintained, have gentle sweeping turns with excellent visibility, have no pedestrians or cyclists, and don't have many junctions.
And there's only 50,000 miles of them in the US. 10 cars, 10 hours a day, 50 miles per hour, and you've driven them all in 10 days.
Much easier than city driving IMHO.
A viable self-driving business plan, on the other hand, has to accommodate taking final responsibility in an accident. That was what got manufacturers into the lobbying game to begin with - they needed to create a public that saw themselves as responsible owners while everyone else on the road was a meanace, and worked towards that reality through both consumer marketing and the financing and regulation systems around autos. Self-drive means that the goal changes to "every ride we provide is a safe one, and we do not serve customers that ask for danger".
And that means that some markets like regional airports and particularly sprawling, car-dependent metros may go unserved for some time, depending on how the strategists feel about their chances, but then the aspect of courting the public shifts towards strongarming governments into more intensive road safety measures, and then to only professional human drivers, and then perhaps to mandated self-drive in urban areas. Having tons of capital to throw around lets you dream very big.
In this way the problem gets redefined incrementally towards something that meets with where the engineering actually is and allows Waymo to compete while retaining its excellent record.
So in this case, you probably want to opt for accidents of lower severity. Metal undents more easily than flesh.
Sure lidar is degraded by rain, but I think the conclusion that when it rains the whole system will degrade to just cameras seems incorrect. It may reduce its effective range but it seems likely to still have some advantage in a wide range of conditions.
They've already got the tech for urban and residential working in their taxi service. It would probably need some work to get it working for a semi truck, but the base technology is there. The only completely new thing would be getting it working for highway travel and it feels like they could take the same approach of starting with a few key arteries, geofencing those, only picking contracts that are in their geofenced area and scaling up from there.
I'm not an expert on this at all, but I'd be surprised if the 80/20 rule didn't apply to shipping corridors as well.
Government should make an AV lane for trucks!
[1] https://electrek.co/2025/05/29/tesla-testing-robotaxi-servic...
It looks like the way this works is that the vehicle first has to come to a stop autonomously, and the control center then gives it instructions on what to do next. It's for getting out of stuck situations.
California's CPUC permit for Waymo operation does not allow teleoperation.[2]
Reviews of the job on Glassdoor indicate that it's more or less OK. No opportunity for advancement, but snacks are included.
[1] https://waymo.com/blog/2024/05/fleet-response
[2] https://www.cpuc.ca.gov/-/media/cpuc-website/divisions/consu...
AI is definitely not even close to being able to safely drive cars using cameras without an observant human driver.
In hindsight, thank fuck. I would have dropped thousands on it in a heartbeat and never, ever seen it. (Also, that car was a nightmare to own and the service centers were really scammy. Getting rid of it was the best move I could have made.)
I wouldn't let GPT-5 drive my car, but I let FSD drive it every day. It's not perfect yet, but I definitely see a day very soon when it will be better than me or any other single human, with all the failures humans have.