One thing Tesla and Comma.ai overlooked in self-driving | Hacker News Reader

One thing Tesla and Comma.ai overlooked in self-driving

deepdream.substack.com

Tesla is probably going to fake it using human remote drivers. Tesla has been running ads to hire the drivers.[1]

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...

No way I get into a car that has a human driver who literally has no skin in the game.

They can always fire the driver if he/she causes a few fatalities.

Are you sure the driver won't have a poster up above his screen like this one? :

https://is.gd/Lw37Su

Poster? Second screen multiplayer.

Yeah, they should have called them "human fleet response agents" instead. [0]

[0] https://waymo.com/blog/2024/05/fleet-response/

Your link says it's the same thing Waymo is doing:

> 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.

Waymo's people don't remotely drive. Waymo wasn't willing to trust cell phone network latency. They just give hints to the car's control system, like "turn around and take another route", when a car gets stuck.

Interesting. I believe you, but do known of any announcements or sources? Just curious to learn more.

"Fleet response can influence the Waymo Driver's path, whether indirectly through indicating lane closures, explicitly requesting the AV use a particular lane, or, in the most complex scenarios, explicitly proposing a path for the vehicle to consider. The Waymo Driver evaluates the input from fleet response and independently remains in control of driving. This collaboration enhances the rider experience by efficiently guiding them to their destinations."[1]

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...

Yes, and your link says that "Tesla's setup appears to be similar." I haven't seen any real evidence that Tesla's remote operators will actually drive the cars.

Here's one of several stories.[1] Job ads for the remote drivers have been seen.

[1] https://electrek.co/2025/05/29/tesla-testing-robotaxi-servic...

Wait... the big insight here is that people live in cities? Brother, this has been talked to death in the "is level 4 good enough" conversation.

Right. In the US, once the top 12 taxi markets have been served, Waymo will have most of the industry.[1] That's all the cities with more than 1000 taxis. Serve the next 10 cities, and it's down to Saginaw, MI, with 41 taxis.

In San Francisco, Waymo has passed Lyft and is gaining on Uber.

[1] https://en.wikipedia.org/wiki/Taxis_of_the_United_States

Given these counts are all individually sourced to different articles (or not sourced at all in some cases), I suspect that wikipedia list is rather incomplete. For example, I suspect Birmingham, Alabama, to pick a city at random, has more than the 41 taxis of the 20th placed in Wikipedia's list.

> Serve the next 10 cities, and it's down to Saginaw, MI, with 41 taxis.

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.

Uh, that list is wildly incomplete. There are 240 metro areas bigger than Saginaw, Michigan in the United States.

City vs metropolitan area vs CSA. Where cities are a legal jurisdiction, and self driving cars need a license to operate fully autonomously, legal city limits are the thing to look at, not the wider area that gets considered when eg LA is colloquially described.

Other have already pointed out that some of the most populous cities in the US aren't on that list, but also if the autonomous can be cheaper and/or more reliable we might see people using them more.

> it will probably not be very safe, otherwise nothing would stop Tesla from rolling out “the safe” autopilot to car owners today, unless you think that something will change in one month and we are getting some kind of breakthrough in software, so it seems like taxi will run today’s software.

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.

The most recent release of FSD is no where near good enough to do what Waymo does.

I use it every day and you are right! But my point is with six months of improvements in the software at the rapid pace demonstrated last year, combined with a boost from bigger models running on more capable compute hardware, I think it's possible that it could be close. Certainly up to the point where Waymo was when they started public rides three years ago.

FSD has been promised every year for the better part of, what, a decade? I’m dubious that a six-month sprint would be make or break for this finally happening.

I'm 100% with you on Elon crying wolf. But it's not really relevant. What they've achieved with FSD 13 is really quite amazing. Given how fast AI is advancing, is it so hard to believe that a system to drive cars using cameras will be feasible this year?

Yes. AI is advancing, but is still regularly making mistakes. Even in the domains where it excels, it still needs oversight from a skilled practitioner. Would you trust AI to run `rm` commands on your machine without your intervention? If not, why would you allow it to run a deadly multiton device on an unpredictable environment with adversarial actors?

AI is definitely not even close to being able to safely drive cars using cameras without an observant human driver.

I wouldn't trust a general LLM trained on reddit and 4chan (literally!) to run `rm` on my machine. But I would absolutely trust a purpose-built model trained on hundreds or thousands of years of data for a specific task. Models trained for specific tasks on huge datasets can be very reliable; certainly more reliable than humans.

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.

More than a decade. I bought my Tesla in 2014, and I recall the announcement because it came after I had bought but before I took possession of my car, and thus I was ineligible for the upgrade.

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.)

As a Comma 3x owner, Comma is not even in the same league as Tesla despite having better SAE level 2 steering. It jerks too much and they have a fraction of the usable sensors.

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.

If Tesla's Robotaxis become Bing to Waymo's Google, that'll be quite an impressive leap from where Tesla is today.

as a Waymo lover and a Tesla owner who won't use FSD, I think they have a long way to go.

> Waymo was frequently mocked that they are heavily geofenced, and “it cannot drive anywhere”-mentality. I have to say that I was also a subscriber to this camp and it seemed like universal “no map”-approach was to go. I was cheering for Tesla Autopilot team.

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?

> Another point is that city driving seems to be a safer in general, there is much less chance if things go wrong when you are in a taxi driving 30-40mph, than cruising with Autopilot 60+mph and hitting a truck on a freeway.

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.

It depends on how you define “safer.” Cities have a higher frequency of accidents, but with lower severity. Highways have a lower frequency of accidents, but with higher severity.

So in this case, you probably want to opt for accidents of lower severity. Metal undents more easily than flesh.

There's an aspect to Waymo's plan that means that they never really have to solve highway driving to capture market share. Highways are appealing to car manufacturers because they incentivize speed, acceleration, braking, and more nefariously, to be the bigger vehicle and more capable of surviving accidents. When automobiles were a new, growing market, manufacturers lobbied for fast streets, wide highways, and ample parking.

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.

The main downside is not just a cost thing but its waves are blocked by moist in air, rain, and dense fog. So, it will also degrades to just use cameras in this conditions.

From the recent tesla acme test, where the tesla runs into a wall containing a picture of a road, they also tested rain and water blocking two mannequins. The lidar equipped vehicle managed to stop before hitting the mannequins while tesla's autopilot did not. The distance before automatic braking triggered was reduced iirc.

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.

That test was not using FSD. When retested on the latest version of FSD, the car stopped. https://www.youtube.com/watch?v=7cxTO8g47_k

Can you provide a link to this?

It wasn’t real rain though. If the vehicle stopped because of the water instead of seeing the mannequin, there is no much value to the lidars.

Around the time that self-driving cars were said to be imminent, my kids were in preschool, and I realized that school pickup would be damned hard to implement as an automated thing. Cars have to line up and advance to the pickup spot using space that’s normally parking lanes. It’s definitely a 90th percentile challenge if not higher, but the sort of thing that literally happens daily in urban/suburban environments.

one point the author seems to be missing as to why we need full self driving everywhere is in the logistics industry. We are talking about interstate driving, from point A to point B, in this case it might involve industria environments such as a truck stop, weighing, loading/unloading etc, something that requires decision making across different environments (urban, highway, dense residential etc). Tesla FSD has the real potential for this unlock, more than just a ride hailing use case of getting people from point A to point B.

I remember Waymo had their "Via" division, seems like they been hitting brakes on that one. I think they had people driving the last mile, which makes sense. Again, freeway crashes are always gonna be the most lethal and the worst, especially if we are talking about semis.

Government should make an AV lane for trucks!

I don't understand why you think Waymo wouldn't be able to take the same approach.

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.

I love how people ignore the fact that Tesla's FSD is no where near good enough to do what Waymo does.

Reply on news.ycombinator.com