A Song of “Full Self-Driving”
thebulwark.com
thebulwark.com
If it's hard, Waymo has an insurmountable lead. If it is easy, everyone can replicate easily what Tesla is doing.
I'm not seeing any strategic moat on Tesla's side
The key ingredient is data - the bitter lesson. It's not about better algorithms, but simply about algorithms that can process more data efficiently (e.g. transformers).
Tesla is one of the few companies that have a data flywheel - a fleet of (non-self-driving) cars collecting real-world data worldwide all the time at massive scale!
Now that is an insourmountable lead. (Along with good engineering, which, believe it or not, is still a competitive advantage - see e.g. German car companies unable to launch a single useful on-board computer, let alone a software-defined self-driving car.)
Google is one of the few companies that could compete, even without Waymo, because of YouTube.
Is data enough? Maybe not, there's a lot of progress on RL now that can do wonders without even more data. Is data necessary? No evidence to the contrary, yet.
Every road, in every country on earth?
No, that is never going to happen.
This kind of "world models" that can understand physics enough to be able to predict short term future (like humans can) are crucial for any kind of real-world AI (i.e. robotics, including self-driving), because they constitute what could be termed as "physical common sense" (that humans have, but also animals, to some degree).
Is it enough for self-driving? No, you also need to understand road rules, communicate with humans (pedestrians and fellow drivers), etc. but it's a good, possibly necessary, step - it allows you to better handle many unpredictable (tail of distribution) situations:
https://www.reddit.com/r/SelfDrivingCars/comments/1g75ftb/wa...
https://m.youtube.com/watch?v=MejbOFk7H6c&pp=ygUJb2sgZ28gY2F...
Reality is the Tesla product in this space hasn’t advanced in a decade. They hit a wall. They have promised that the next big thing is just over the next hill, but like the Roadster, hasn’t quite arrived yet.
The point is that automated driving with a fleet of sensors can be an order of magnitude safer than just using cameras. The current slow improvements at Tesla do not correlate to an unlimited amount of improvements.
Does that mean I think it’s the future or they have won? Nope, not at all. I don’t like Musk but I also don’t let me feelings for him cloud the discussion.
I have no idea who is going to win or what that strategy will be, vision only, lidar, combination etc. but it’s false to make claims like yours. I get it though it’s easy to take your hate for an individual in an extension to other adjacent discussions. I don’t like the guy either but not worth making up silly claims.
“Tesla has been getting closer to FSD the same way climbing up a ladder gets you closer to the moon”
This is a silly claim that perpetuates your dislike of Musk. FSD has absolutely gotten no where close to his claims. I have no idea if vision only is the right path and Tesla probably does play too fast and loose but their FSD program has absolutely been making improvements in, especially in the past few years. How about you back up your claim instead of using hyperbole?
I stated my reasoning, the sensor stack in insufficient. The evidence is they've been trying at it for close to a decade now, and have had to continually move the goal posts, missing their self-proclaimed milestones, delaying their shipping dates by years, reneging on their promises and projections, and leaving a trail of death and destruction in their wake (and that's not hyperbole); whereas Waymo went with a better sensor stack (the one roboticists have been saying was the way to go), and is actually shipping a true full self-driving service, one that won't run off the road if you fall asleep.
So we already know what does work (orthogonal sensors like LIDAR+RADAR+Camera) and what doesn't work (Camera only). We learned this in 2007 at the DARPA Urban Challenge, and it's been affirmed by Tesla's continued track record of failure in 2025.
It’s far more likely that you’ll be commuting in a Waymo someday.
Like some other tech moguls, you can’t separate the man from the company. The company lacks meaningful governance, so for good or ill it’s an expression of Elon’s hopes and desires. Facebook is similar, but Zuck is more calculated and rational. Elon is obviously a man in need of help who is degrading before us.
Meanwhile, Google actually delivered the robot taxi in SFO, and it’s amazing.
The sensor strategy really says it all. Why skimp on a minor part of the BOM, which enables a use case with 50-100x value?
The whole Tesla schtick is such a bizarre Lucy and the Football scenario, over and over again.
It’s getting better. Slowly, but surely.
The problem with Teslas is given the deficient sensor suite Musk has insisted on, they can't be programmed to not do things, like not run though a picture of a road like Wile E Coyote.
Given how quickly things are changing, I don’t think it’s useful to say iPhones are crap and always will be because you just got a 6 running super old iOS.
Things are not changing rapidly, this has been a failure mode on Teslas for almost a decade; it's been a problem at least since 2016 when Joshua Brown was decapitated by his Tesla running on AP. People keep saying "no, that was the last version. The new software/hardware version fixes it", but the looney tunes test demonstrates Tesla's sensor stack continues to be fundamentally insufficient.
And this is why: https://insideevs.com/news/658439/elon-musk-overruled-tesla-...
Musk thinks he knows better than his engineers how to solve this. And no, I don't consider Musk to be an engineer, he's a showman.
This is literally the definition of progress.
Tesla is currently developing HW5 which has 10x the inference capability of HW3. More progress.
https://electrek.co/2025/05/23/tesla-full-self-driving-veers...
My chin split open, and I had to get 7 stitches. After the impact, I was hanging upside down watching blood drip down to the glass sun roof, not knowing where I was bleeding from. I unbuckled my seatbelt and sat on the fabric interior in the middle of the two front seats, and saw that my phone’s crash detection went off and told me the first responders were on their way. My whole body was in shock from the incident.
Wally said he was on Tesla FSD v13.2.8 on Hardware 4, Tesla’s latest FSD technology. He requested that Tesla send him the data from his car to better understand what happened.
This person is just too embarrassed to admit they fell asleep at the wheel and the weight of their hand disengaged FSD. This could easily be cleared up if he showed the driver facing camera the car records in a crash.... But something tells me he won't do that.
You don't fact check things very well. That's two blatant falsehoods you've swallowed. What else are you wrong about.
My initial post said: "they can't be programmed to not do things, like not run though a picture of a road like Wile E Coyote."
Then someone said: "That video was not FSD, but an old version of “driver assist”."
Then I linked a video showing that in fact FSD was driving through the wall.
Obviously if you have to upgrade the hardware to not run through the wall, then they can't program their current cars to not do so. Because, as I've said since the beginning: their sensors stack is insufficient.
And that's the standard line with Tesla fans is constantly shifting the goal posts. It was supposed to be the case that in 2016, all Teslas sold would be equipped with hardware capable of self driving, because Elon told us that cameras were enough. And yet, they have required hardware refresh after hardware refresh after hardware refresh (now on V5) because he's been wrong time and time again. Because cameras aren't enough.
And what is this progress you speak of? Just special casing edge cases while leaving fundamentally flawed hardware and software on the road. Now it's 2025 and maybe finally HW5 will bring it? I guess we'll have to see because nothing Elon says can be trusted as fact.
> This person is just too embarrassed to admit they fell asleep at the wheel and the weight of their hand disengaged FSD
Ah yes, a driverless system that actually doesn't prevent you from crashing because you fell asleep. I don't know why you feel this alternative take makes Tesla's product look any better. Such progress. Much driverless. It's like a sailboat that sinks instead of floating.
FSD will not randomly drive off the road and your post showed a human turning the steering wheel by accident.
Progress:
HW4>HW3 and HW4 cars can see painted walls HW5 exists and will be deployed soon.
Goalposts shifted by you = 2
How do you know? The point of the video I posted from Youtube was in response to a chorus of people claiming the looney tunes test wasn't valid because FSD wouldn't be affected by it. And yet, they demonstrated FSD plowing through the wall. So Tesla had to respond by upgrading the entire HW stack, because a HW3 car apparently doesn't have the necessary hardware to see the wall. That's called moving a goalpost.
We have Teslas driving into walls, Teslas driving into firetrucks, Teslas driving under tractor trailers and decapitating their passengers. Presumably they weren't supposed to randomly do that either, yeah?
So with what confidence can you tell me they won't drive off the road? Has Tesla proved this? Is there independent testing that has verified it won't? Is there any guarantee Tesla can offer that they won't? Does Tesla even make this claim?
> HW4>HW3 and HW4 cars can see painted walls HW5 exists and will be deployed soon.
The entire existence of HW4 shows us that cameras are not enough. The entire thesis of Musk is that this should have been possible with the original Tesla sensor stack, so why are we on HW5?
“The Model 3, Model Y, and S and X that we make today, will be capable of full autonomy, unsupervised.” If that sounds familiar, it is because he said almost exactly the same thing several times in the past, as far back as 2016. Musk reiterated this sentiment after the release of HW3/AI3 in 2019, and as we all know, Tesla has since released HW4/AI4, and has HW5/AI5 on the horizon.
You want goalpost moving, that's goalpost moving.
The key ingredient is indeed data, and also, depending on the stack, hardware. If "true" level 5 self-driving is only possible with LIDAR and up-to-date HD maps, then it won't happen for some time. I foresee that unbounded "full" self-driving will either never happen or with severe boundary conditions only.
The bitter lesson isn’t fundamentally even about data. the key ingredient is computation (which does scale with data in modern deep learning). There’s even a whole theme on search outperforming learning - until learning methods changed to leverage computation!
I think in a lot of applications, richer data yields better scaling performance. So the bet for self driving is that the complexity of multiple sensor fusion gives a much better constant factor or exponent in the power law scaling of performance with data and compute
https://electrek.co/2025/05/23/tesla-full-self-driving-veers...
This was literally 2 months ago, so miss me with the "you're wrong, the latest FSD solved it all this time bro, for real, trust me (the 100th time I've heard this)".
> Despite its name, Full Self-Driving (FSD) is still considered a level 2 driver assist system and is not fully self-driving.
How the hell has Tesla not been sued into the ground by now?
Musk equates machine vision to human vision, but that is an over simplification, and the best MV algorithms and methods are still miles away from human capabilities. FSD is very reliant on depth perception, which is much easier to solve with sensors other than stereo vision.
I also don't get your German car company statement, Mercedes has been a technology leader in the automotive space across a number of fronts for quite a while.
Waymo isn't the only player in the market and unlike Tesla and their Chinese competitors, they aren't really structured as a manufacturing operation. They are modifying consumer cars made by others, probably at great cost. That raises the question about how they are going to scale. Waymo is so far not a scalable operation. They can manage thousands of cars. But can they handle millions? Or make a profit.
Chinese manufacturers, like Tesla, have a ready answer to that by virtue of their proven manufacturing skills. And it seems they so far don't need Waymo's software to get their cars to drive themselves. Tesla is late. But they can build cars at scale. So, I'm not ready to claim they are too late yet.
I think Waymo might have a bright future ahead licensing their platform to manufacturers that are, so far, not showing many signs of having their own version of this.
As for whether this is technically possible or not, I think it's hard to argue that it's not from the back of a self driving taxi. I've not had that pleasure yet (none around near where I live) but I've seen enough to not be that foolish.
America doesn’t have a monopoly on challenging driving situations, y’know. :-)
If so, I wonder if or how that affects the algorithms for pedestrian avoidance. :)
AV demos are easy. Scaling AVs is hard if you don't have mature technology. You need way too many remote supervisors to be able to scale, even in China where it's relatively cheap to hire humans to fill the gaps in your AI.
I'm not arguing for complacency. The opposite really. But without realistic assessments, it's easy to get the impression that something is a lost cause or that protecting technologies doesn't matter.
It's just not that expensive or hard to add sensors to an existing car and use a new platform. Byd and other Chinese automakers have similar fsd with lidar for $30k or less. It's certainly true that Tesla can crank out a lot more cars as their factory utilization goes down by the day. The millions of immigrants with low cost vehicles driving for Uber are also going to be hard to outcompete.
My car doesn't even have LIDAR but it has 5 radar sensors (corners & forward) so it can do all sorts of neat stuff. For example, it has actual working adjacent lane blindspot detection, which Tesla never dialed in. It also can warn you about oncoming cars at intersections or when backing out of parking spots. It even flashes lights on the doors if you try to open a door when a car is coming in the adjacent lane.
None of that is full self driving but it leaves me wondering what Tesla can ever accomplish with cameras-only.
Is this not standard? I don’t think my Hyundai Ioniq 5’s blind spot detection has ever failed to warn me about a car in my blind spot
Current Teslas at least actually have some sort of light in the sill, but again its using cameras only, no radar, sonar, whatever.
They still are ahead on some technologies, with the Cybertruck's one and only interesting technology its drive-by-wire system, but the industry has caught up and overtaken Tesla with the relentless progress of the year model.
And as others have pointed out - not only is Musk insistent on camera-only, but they are not even good state of the art cameras.
For years they were using cameras so bad that even though it provided a "security system recorder" for your car, you couldn't make out license plates most of the time. Kind of useful to be able to read the license plate if you want to actually have anything to give to police when a car hit & runs you.
They only in a recent year updated the cameras to.. still not even 4k.
Sure reading license plates is not the primary purpose of these cameras, but then again .. I'm supposed to trust my life to a car that "sees" with 2010 era iPhone cameras?
California minimum vision requirement to operate a vehicle is 20/40 vision [1] which corresponds to a arc resolution of 30 pixels per degree of field of view [2].
HW3 cameras have a horizontal resolution of 1280 pixels [3]. There are three front-facing cameras with field of view (120, 50, 35 degrees) and different focal lengths with optimal distance (60, 150, 250 meters) placed next to each other preventing usage of binocular vision meaning their front-facing cameras are “one-eye” driving.
The cameras have a arc-resolution of ~11 pixels per degree up to 60m, ~25 pixels per degree up to 150m, and ~36 pixels per degree up to 250m. That corresponds to ~20/120 below 60m, ~20/50 at 150m, and ~20/30 at 250m.
In comparison, you are considered legally blind if operating with 20/200 vision which their below 60m camera just barely surpasses. Up to 150m, their cameras fail to meet minimum vision requirements to operate a vehicle in California.
Even on HW4, the 120 degree camera for below 60m, which is the majority of complex high-acuity maneuvers, only has a horizontal resolution of 2896 pixels which is only ~24 pixels per degree corresponding to ~20/50 vision which is below the minimum vision requirements in California.
[1] https://eyewiki.org/Driving_Restrictions_per_State
[2] https://en.wikipedia.org/wiki/Visual_acuity
[3] https://www.blogordie.com/2023/09/hw4-tesla-new-self-driving...
My car just has really really good highway ADAS. Which is all that most people need.
How's it do for bikes? If it can catch them as well, mind saying the brand?
Ford does not suddenly have several million cars with 8-9 cameras to tap into for training data, nor does it have the infrastructure/talent to train models with the data it may get. I think you are underselling the Tesla moat.
Its the same reason why there are only 3-4 "frontier" AI labs, and the rest are just playing catchup, despite a lot of LLM improvements being pretty well researched and open in papers.
* treating red lights like stop signs
* completely ignoring flash reds
* stopping at flashing yellows
* ignoring school zones
* driving 10mph under the limit
* driving 20mph over the limit
* ignoring turn only lanes resulting in being dumped into oncoming traffic until it leisurely moves back into the correct lane
* waiting until the last minute to get into the correct lane to make a turn/take an exit
* ignoring signs indicating a lane ends and having to force its way into traffic
* ignoring signs a lane ends and changing into the lane only to have to change back immediately
* pulling out in front of cars, especially at night or other camera obscuring events
* ignoring cars merging onto the highway and blocking them until I take over and get out of the way
I treat it like cruise control. And based on the recent accident during which FSD decided to run itself off a perfectly straight, open, clear rode and into a tree, I keep my hand on the wheel.
Even in a geo-fenced environment, FSD is NOT going to work for unsupervised robo-taxi applications. Musk needs to get out of the way and let them stick LIDARs on these cars.
All that being said, it has also handled really complex driving in/around DC/Baltimore including some tricky merges/transitions that are hard to track as a human driver. It's knowledge of all of the surrounding vehicles allows it to make these maneuvers with far more confidence that I would. It also frees me up to track the bigger picture of traffic while not having to manage my speed/lane position. It's like driving with a friend who's not necessarily the best driver, but is basically competent. Let's just hope it doesn't decide to drive me in to a tree for no particular reason.
And the average human driver is terrible at driving
There's tons of people on the road, so there's plenty of quite bad drivers you'll encounter every day.
But driving is relatively easy. Saying that the average driver is terrible seems like a bit of a stretch.
I'll give you that there's millions of terrible drivers. That doesn't make the average terrible, though.
The fact that we're all not constantly getting smashed into means that on the whole drivers are competent.
On a typical Texas highway not only is everyone driving 10 over, and the typical following distance is about ~1 car length. At 80 miles an hour. Lord have mercy if anything happens on the road.
Elon is a great promoter and does have visionary ideas and has often been able to execute them when he is paying attention.
He is also running a harem to get +100 or more children, seems racist for many reasons, over promises (but this comes with being a promoter), and has biased X/twitter to favour whatever ideas he currently likes.
Cybertruck and X are what you get when Elon Musk has his limiters off. SpaceX and Tesla are what you get when a competent team knows how to manage him and his ego.
Also the ideas are not completely his mostly because this isn’t novel basic research. This is applied engineering and thus one needs to grab onto ideas who are ready for implementation at scale. This means they are well known ideas.
I don’t disagree that Elon messes things a lot, especially recently. DOGE cuts are being undone via court ruling nearly as fast as he did them.
Jeff Bezos does get credit for Amazon's work like Amazon, AWS and Kindle among others. Because he did allocate the resources to this and manage the company in this direction.
That's not true. Lane-departure detection uses mostly radar and cameras, adaptive cruise control usually uses radar and sometimes cameras, and emergency break assist is often just a brake pedal sensor. Front-collision avoidance uses lidar in some cars from the last 2 years.
Lidar might very much be needed for full self driving, but it not yet used in many cars on the road today.
The writer does not have all his facts straight.
I assume they're capturing more than just images, does anyone know for a fact?
My guess is they have a list of every MAC address of every device they can find, geolocated. And then they match that to data from all those apps that ask to discover devices on my local network. Now they know how old my tv and lightbulbs are, etc etc
Wikipedia urls ending in punctuation are unreliably broken or not depending on caching and the platform, so if you put a # on the end to escape it, it fixes it, without having to worry about percent encoding.
Per the HN formatting documentation: https://news.ycombinator.com/formatdoc
<https://en.wikipedia.org/wiki/Joffe_v._Google,_Inc.>
https://en.wikipedia.org/wiki/Man_or_Astro-man%3F is another story -- since "?" is a reserved character in URLs for CGI queries. Enter the question mark anyway, and the article comes up! Why? There's a redirect without it!
For Wikipedia's gory details on technical restrictions for article titles: (note that HN properly parses this article title ending in a right-paren)
https://en.wikipedia.org/wiki/Wikipedia:Naming_conventions_(...
That’s the first problem.
> It would be simply bizarre for any browser or web server to choke on a perfectly legal dot to end a URL.
I agree? I never said anything like this. My original comment was:
> Wikipedia urls ending in punctuation are unreliably broken or not depending on caching and the platform, so if you put a # on the end to escape it, it fixes it, without having to worry about percent encoding.
I mentioned the platform specifically, which in this context could be either the server context or client context. You mentioned server/client context, as in what HN serves the user or vice versa. I mentioned that and client context inclusively. If you’re correcting me, assume I need you to show my error.
> Enter the question mark anyway, and the article comes up! Why? There's a redirect without it!
That’s the second problem - the site in question as typed - there is no redirect from https://en.wikipedia.org/wiki/Joffe_v._Google,_Inc to the version of the article that has a period, which is https://en.wikipedia.org/wiki/Joffe_v._Google,_Inc%2E if you use percent encoding.
I’m not sure what the point of your comment is.
If you put a hashtag at the end of a Wikipedia URL, then I suppose it works, until the URL already has a hashtag in it, because these are used for section headings. It's not called "escaping" anything, it's just... an empty URI fragment: a link to the top of the article?
There is also nothing preventing a Wikipedia editor from creating a redirect from the title y'all linked. In fact it's a perfectly fine idea for a redirect. The fact is that the canonical title is in US English, and in US English, "Inc." takes a period as an abbreviation.
There's nothing wrong with your workaround or your percent-encodings to escape some dubious glyph, but I hoped to clarify things and derail the thread further on pedantic technicalities. Thank you for coming to my TED talk.
It’s an escape from pedantry.
I appreciate your gentle needling, as imprecision in my words reflects an imprecision in my rhetoric, making it vulnerable to nitpicking. It’s okay to be wrong if it allows me to make a larger point in favor of my position, but at a cost to readers’ time and patience.
Thanks for your close reading and feedback, it helps.
I'm not sure what they're doing now, but I assume that they're fingerprinting everything by MAC addresses and SSIDs and other identifiers, but for their sake I hope they have stopped intercepting unencrypted wifi traffic, which is apparently a thing that was proven at court that Google Street View cars did do.
https://en.wikipedia.org/wiki/Joffe_v._Google,_Inc%2E
> Joffe v. Google, Inc. is a federal lawsuit between Ben Joffe and Google, Inc. Joffe claimed that Google broke one of the Wiretap Act segments when they intruded on the seemingly "public" wireless networks of private homes through their Street View application. Although Google tried to appeal their case multiple times, the courts favored Joffe's argument. Ultimately the Supreme Court declined to take the case, affirming the decision by the United States Court of Appeals for the Ninth Circuit that the Wiretap Act covers the interception of unencrypted Wi-Fi communications.
> Musk was sued by a group of shareholders who claimed that the Tesla boss had defrauded them with his lofty claims about the capabilities of Tesla’s advanced driver assistance tech, including Autopilot and Full Self-Driving. [1]
> After hearing the case, U.S. district Judge Araceli Martinez-Olguin claimed that the plaintiffs had failed to prove that the Tesla boss had acted with “deliberate recklessness,” .... That’s because Musk’s lawyers didn’t decide to argue that Tesla’s claims about its self-driving abilities were perfectly accurate. Instead, the legal team representing Musk basically said that nobody would realistically believe what Musk was banging on about... [1]
> In a mind-numbing statement, Musk’s lawyers argue that his claims about Tesla Autopilot safety were “vague statements of corporate optimism are not objectively verifiable” [2]
> The lawyers even argued, successfully, that “no reasonable investor would rely” on many of the alleged misleading statements because they are “mere puffing” [2]
Never trust Elon's statements.
[1] https://qz.com/tesla-robotaxi-lawsuit-elon-musk-lawyers-clai...
[2] https://electrek.co/2024/10/02/elon-musk-celebrates-winning-...
https://m.youtube.com/watch?v=LpxT9TLGoLI&pp=ygUScGVuZWxvcGU...
FTA:
> we’ve turned over large portions of America to insanely rich people who have no idea what they’re doing
> It’s about oligarchy
> What happened? Elon Musk was stupid. That’s what happened.
> the stock market pretend that Musk hasn’t failed
Anyone who's seen recent FSD videos on YouTube will understand the incredible progress that has been made, and how FSD is a real-world solution whereas Waymo (geo fences, "safety drivers" etc) isn't comparable
What a sad world we live in.
It's more than that-- many people are spending a lot of time just testing FSD, filming it, actively looking for failure modes, and being quite methodical about it. We don't have the same data Tesla has about its fleet obviously, but we can all go and observe the progress being made.
Is it FSD, though? Understand that Waymo has remote tele-operators, who can intercede if necessary. How many are there? How many cars does each handle? I’m not sure the company makes that public.