A Song of “Full Self-Driving”

https://www.thebulwark.com/p/elon-musk-self-driving-fsd-tesla-tony-stark-michael-scott

by latexr • 1 year ago

100 102 1 year ago

ddalex3• 1 year ago

Full self driving is either a hard problem - requiring advanced sensors and algorithms; or an easy problem - solvable with cameras and neural nets.

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

rcxdude• 1 year ago

The notional moat is that it's easy given enough training data, which Tesla has a significant hoard of. I'm skeptical of this (they've been in that situation for a long time now), but that's the argument.

Zigurd3• 1 year ago

Tesla has imagery from lo-res cameras. If it turns out that LIDAR and/or radar is needed for AVs, the value of that data is questionable.

tomp3• 1 year ago

You simply have no idea how AI works.

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.

dboreham• 1 year ago

It will turn out there isn't a linear relationship between data volumes and successful self driving. There will be a wall somewhere that prevents it from being achieved regardless of data (assuming current AI techniques and hardware available in the foreseeable future). So what will happen is Musk and co will demand that roads be redesigned to accommodate their tech (e.g. cars network and human drivers are banned from major highways).

aatd86• 1 year ago

How would youtube help with driving?

Spooky234• 1 year ago

That’s the Muskophile babble. Cool story, where is it?

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.

ricardobayes• 1 year ago

It's not that car companies are unable, but rather, margins on new cars are pretty thin so they are trying to save every last penny.

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.

rsfern• 1 year ago

This is sort of a misguided take IMO - as if LiDAR and other sensor streams are somehow not large scale data while video is

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

brk• 1 year ago

Almost every vehicle has multiple cameras in it today, plus cellular connectivity. Tesla doesn't have a major advantage here. Plus, the data isn't immediately valuable on its own, it needs to be labelled and fed back into the training. This is all very far from a data flywheel analogy.

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.

fsh• 1 year ago

Transformer LLMs are good at generating text since they were trained on huge volumes of high quality texts. There is no reason to assume that feeding a neural net with unlabeled garbage data (i.e. video recordings from random cars) is going to lead to anything.

jillesvangurp• 1 year ago

I agree with you on the moat side. Tesla is very late to this market. But, I would also argue that Waymo's moat isn't that wide either. Waymo is probably behind their Chinese counterparts that are already operating at a larger scale. More rides. Larger cities. And more of them.

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.

jonstewart• 1 year ago

I wonder how long a Chinese FSD car would last on the BWI parkway coping with Maryland drivers before it pulled over for a good cry.

Zigurd3• 1 year ago

Remember 10 foot tall Ivan? Lots of people are thinking about the Chinese similarly. Waymo is probably not behind the Chinese. Huawei has the potential to make good AV software, but they're not caught up to Waymo yet.

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.

Alive-in-2025• 1 year ago

There's also no moat because zoox and waymo will compete and absorb /reduce any profit that Tesla finds if they ever deliver a working system. waymo has been scaling and has 250k paid rides a week.

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.

steveBK1236• 1 year ago

I had a Tesla 2018-2022, and switched to a competing EV brand at that point. The fact that Tesla has removed sensors since then while competitors have added says a lot.

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.

rimunroe2• 1 year ago

For example, it has actual working adjacent lane blindspot detection

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

ChoGGi2• 1 year ago

It even flashes lights on the doors if you try to open a door when a car is coming in the adjacent lane.

How's it do for bikes? If it can catch them as well, mind saying the brand?

onlyrealcuzzo2• 1 year ago

How hard the problem is may depend on how successful you are at lobbying politicians to force your solution down the public's throat...

boshalfoshal• 1 year ago

You are thinking about "hard" and "easy" in the wrong frame of mind. What Tesla does is not "easy" either. Their moat is manufacturing and the R&D they've spent on codesigning their HW and SW stack, and their insane supply chain.

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.

walthamstow• 1 year ago

The thing about Tesla having more cars on the road is obviously true, but Google Street View captures my house in London at a decent frequency, about twice a year, so those cars are driving around capturing data all the time in major cities.

I assume they're capturing more than just images, does anyone know for a fact?

xnx• 1 year ago

Lidar was added in 2017 https://en.m.wikipedia.org/wiki/Google_Street_View#:~:text=L...

flir• 1 year ago

Thanks! I thought the 3D models in Google Maps were just Extremely Clever Math. Lidar and Extremely Clever Math makes more sense.

prepend• 1 year ago

They were sued, and lost, for capturing and analyzing data off all the WiFi networks they drove by. [0] I assume they are recording all the metadata and data that they can to stay compliant with this.

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

[0] https://en.wikipedia.org/wiki/Joffe_v._Google,_Inc.

aspenmayer5• 1 year ago

[0] https://en.wikipedia.org/wiki/Joffe_v._Google,_Inc.

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.

aspenmayer5• 1 year ago

I assume they're capturing more than just images, does anyone know for a fact?

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.

aspenmayer5• 1 year ago

Scooped by ~2 min by users xnx and prepend, but I like to think my flavor text was worth the wait. I'm likely biased on that point.

amarcheschi• 1 year ago

The title made me think of something different, well, there actually is a song about musk and its cars, although it is from 4 years ago

https://m.youtube.com/watch?v=LpxT9TLGoLI&pp=ygUScGVuZWxvcGU...

bhouston3• 1 year ago

Maybe Elon isn’t Stark or Scott and the world is more complex than simple analogies.

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.

krapp2• 1 year ago

I think it's more accurate to say other people have been able to execute on his ideas (which aren't even entirely his) when he isn't allowed to interfere.

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.

bhouston3• 1 year ago

Yeah other people have executed but this actually true of anyone who starts companies and employs tons of people. He isn’t a sole practitioner in a lab.

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.

kklisura• 1 year ago

For years, Elon Musk has been promising that Teslas will operate completely autonomously in “Full Self Driving” (FSD) mode. And when I say years, I mean years...
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-...

cbeach2• 1 year ago

To save your time, this is another article that focusses on the author's personal (and political) affectations with Elon Musk, as opposed to being a substantive technical review of FSD.

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

sega_sai• 1 year ago

Are you saying that the factual content of the article is incorrect? If not, yes the author may have views on Musk, but that is irrelevant in the context of the article.

epgui2• 1 year ago

Look, I can see just as well as anyone else that Musk is wildly off on his timelines, and I really don't like the guy either. That said, the article is full of facts that are decontextualized or presented in a misleading/misled manner.

zippothrowaway• 1 year ago

Becasue youtube videos now pass as data rather than self-interested anecdotes.

What a sad world we live in.

cbeach2• 1 year ago

Hundreds of thousands of hours of people physically demonstrating FSD technology in the real world, with production cars and real users (who are not employed by the manufacturer) is meaningful evidence, even if it happens to be published on YouTube

epgui2• 1 year ago

self-interested anecdotes

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.

jp57• 1 year ago

“Of course we already have full self-driving technology. It’s called Waymo…”

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.

ddalex3• 1 year ago

They demoed what the operators do, they can only review the data and select one of a few possible options, they never take control of the car

420official• 1 year ago

Waymo is firmly SAE self driving level 4 "automated driving" vs Tesla at level 2 "driver support". You're correct that waymo has operators that can jump in in the event something unusual happens, but the vast majority of the time the vehicle is operating autonomously within their defined area. All that is described by SAE level 4 where no human is driving while automatic driving is engaged.

https://www.sae.org/blog/sae-j3016-update

Zigurd3• 1 year ago

Google still carries trauma from buying Motorola, and all of Motorola's employees. They would never have expanded Waymo if it meant hiring buildings full of remote operators. And they've demonstrated that that's not how it works anyway.

x187463• 1 year ago

FSD is incredible in that it works everywhere, most of the time. The problem is it works most of the time. I do not think the current FSD sensor suite (camera only) will ever be safely capable of true unsupervised FSD. I use FSD every day. I would seriously estimate FSD has put 95% of the miles on my car. I regularly encounter the following:

* 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.

op00to3• 1 year ago

I don't want to be seen as defending Tesla's overpromising and underdelivering, but everything you mentioned I see on a daily basis from all makes and models of motor vehicles on the roadway. Of course, FSD should be better than humans, but it's somewhat humorous to me that right now it's got the same problems as people.

ddalex3• 1 year ago

It’s not humorous it’s expected - when you train your system on data from millions of average drivers, why would you be surprised that it behaves like an average driver?

And the average human driver is terrible at driving

Coeur• 1 year ago

"Lane-departure detection, front-collision avoidance, adaptive cruise control, emergency break assist—all of these features are powered by lidar."

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.

A Song of “Full Self-Driving”

Related Stories