Hacker News Reader

Comment by ra7 | Hacker News Reader

Tesla has been kept alive only because they are scared to even let drivers take their hands off the wheel. They have very low confidence in the system to allow for driver inattention, let alone completely remove the driver. Numerous software “rewrites” over the years with many buzzwords attached, and they still haven’t clocked a single driverless mile.

They also have a huge fan base (investors) who are happy to babysit FSD in order to provide more “training data” and they’ve rationalized themselves whatever little benefit it offers is worthy of the product name.

View on HN

To be honest, I'm surprised quite how poorly teslas FSD performs. They have plenty of funding, plenty of smart minds, orders of magnitude more training data than anyone...

Sure, they have hobbled themselves with lack of LIDAR - but it seems that planning and decision making is the stumbling block for most systems, not object recognition.

But still, I think there must be some poor management or a lot of bad luck in that team not to see more success.

They have a relatively small team when compared to Waymo, at least at the research/state-of-the-art level. Don't quote me on this but just look at the the difference between autopilot team size vs Waymo employees size who have a PhD in a related area. It'd at least 10x more ay Waymo, probably even more.

Hardware and HD maps makes a lot of difference too. Planning and decision making always becomes easier when you have a perfect estimate of the environment and good predictions of other agents behaviors. Intent matters a lot more than just recognition/detection in this setting.

To be fair I'd say Autopilot team is doing pretty good for the team of their size. They are no where close to Waymo, that's super clear, but if they focused all of their stack on making driver assistive tech rather than trying to push towards FSD they'd probably be better off.

I also see questionable engineering decisions. For example, some reverse engineering showed that there was an overlay map of problematic stop signs and road signals that was downloaded per-road on all navigation routes. Surprise surprise, as that data got out of date, the car would randomly slam the brakes on as the overlay data told it there must be a stop sign, when in fact it was a temporary sign from construction months ago.

If you're aiming for end-to-end neural networks, you shouldn't have hard overlay facts like this. Ideally you get all the behaviour you need from curating training data, but if you can't do that the most you should have location specific info-vectors that nudge the network into making the right decision if needed. Such info-vectors can be used for all kinds of things, like "this city has aggressive drivers", "this street has a blind corner", "there's a dog at this house who loves to run under wheels", "this country drives on the other side of the road", etc.

If the list is of objects they can't yet reliably detect, then how would they implement your suggestion to detect those objects? I'm sure they have a lower tolerance for false negatives, so of course the known problems must be hard-coded.

Lets imagine there is a stop sign that is half hidden by a bush and pretty faded, so the stop sign detection logic says "only 5% chance it's a stop sign". That in turn isn't enough to make the car stop.

The hardcoding approach says "This is a sign. 100% sure.".

The vector approach says "There's a hard to see stop sign around here, boost up the probability of anything stop-sign-ish a bunch".

The difference functionally is that nothing in the real world is ever 100% certain. So you should never tell any bayesian machine (which a neural network effectively is) that anything is 100% true.

The vector approach I outlined is far more general than the above though - it allows any behaviour of the car to be tweaked automatically or manually. location-specific vectors can be learned from data, and/or put in by operatives. The way the neural net trains, the meaning of a vector could 'evolve' too - for example, whenever a human puts in that there is a hidden stop sign, the neural net might learn that that means other human drivers might occasionally fail to see the sign and stop in those locations. Even though it had never witnessed a human failing to stop in this specific location, it has learnt that is part of the meaning of this vector.

I'm saying it's very likely that the business has exactly zero tolerance for the detector missing those signs while there are known, unresolved issues. They can't have something happen when they "knew it could happen".

After "resolving" the issue, then they can resume tolerating the normal probabilities.

In the US, it might be more of a precaution. In Europe, it could be more of a legal obligation. Either way, I doubt engineering has a say in the matter.

> They have a relatively small team when compared to Waymo, at least at the research/state-of-the-art level.

Sure, but in every independent test, Tesla's smart driving capabilities get completely trounced by Subaru EyeSight. Does Subaru also have vastly more PhD researchers than Tesla?

Subaru EyeSight works very well in my experience, but it's much less ambitious. It basically just does lane keep assist, automatic braking, and dynamic cruise control. It can't be used at all in heavy precipitation and will automatically deactivate. The cruise control also has an annoying tendency to unnecessarily slow down on the freeway when you're in the slow lane and the vehicle in front takes an exit ramp.

I'm sorry but there's no way that Subaru EyeSight can "trounce" FSD. Maybe it is slightly better in some scenarios than legacy Autopilot which came out many years ago and hasn't been updated since. Tesla FSD is vastly more capable and can handle city driving with traffic controls whereas EyeSight merely just follows a lane.

> I'm sorry but there's no way that Subaru EyeSight can "trounce" FSD.

I mean that's what literally all of the data shows, e.g.:

https://www.caranddriver.com/features/a24511826/safety-featu...

AEB has nothing to do with self driving or Autopilot.

It's literally the most foundational part of self driving.

Not really.

1. the AEB may run on a different codepath than FSD

2. AEB needs to decide when it is appropriate to override the driver, e.g. if the driver is behaving as though they are paying attention, e.g. by repeatedly stepping on the accelerator, it is undesirable to override the driver's deliberate actions even if it appears that the driver is about to run into something. On the other hand, when the car is driving autonomously, it has full control to avoid running into anything at all.

Also, instead of trusting unscientific tabloid news with inconsistent methodologies, why not use the standard assisted driving scores from official government tests, such as the Euro NCAP? Teslas score by far the highest. https://www.euroncap.com/en/ratings-rewards/assisted-driving...

You have made a solid argument for why we should believe the Euro NCAP scores so I will accept their assessment that Tesla is by far the worst.

Your link shows Tesla scores 36/100 in assistance competence, the lowest score by far. The second lowest is 52/100 and the average over all models (excluding Tesla) being ~67/100.

Oh weird haha idk why that link doesn't show the later scores but the Model Y and Model S both score 98/100 in Safety Assist:

* https://www.euroncap.com/en/results/tesla/model+y/46618

* https://www.euroncap.com/en/results/tesla/model+s/47760

which is by far the highest out of all models tested.

Your original link was to the “Assisted Driving Gradings” which evaluates ADAS systems such as Tesla Autopilot.

What you just linked is the AEB (and other collision avoidance/warning) rating and “AEB has nothing to do with self driving or Autopilot” as you said, so is totally irrelevant.

So, I stand by your position that we should trust Euro NCAP which means concluding Tesla Autopilot is by far the worst system.

In that "worst system" result it actually scored full marks on Adaptive Cruise Control Performance and got dinged points on subjective points such as "Driving Collaboration". The conclusion states that it excels at driving but the design may cause over-reliance. Since the 2020 test, Tesla has added a lot of new features such as driver monitoring via in-cabin camera which is extremely good at detecting whether I am focused on the road or looking at my phone.

For what it's worth, I have lots of experience with both Autopilot (both legacy and FSD) and various cruise control implementations. Tesla is honestly just vastly better.

So, we should “use the standard assisted driving scores from official government tests, such as the Euro NCAP” when you thought “Teslas score by far the highest”, but now you say the test sucks and has lots of methodology problems.

Yes I am very bad at arguing. But it doesn't change the facts which is that Tesla FSD is vastly better at driving than anything else out there.

It looks like of the features both systems have, Subaru EyeSight performs better... But Tesla FSD has vastly more abilities (some of which it performs well at, some middling, and some poor).

I'm not following the news on this, but what metrics do these independent tests use and what aspects do they cover? Are they driver assistance features?

they don't use lidar, which seems like a strange choice at this point.

On the contrary I'm quite surprised how well FSD performs. Of course, it's not fully autonomous like Waymo, but here in the Bay Area it works remarkably well on the suburban stroads outside of San Francisco. I basically never have to intervene at all. Moreover, when you read news like how Cruise actually has a manual intervention every 5 miles [1], it seems that actually Tesla is not really as behind as some naysayers claim.

[1] https://www.cnbc.com/2023/11/06/cruise-confirms-robotaxis-re...

Tesla cannot abandon FSD, if they did Musk, and by extension (and due to all kinds of financial connetctions and dependencies) all his other companies would be in very serious trouble.

Why is everyone so overly dramatic about this stuff. I remember when FSD was announced on HN everyone knew it'd be way longer and he was just pumping it up cuz that's how he is. It wasn't some big surprise.

Now everyone is acting like it's do or die, as if everyone bought Tesla's just so they won't have to ever touch a steering wheel at some point in the future. Only 19% opted for it and plenty of those are probably wealthy people who can afford novelties and all that matters at the end of the day is if you're competitive with the alternative options on the market (holistically not just feature to feature). Ultimately it's still just a car people bought instead of a BMW or whatever. They're not making car purchasing decisions comparing it to some beta project only running via taxis in silicon valley and some random small cities.

Because Teslas market gap is based on the assumption that it is a tech company. This assumption is, basically, only FSD and Musks image as a tech genious. And a lot relies on Teslas valuation, including for a time SpaceX funding.

Officially announcing that FSD is a dead end takes allnof that away, and Tesla might very well be valuated as car company (which it is). And that share price adjustment could, or will, be a very serious problem Musk and his enterprises.

> Teslas market gap is based on the assumption that it is a tech company. This assumption is, basically, only FSD

Energy. Batteries. Charging network. Top selling vehicles (Model Y is best selling car worldwide)…

Energy is a tiny part of Tesla. Auto revenue was $19B last quarter, energy and storage was $1.5B.

Yes, which is why I also said they sell the #1 selling car in the world. I was responding to someone who said their valuation is based on FSD.

There was a time VW sold more cars in a year then Tesla did in its entire history (they still do by the way, a lot more), had the best selling car in the world, the VW Golf, for years in a row, was profitable, on its way to surpase Toyota as the biggest car manufacturer by units sold and had nowhere near the multiples in the valuation that Tesla has.

So no, having the best selling car is by no means a reason Tesla share prices are where they are.

Toyota has four of the top ten selling cars and is worth $290B. Tesla has two and is worth $792B. Tesla the carmaker is a valuable company but not more than Toyota.

You could make the argument a few years ago that Tesla was only electric car maker and would capture the market as it grew. But now lots of companies make electric cars. Electric cars are booming, Tesla is growing fast, but they are losing market share.

That's just rubbish speculation on your part. All evidence points to FSD having a very marginal effect on Tesla's market cap. The market has already discounted it as nothing more than a driver assist system.

Most of Tesla's valuation is easily attributable to the fact they sell a ton of cars, with very high profit margins, and are on track to sell even more in the future.

There are nicer ways to challenge the GP's argument.

Because that's what Musk has promised many years going. It's snake oil sold to the gullible because it's highly unlikely that the processors and sensor array in the production cars sold (with this feature) will ever be able to actually do FSD - no matter how advanced the software becomes.

If a company consistently lies to its customers and the market as a whole, shouldn't there be consequences?

HN is just generally awful on this topic and not worth commenting in imo.

Tesla is very well positioned even if they did zero on FSD (which they won't). The Model Y was the best selling car on earth across all categories (including ICE and EVs!) - not sure what more you could want for car company success.

I know that to be the case. I've checked the numbers from reputable sources numerous times.

But my mind still fails to comprehend this. It's a $80k cad luxury idiosyncratic vehicle. How's this out-selling civic and corolla, kia and hyundai base models? Have millionaires outnumbered regular folks by that much worldwide? I just don't understand that number if it's talking by units sold (and not revenue or profit). My entire world-view fails to make sense of that factoid.

Model Y is definitely not a luxury vehicle in the US. I'm not sure about in Canada, but in the US the Model Y AWD only costs about $15k more than a mid-tier AWD Honda CR-V or Toyota RAV 4 trim (I just recently priced out options for our next vehicle). You'll make up that difference over the first 5-10 years in gas and maintenance savings, and then tax incentives shave off another $7.5k or more depending on the state at low/mid income levels. In other words, it's at most marginally more expensive than a Civic or Corolla, and on par with compact SUVs. The same is more or less true in China. See tons of Teslas there.

Tesla Model Y, according to website, starts at 70k CAD.

Toyota RAV4, according to website, starts at 36K CAD.

So that's plus or minus twice the price :-/

Other way of looking at it - even if it were "Only $15k USD difference" in USA, that is an entire car worth of difference (not only in pre-COVID times, when you could buy a new car for $10k! We just bought a $17k Kia Rio 5. Tesla Y is four time the price, and ~2.5 times the price of Corolla or Civic, which again I wouldn't think of as "marginally more expensive". (Edit: And this does not include the money one would presumably need to spend to install chargers etc at home).

Please don't take this the wrong way, but it's a somewhat Hacker-News specific perspective that Tesla isn't luxury, or that $15k USD difference isn't much. Acura RDX starts at $57K CAD, which is 12K CAD less than the Tesla Y, and I'd certainly consider that a luxury car. Other people might not! But that underscores my point that for high earners, "luxury" and "rich" creeps up. (same for me too! I used to think $100k a year was "rich" until I got there and then I conveniently moved up my own personal definition. I don't consider our Honda Odyssey a luxury internally - I consider it a "necessity" or a "practical choice" until I have a "whoa" moment and remember that my parents took five people around in a little Citroen AX and Zastava 101, and that vast majority of people cannot afford a $55k CAD family car - and even we got it used from a 2 year lease because $55k CAD for a car is just a lot lot lot of money)

Yeah, this is basically the reason.

The price is comparable, the costs of ownership are way lower, and the product is at least twice as good.

There's basically no downside which is why they're selling so well.

It's cherry-picked statistics. It's just talking about a single model. Tesla has far fewer models than the larger automakers, so total sales are more concentrated in any one model.

Tesla has sold about 5 million vehicles in its entire history. The last time Toyota sold less that 5 million vehicles in a single year was 20 years ago (they do close to a million vehicles a month now [1]). VW Group is not far behind Toyota.

----

1 - https://www.investopedia.com/toyota-global-output-hits-recor...

I've sent it reported by several websites but I've never seen the primary source of that "fact". Lots of websites source Motor1 (some fly by night website I've never heard of) who source but do not link JATO Dynamics, another source I've never heard of, who has supposedly compiled "multiple sources" to find this figure.

I share your scepticism.

A lot of people are no longer buying new cars. Of my groups of friends, only the very wealthy and well paid ones have done so. Half of them very reluctantly because the used car market getting so crazy. In fact a lot of the craziness of the used car market is down to increasingly people can’t afford a new car.

Base Model Y starts at $43k USD. Then there's probably some electric incentives to bring the even lower depending on where you live, so presumably you can get one in the mid to upper $30's.

That's not bad for a new car, especially when the general buyer I think mostly considers it a luxury brand.

Also the floor for car prices has gone up. The days of the near $10k new car are over.

> Have millionaires outnumbered regular folks by that much worldwide?

I'm far from a millionaire and a $40-50k car is not unreasonable, my wife's car is in the low $50s. But I have a Kia Forte that was like $25k. A small Kia for $25k! But it was the cheapest car I could find on shortish notice that had lance centering and whatnot.

USA and Canadian pricing is clearly different, fascinating! We just got a Kia Rio 5 for $20K CAD (starting price is $17K CAD). Which is less than 25% the price of Tesla Y starting at $68-70k (see my sibling comment a bit up for breakdown details). Absolutely different cars in size, features, capabilities, etc, yes! But for me, in Canada, Tesla Y is very much in the luxury category :-/

a 2024 base model civic in the burbs 2 hours out of NYC goes for $23,950, and will almost certainly have better QA than the tesla.

the media salary in the US is like 55k is a lot, esp. given the prices of everything going up, and given the alternatives.

Yeah I wouldn't buy one, was just trying to make the point that it wasn't so expensive that it didn't have potential to sell well just based on price, given the parent millionaire remark.

HN has a lot of people that hate Elon Musk because they disagree with his politics. It's clear to anyone that is on here every day since his twitter became popular.

If you look back, 10 years ago HN loved him (see convo about first time 2 falcon 9s landed), but now that he is the richest man in the world the tides have turned.

Ten years ago was before he started calling rescue divers pedophiles.

and ruining twitter by recreating voat aka parlor 2.0

and don't forget all of the anti-union causes he donates to

You say that like there's something wrong with disagreeing with anti semitism. Of COURSE people disagree with his politics. You'd have to be fucked in the head not to. HN isn't Stormfront, you know. His behavior changed and new facts arrived since he used to be popular, so reasonable people changed their minds about him, and unreasonable racists are now his new biggest fans. The reason people hate him is not that he's the richest man in the world, it's that he's a bigot.

I'm stunned that I need to explain that to you. Is it actually news to you? If so, welcome out from under your rock. Or are you just pretending not to know, in order to troll?

How is he a bigot? What has he specifically done that is racist or anti-semetic?

I don't care that he called someone a pedophile. He apologized for that: https://www.pbs.org/newshour/world/elon-musk-apologizes-for-....

He is guilty (as in they lost this lawsuit for millions in damages) of letting the Fremont, California factory run like this [1]:

Robert Hurtado, a supervisor at Tesla, called Diaz the N-word over 30 times, including by saying “[N-word], hurry up,” “[N-words] are lazy,” and “I wish I can get all you [N-words] fired.” Tr. 3:604:23-606:8. Ramon Martinez, another Tesla supervisor, called Diaz the N-word over thirty times, said “I hate you, [N-word],” and told him “to go back to Africa.” Tr. 3-603:11604:20. Diaz's supervisor, Tamotsu “Tom” Kawasaki, testified that Martinez often used the words “mayate” and “chongo,” both racial slurs in Spanish. Tr. 2-281:9-12. Wayne Jackson, a Black program manager at Tesla, confirmed that Martinez used the word “mayate.” Tr. 2-381:17382:7. Martinez denied all of this. Tr. 3-584:18-585:1.

Diaz also testified that he complained to Romero several times about being called the N-word. Tr. 3-613:5-14. Romero agreed that he learned Diaz had been complaining about racist conduct, see Tr. 3-542:22-24, 543:13-14, though Romero said Diaz did not complain directly to Romero at least until the Timbreza incident, Tr. 3:561:1-4. Romero testified that he did not personally investigate Diaz's complaints and was specifically told by Tesla to not personally investigate the Timbreza incident. Tr. 3-541:22-542:1, 544:21-545:2, 549:13-550:16. Diaz and Romero both testified that Tesla never fired anyone for using the N-word. Tr. 3-552:10-19, 615:13-15.

Jackson testified that the use of the N-word was “[r]ampant,” as in it was used “[e]very day,” “[m]ore than several times a day.” Tr. 2-381:3-15. He found the N-word “[e]xtremely” offensive and “very derogatory” when it was used, regardless of how it was said. Tr. 2-383:15384:2. He said that these words should not be used in a workplace, but that there was “nothing” he could do to stop their use “[b]ecause Tesla allowed it. They didn't stop it.” Tr. 2-384:12-22.

[1] https://casetext.com/case/diaz-v-tesla-inc-7

That case that focuses on allegations against Tesla as a company, not directly against Musk.

There are uncountably many horrible things I don't need to insincerely issue a non-apology apology for at the insistence of my legal department, because I never did them in the first place.

You've got a lot of catching up to do from your 10 year absence from reality. I'm not going to lead you to water because you'll refuse to drink. Go google it for yourself, and stop trolling.

You didn’t answer any of his questions, which are not as ludicrous as your scathing reply implies.

Read the Jean-Paul Sartre quote below.

Let me ChatGPT that for you:

This quote is from Jean-Paul Sartre's work "Anti-Semite and Jew" (originally "Réflexions sur la question juive"), written in 1946. In this passage, Sartre analyzes the behavior and motives of anti-Semites. He argues that anti-Semites are aware of the absurdity and frivolity of their arguments. According to Sartre, they engage in such discourse not to persuade through rational debate but to amuse themselves and to disconcert their opponents. They do this by undermining the seriousness of any conversation, thus avoiding genuine discussion. When pressed, they may cease the discussion, indicating that they are above argumentation. This approach, Sartre suggests, is a form of bad faith, a concept he discusses in his existentialist philosophy, referring to acts of self-deception or dishonesty in one's dealings with oneself and others.

Can you connect this quote to the conversation at hand?

I will save you the time, no they cant.

Nice attempt at deflection, but it applies directly to YOU.

Exactly the response I was expecting. Throw out some huge accusations, and dont back them up with a single example. Solid work Don!

Have a nice day.

Exactly the response I was expecting from you too, since I just said you were pretending not to know in order to troll, and that confirms it.

I don't owe you any examples, because you know very well but are pretending not to know what I'm referring to, you're not a serious person, you're intellectually dishonest, and you're trolling, so your bad faith arguments don't matter, and you don't deserve a response, troll.

“Never believe that anti-Semites are completely unaware of the absurdity of their replies. They know that their remarks are frivolous, open to challenge. But they are amusing themselves, for it is their adversary who is obliged to use words responsibly, since he believes in words. The anti-Semites have the right to play. They even like to play with discourse for, by giving ridiculous reasons, they discredit the seriousness of their interlocutors. They delight in acting in bad faith, since they seek not to persuade by sound argument but to intimidate and disconcert. If you press them too closely, they will abruptly fall silent, loftily indicating by some phrase that the time for argument is past.” ― Jean-Paul Sartre

Yeah the irony is he used to get a lot of hate from the right (because of environmentalism and EVs being left-coded) while the left loved him, but it's flip-flopped. I think partly because the media is directly threatened by twitter - there's a conflict of interest, but also because the progressive left has gotten worse (though I think the 'hate the rich' bit is part of it).

The same types of people (inaccurately) labeling Musk an antisemite are often simultaneously pushing/defending DEI offices and ideology at universities that cause actual direct threats against Jewish students.

The biggest reason why I haven't traded in the Tesla for something else is FSD. No other system compares. With FSD, I can do 6-8 hour segments where I actually drive the car less than ten miles total.

The Supercharger Network is nice as well but Electrify America and evGo have been scaling hard as well and they work for the most part. Not like that matters because the Supercharger Network is being opened up.

Why would they want to abandon FSD? I look up videos whenever these threads come out and over the last few years it seems to have made astonishing progress.

2 million miles per day of FSD driving is impressive. The resulting training data from every corner of the USA is a bonkers advantage.

Only if you think this is actually a model training problem.

Their existing system is pretty strong evidence that it was in fact a model training problem -- do you mean to say that you think training will stall out before it gets to the point of being viable?

They still haven't removed the driver, which Waymo did in 2018 when AI was much less advanced than today. Therefore it's pretty clear there's some fundamental limitation that probably has to do with hardware and/or inability to build proper guardrails for their AI, as opposed to the capability of the models. In other words: building AI models that can drive impressively in the common case is likely the EASY part of the problem, and they haven't made much progress on the hard part, which is likely dealing with the long tail.

Should you remove the human and then expand scope (Waymo)? Or expand the scope and then remove the human (Tesla)?

I dunno. I personally think the Tesla approach is fundamentally better because it obtains more data, faster. Millions of miles per day rather than millions of miles per year, with better macrodiversity to boot. However, that's just a heuristic: we've got a Tortoise and Hare scenario and as the parable teaches us, it's not so difficult to imagine a scenario where slow and steady wins the race. Even though I favor the Tesla approach, I could see this going either way.

That said, I do think it's uncharitable to read failure into Tesla's high level decisions. They could have chosen the Waymo path and dumped all their effort into maximizing reliability on a small scenario subset. They chose not to. There are Andrej Karpathy interviews about it. They wanted to data-max. They data-maxed. Now they have 300x the data collection rate of Waymo. That's not failure, and it may even be a killer advantage. We'll see.

> They wanted to data-max. They data-maxed. Now they have 300x the data collection rate of Waymo.

Citation needed.

What Tesla lacks is quality data, which Waymo has due to their high fidelity sensors. Tesla has been "collecting data" for years without much to show for. They still have embarrassing failures so basic in nature [1], you know they're going nowhere. If all it took was data, they would have something meaningful by now.

[1] https://twitter.com/elaifresh/status/1727889381309239731

Tesla's data isn't just more volumnious than Waymo's, it's also higher quality. Macrodiversity is important. LiDAR isn't. Scene reconstruction got good in the last few years, and that's what LiDAR was competing with. Meanwhile, there's no substitute for macrodiversity.

Anyway, you asked for citations. Here they are.

Waymo:

> Waymo was the clear leader in 2022 with 2.9 million miles with safety drivers

https://www.eetimes.com/waymo-cruise-dominate-av-testing/#:~....

Tesla:

> FSD Now Drives 1 Million Miles Per Day

https://www.notateslaapp.com/news/1327/tesla-fsd-milestone-1....

> Graph showing 2M miles per day

https://digitalassets.tesla.com/tesla-contents/image/upload/...

> Tesla has been "collecting data" for years without much to show for.

Except, you know, millions of miles per day of self-driving in situations that Waymo isn't even planning to attempt for years.

Waymo has millions of driverless miles at this point. Tesla has zero. I'm not sure how Waymo's safety driver miles is relevant here.

> Except, you know, millions of miles per day of self-driving in situations that Waymo isn't even planning to attempt for years.

And what exactly have they got to show for it? All that data collection and they can't handle the easiest of situations in San Francisco. You don't need nationwide data not to run stop signs or swerve onto oncoming traffic.

Safety-driver miles are more useful for training. Driverless miles are more useful for proving. Like I said, different paths: remove the human then expand scope vs expand scope and then remove the human. Waymo is undeniably further along path 1, Tesla is undeniably further along path 2. Which path is shorter? I'm pretty sure it's path 2, but time will tell.

> All that data collection and they can't handle the easiest of situations in San Francisco

Unlike Waymo, they haven't limited to and exclusively trained on the easiest situations in SF for the last decade. Just like Tesla hasn't even attempted true driverless miles, Waymo hasn't even attempted unmapped back-alley navigation, let alone unmapped back-alleys in the middle of nowhere. You need scope and reliability to win this race, and it seems to me that the Tesla formula for improving reliability scales better than the Waymo formula for expanding scope. Time will tell.

> Safety-driver miles are more useful for training.

Waymo does most of its training in simulation. They do 20+ million miles of simulation every single day with 25,000 "virtual" vehicles. They don't even have to put out many cars for training.

> Unlike Waymo, they haven't limited to and exclusively trained on the easiest situations in SF for the last decade.

I've got a bridge to sell you if you think driving in SF is easy or that Waymo is only doing it in the easiest situations. They cover the entire city 24/7, and in rain and fog.

> Just like Tesla hasn't even attempted true driverless miles, Waymo hasn't even attempted unmapped back-alley navigation, let alone unmapped back-alleys in the middle of nowhere.

And they won't because maps are clearly a big part of their tech stack and they've found a way to efficiently maintain them. That's what gives them the confidence a.k.a reliability to remove the driver.

> You need scope and reliability to win this race, and it seems to me that the Tesla formula for improving reliability scales better than the Waymo formula for expanding scope.

So far, evidence suggests Waymo is expanding both scope and reliability faster than Tesla. They just announced airport terminal rides in PHX today. Community trackers (absence of Tesla sharing data) show Tesla's intervention rate hasn't really improved much at all and they still seem to make extremely basic mistakes. So the formula clearly isn't working as of today.

Tesla has simulations too, of course. Thing is, the world has fractal complexity and synthetic data only captures the variation that you tell it about. Simulations for synthetic data generation are literally my job -- this is where my intuition is at its strongest.

> if you think driving in SF is easy

I'm not here to trivialize SF driving, I'm here to say that it's a small corner of a big weird world that Tesla has done a much more thorough job of exploring. Google and Waymo are used to having a scale advantage, but Tesla's fleet utterly dwarfs them.

> Waymo is expanding both scope and reliability faster than Tesla. They just announced airport terminal rides in PHX today.

That's nice, but FSD already drives almost everywhere in the USA. It can usually even figure out poorly mapped rural roads in flyover country. What's Waymo's timetable for those?

> So far, evidence suggests Waymo is expanding both scope and reliability faster than Tesla.

FSD has gone from private to "mostly works everywhere in the USA" in the last 3 years. It has taken Waymo 10 years to start escaping the hubs one city and route at a time. They don't even use freeways, last I checked. Slow and steady is a fine strategy, I wish them luck in the marathon, but it's a bit early to call this one for Waymo.

I'm used to holding back the enthusiasm in Tesla circles where we are naturally predisposed to call the race early for Tesla, so this is a fun change of pace.

> I'm not here to trivialize SF driving, I'm here to say that it's a small corner of a big weird world that Tesla has done a much more thorough job of exploring.

Then explain to me why it can’t stop for stop signs in SF or can’t drive in a straight line without swerving onto oncoming traffic. Surely, they have a lot of data from the city with their large fleet? Either the data is helping or it’s not.

> That's nice, but FSD already drives almost everywhere in the USA. It can usually even figure out poorly mapped rural roads in flyover country. What's Waymo's timetable for those?

Yes, with a driver. And still with poor intervention rates after years of development. What’s Tesla’s timetable for getting their reliability up and removing the driver? That’s really the ballgame in self driving.

> FSD has gone from private to "mostly works everywhere in the USA" in the last 3 years. It has taken Waymo 10 years to start escaping the hubs one city and route at a time.

“Works everywhere” is doing a lot of heavy lifting here when you rely on a driver to prevent crashes. Literally does not pass the true test of self driving.

It took Waymo 10 years to completely remove the driver. Now they’re announcing new cities and capabilities constantly, with freeways coming soon. Let’s see how many “rewrites” it takes for Tesla to tell drivers not to pay attention, let alone completely removing them. I wish them luck (as a Model 3 owner), but they’re working with their hands tied behind their backs on this one.

As a vehicle user, and not a stock holder, there is only heuristic that will determine whether I get into either of their vehicles: Which one is more likely to kill me.

I'll take a ride in a self driving taxi through residential streets that might hit a dog.

I will not pay for a vehicle that will hard brake on a highway or drive under a tractor trailer.

Maybe after more deaths and data Tesla will pull ahead in that risk calculation but I will not pay to be part of that.

Of course. I feel the same way. Nobody is asking us to behave differently.

I'll admit that I am by no means an expert and might be completely wrong, but yes, I believe that model training cannot fully solve driverless vehicles.

All models that we train today have edge cases where the model produces wildly inaccurate output. As far as I'm aware, nobody yet has managed to squeeze this problem out.

ChatGPT is still an overall useful tool with the occasional hallucination. Those types of edge cases in software driving a car are unacceptable.

ChatGPT intentionally plows through low-confidence scenarios. It just tries to do its best because "I dunno" is reliably low-utility while a hallucination doesn't gamble lives. It errs on the side of hallucination by design: if you alter the RLHF feedback, you can alter the confidence level.

In FSD, as you point out, the stakes are higher. It makes a different design decision: it slows down in low-confidence situations. Just like a person. If you watch a FSD video, you'll see that most of the human interventions happen because FSD lost confidence in a tricky scenario and halted the car even though the pathing ultimately wound up being correct. This won't solve every edge case, but neither do humans -- the bar is (or should be) "better than human," not "perfect," otherwise we are sacrificing lives on principle.

I have no idea if training will stall out half way through the march of 9s. It might, but if it does, it won't be because there is no high level strategy to control hallucinations. That already exists.

Tesla sold their investors and customers on "full self driving", not "self driving where you must always have your hands on the wheel ready to take over".

Customers are willing to tolerate the latter right now because they believe the former is still in development.

Do you think it isn't?

I'm not here to argue that the overly enthusiastic timetables weren't cringe. They definitely were. I'm just here because there seems to be consensus that FSD is doomed and it doesn't remotely seem that way to me.

Of course it's in development, but if I'm right that ML cannot solve this problem then a lot of that development is wasted.

My point is that in the long run, Tesla cannot solve the ML edge cases by having a human available to take over at a split second notice.

Maybe, but I think that feedback and training data from supervised driving will continue to propel Tesla along the march of 9s until FSD has genuine superhuman performance. Of course, there is only one way to find out. Time will tell.

Right, said another way, is driving a completely deterministic activity or does it sometimes require general intelligence. If it's not completely deterministic, then you might need a human in the loop.

I agree, but i think it's getting close to be much safer than humans driving on highways.

Not at all, Tesla has enough cash on hand to pay everyone out a few billion if there was a judgement wrt FSD. They outran the liability with the brand and sales.

The comment you're replying to is not talking about court-enforced liabilities. If the pull out of FSD they're basically Mercedes with a 5 year lead on manufacturing. That is categorically not a trillion dollar company and the house of cards leveraged against Tesla stock crumbles.

I don't understand what you mean. Mercedes-Benz is currently ahead of Tesla on autonomous driving. They have a working SAE Level 3 system already. It is currently very limited but will expand over time.

https://www.caranddriver.com/reviews/a45326503/mercedes-benz...

    DRIVE PILOT is ready to chauffeur you under conditions that help ensure a secure ride.  Conditions include:
    Clear lane markings on approved freeways
    Moderate to heavy traffic with speeds under 40 MPH
    Daytime lighting and clear weather
    Driver visible by camera located above driver's display
    There is no construction zone present.

and from the manual:

    The distance [to the vehicle in front] is neither too small nor too large

How is this remotely comparable to FSD? Drive Pilot is less technologically ambitious than the 1995 CMU No Hands Across America vehicle, which happened 28 years ago. Polishing it enough to accept liability is cool, but there is no technology path from follow-the-leader to general purpose self-driving.

Mercedes Drive Pilot is aimed at a different use case than Tesla FSD so it's difficult to make a direct comparison, but independent reviews generally indicate the Mercedes system is superior. Tesla doesn't have a viable technology path to general purpose SAE Level 4+ autonomous driving either, so I don't see any difference there.

https://www.roadandtrack.com/news/a39481699/what-happens-if-...

lol

5 year lead on manufacturing vs Mercedes? Are panel gaps and technicians swapping USDM headlights to EUDM on cargo ships transporting their cars to Europe a 5 year lead now?

They’re definitely ahead in EV process. I don’t like them or Musk, but that’s undeniable. Whether that leads to a better finish is debatable.

You realize you can take this confidence and bet against TSLA right?

If you're right it'll be a relatively great payday for you.

Gamestop proved to me that the market can remain irrational much longer than I can remain solvent.

Many are shorting Tesla and continue to lose money because of the markets idiocy. If I were a billionaire I would short it, but as it stands I cannot risk losing money just because of stupid prolonged hype.

Have you ever been wrong about something or was it always attributed to someone else's idiocy?

I'm not betting against anything, my money is going to real estate me and my family will live in, not WSB gambles. Also, betting against hordes or Tesla fanboys does not seem rational at all, doesn't mean the product is any good or has a bright future still.

Nah, there is still enough collective delusion in the brand to keep the stock somewhat propped up vs legacy auto’s slow death. BYD and Tesla are crushing it comparatively and have the supply chain to keep ramping. No one else is selling meaningful volume at this time. Mercedes is certainly not the future in any meaningful way.

Stock would take a nose-dive.

Correct, FSD is baked into the future success of Tesla. It is the hype engine that has helped keep the company alive all these years. However you notice they are starting to pull back from all there promises of robotaxi's and now Musk is calling it "supervised" FSD.

If they don't think it's achievable, the smart move at this point is to shrink the team down to just a few people and claim to still be working on it, while pushing incremental improvements every few months while telling customers and investors that 'its coming'.

Any other move causes a precipitous drop in stock price as investors see that they won't be taking over the entire transportation industry.

Oh wait, thats exactly what we see.

> they still haven’t clocked a single driverless mile.

This seems maybe technically true but extremely misleading. There is of course always a driver in the car, but they have absolutely clocked many miles with a driver in the car not needing to intervene.

"Technically true" == true

"Always a driver in the car" != driverless

You often see waymos cruising around sans driver. That's real driverless.

> You often see waymos cruising around sans driver. That's real driverless.

Very cool. Where? I don’t think I’ve ever seen it anywhere in LA.

They are on a promotional tour in a LA, giving rides in specific neighborhoods. Other than that, they operate all over SF and Phoenix.

Waymos offer truly driverless rides in San Francisco. I just took one last week, without anyone in the driver seat. It's great!

On December 17th, they'll be at the Original Farmers Market in WeHo from noon til 5pm.

https://waymo.com/tour/#schedule

Join the waitlist asap so you can get access when they launch in socal.

As someone with a Tesla FSD beta - the "not needing to intervene" is a little bit of a stretch.

I would describe it as a system for rapid reaction and lane assist. It's very... very poor at self driving, outside of an empty interstate.

Very much depends where you are. It can handle most of CA well in my experience (and from a lot of videos posted on Twitter).

It's not misleading. It's the true test of L4+ self driving — do you require a driver ready to intervene or not?

I think what's interesting is how many end-to-end trips, or at least meaningfully distinct segments (eg. the highway portion, the city portion, etc.) driven without any intervention. And how repeatable that is.

You could start me in goal for the Leafs and I bet I could clock many hours of scoreless game time in a season.

There's a massive difference between needing a driver every x miles and never. Waymo has served more than 700,000 rides this year. FSD cannot drive intervention free 700,000+ times in a row (on many different routes), I doubt it can even get close.

Eh, driving the 99.99% of miles where nothing happens is easy.

It's the 0.01% of miles where an empty carrier bag blows across the road, or a fire truck is stopped in your lane, or a kid crosses the road without looking, that makes self driving hard.

as it does for real people - especially those that are on their phones while they drive.

FSD doesn't need to be perfect, it just needs to be consistently better than the typical/average driver and cause or get into less accidents than real people do.

That day is coming - it may be TSLA, it might be someone else, but sooner or later you won't even be able to afford the insurance for a car that is powered by a human if FSD provably is safer in agrregate.

Yep, this is it. It's why it's such a shame about the Cruise incident in SF - people act like it's shocking that an autonomous vehicle didn't do the correct thing when a human driven vehicle launched a person into their windshield. Of course it's terrible that the car drove over the person (and unacceptable that Cruise tried to conceal that fact), but in terms of thinking about the safety of the vehicle you really have to ask yourself whether humans would be able to handle that situation well.

Would I definitely react in the optimal way if a person hit my windshield? I wouldn't bet on it, and I would say there is a nontrivial likelihood that I would accidentally step on the gas and run over the person. That's just the nature of dealing with a truly one in a million situation that I'm totally unprepared for (just as the autonomous car was totally unprepared for it).

This is a big dramatization of the Tesla situation. No consumer vehicle has a driverless system that tolerates inattention, they are merely driving aids. Nobody on the consumer market is claiming to be anywhere near level 5 fully autonomous driving.

Tesla’s full self driving system is the only one on the consumer market that will drive on city streets for you in the form of beta software. Yes, it’s going to make you touch the wheel and prove you are alert. Which competitor doesn’t? Yes, it’s a beta that isn’t done. Yes, it will disengage at some point. Are there any other car makers with software that is further along?

A Waymo driverless cab has a roof rack and all four corners of the vehicle covered with equipment that might cost up to five figures to produce and install. It isn’t a consumer product. The whole reason consumer automakers like Tesla don’t want to bother with that equipment is that it isn’t something that will ever make financial sense. Consumer personal vehicles are cost control exercises where every dollar and cent counts. When you buy a base model Tacoma and it has an 8” screen instead of a 12” screen, that’s a change that saved Toyota $5-10 at most, and that $5 or $10 matters.

Your comment makes it sound like Tesla is in some kind of financial or competitive danger because of its autopilot/FSD feature, but I think that just isn’t true from a dollars and cents standpoint. If anything Tesla’s strategy has much higher profit margin potential due to its reliance on consumer grade equipment.

> No consumer vehicle has a driverless system that tolerates inattention...

Mercedes's Drive Pilot (L3), which you can buy in a production car today, is explicitly advertised as allowing drivers to take their attention off the road and do other things in specific situations.

It's not L5, but it'll absolutely tolerate inattention and let you e.g. pull out your iPad to wage an HN comment battle while it drives you from Berlin to Potsdam.

> while it drives you from Berlin to Potsdam.

This is maybe technically true but in practice not really correct. Some conditions to use it are (per https://www.mbusa.com/en/owners/manuals/drive-pilot):

Clear lane markings on approved freeways Moderate to heavy traffic with speeds under 40 MPH Daytime lighting and clear weather

So if it's night or raining or foggy, you can't use it for that drive. If there's no or little traffic, you can't use it for that drive. In circumstances in which you can use it, you're gonna be doing 40MPH in the right lane.

> If there's no or little traffic, you can't use it for that drive.

So it needs other vehicles to orient itself?

It doesn't, but it can't go faster than 40 mph so it wouldn't be pleasant to go 40 mph while people are zooming past you at 70 mph on the empty autobahn.

Jesus, that's REALLY strict.

Can we stop pretending that other manufacturers have better self-driving capabilities when they have such SEVERE restrictions?

I can use Autopilot at night with light rain at 80 mph. It's actually amazing how well it can read lines.

> On suitable freeway sections and where there is high traffic density, DRIVE PILOT can offer to take over the dynamic driving task, up to the speed of 40 mph.

https://group.mercedes-benz.com/innovation/product-innovatio...

> Mercedes's Drive Pilot (L3), which you can buy in a production car today, is explicitly advertised as allowing drivers to take their attention off the road and do other things in specific situations.

So Mercedes is going to take 100% liability if their system causes a crash, right?

> in specific situations

Very specific situations.

only if there's enough traffic to keep it under 40 the whole way, right?

> A Waymo driverless cab has a roof rack full of equipment that might cost up to five figures to produce and install. It isn’t a consumer product. The whole reason consumer automakers like Tesla don’t want to bother with that equipment is that it isn’t something that will ever make financial sense.

Baloney. I, and many others (and as you point out, many Tesla owners already have) would pay considerably for true self driving that actually worked. That is, if I could get true self driving (meaning I could take a nap or read a book), I would easily pay for lidar and other advanced sensors.

The reason I'd never trust Tesla's FSD claims is that I believe (like many of their own engineers) that their (really Musk's) "camera only" approach is foolhardy and unsafe.

Tesla hardware 4.0 has a radar module.

If Tesla gets their solution across the finish line, Waymo’s solution is instantly uncompetitive just from a hardware cost standpoint. The fact that they’re first to get higher level self driving working would become instantly irrelevant.

I think of it like a format war. The winner isn’t usually the one that is the most technologically advanced.

Elon is an idiot but he wasn’t wrong about the general idea of having a very low hardware cost. There’s almost no point in getting something working with exotic equipment.

> I think of it like a format war. The winner isn’t usually the one that is the most technologically advanced.

The stakes are a tad different. I can accept the slightly lower visual quality of VHS compared to Beta. For self driving, if the lower technological level results in a slightly higher chance of "you're dead", I'm not going to be cavalier about what I choose.

But that's not really the right analogy. A VHS tape didn't have an analogy to "you're dead." It worked fine, was just as reliable, but it happened to give you a longer recording time for the money. Any negative trade-off it had was basically negligible to the average person, which is why it won.

What I mean here is that if Tesla and other automakers are developing cheap-to-mass-produce autonomous driving that ends up being 5 years behind Waymo, their lack of first-mover status is moot. The moment that the cheaper technology it catches up in capability well enough to be reasonably safe and equivalent, Waymo is stuck with a more complex and costly solution.

Waymo will be making the hand-crafted mechanical watch while the automakers will be making cheap Quartz watches. Both tell time equally well.

> But that's not really the right analogy.

I know it's not the right analogy, that's why I used it as an example of what you proposed, i.e. "I think of it like a format war. The winner isn’t usually the one that is the most technologically advanced."

The point being, in normal "format wars", people are willing to take certain tradeoffs because the stakes are low - I'm willing to be cheap because the difference in video or audio quality or whatever is negligible to me.

That dynamic doesn't exist with self-driving cars. If one system is able to drive just .5% better, I'd pay a ton more for that because my life is on the line. And it's not at all like Waymo is building a "hand-crafted mechanical watch" - they're building a system with better sensors to provide a better experience. I sincerely doubt leaving out these sensors will ever lead to a safer experience.

> A Waymo driverless cab has a roof rack full of equipment that might cost up to five figures to produce and install.

"Full Self-Driving Capability: $12,000"

Source: tesla.com

FSD upgrade on a tesla is already low 4 figures...

It doesn't matter what other consumer vehicles or Waymo offers, or their financial state (which wasn't point of my comment). Tesla is the one who has continuously promised they'd be L5 and have robotaxis "by end of the year" for years now.

When questioned on their own promises, you can't turn around and say "It's a beta; no other car maker is further along".

Of course, it's a successful business already because many consumers believed Tesla's (false) promises. If that's your bar, then very well. But they still haven't met their original stated goal and it doesn't look like they're going to anytime soon.

> Tesla’s full self driving system is the only one on the consumer market that will drive on city streets for you in the form of beta software.

Why do people ignore comma.ai? They call it ADAS instead of FSD? Their marketing is confusing because they just provide hardware for MIT license OSS?

> that $5 or $10 matters.

That matters, but what almost surely matters more is providing a differentiation among the trim levels to justify the four-figure difference in prices.

> Tesla has been kept alive only because they are scared to even let drivers take their hands off the wheel. They have very low confidence in the system to allow for driver inattention, let alone completely remove the driver. Numerous software “rewrites” over the years with many buzzwords attached, and they still haven’t clocked a single driverless mile.

That is quite backwards.

When they first launched, the steering wheel checks were every fifteen minutes. They reduced it to five minutes, and slowly had to be dragged kicking and screaming due to threats from the NHTSA to get to where it is now. Even now, most of their attitude to it has been very nudge nudge wink wink, "yeah, they make us tell you to pay attention but you don't have to" (until you do get in a collision, and they throw you under the bus, of course).

> until you do get in a collision, and they throw you under the bus, of course

You made my point for me. It's less about steering wheel checks and more about taking liability. They can remove the driver and take liability only if they are confident in their system. They are not.

Not really. They don't care about the liability until they have to. That's why they didn't care about steering wheel checks either. Tesla were not scared to "even let drivers take their hands off the wheel". It was the NHTSA who said "No, actually, you need to be", and started forcing their hand to actually implementing the checks (or rather with a reasonable interval).

Reply on news.ycombinator.com