This is a gotcha headline that is contradicted by one of the quotes from Tesla they list in their own article:
Please note that Tesla’s development of true autonomous features (SAE Levels 3+) will follow our iterative process (development, validation, early release, etc.) and any such features will not be released to the general public until we have fully validated them and received any required regulatory permits or approvals.
TL;DR: Tesla does not want their upcoming release of FSD Beta to be considered to be above Level 2, and their lawyers are submitting documents to California saying that it shouldn't be considered above Level 2 and thus they shouldn't have to apply for special permits in California to release it. They continue to work on autonomous driving features.
He's such a good marketer. There is a case to be made that he's participating (often) in fraud to pump up the stock price, but people keep buying it so...whatever.
I don't think of it that way. It's investing $10,000 in a feature that is currently under development and will go up in price as it improves. It also includes any hardware upgrades that are necessary to keep up with FSD requirements.
Tesla fans know that FSD isn't ready. They know that the $10,000 upgrade doesn't actually grant FSD, but the purchase of it indicates that they believe it will in the future. You can call it a $10,000 gamble if you wish, and you can express your belief that it will never happen, but to say "it's a $10,000 option that doesn't work" is only a half-truth, IMO.
Musk’s reality distortion field must visibly bend light. He manages to accomplish so much, but also manages to peddle a lot of absolute monorail-level bullshit.
TSLAQ is a coordinated effort by TSLA short-sellers to generate as much bad press about Tesla as possible in order to tank the price and make money on their short positions.
Ensuring said bad press is accurate/fair is certainly not the primary goal.
> TSLAQ is a loose, international[1] collective of largely anonymous short-sellers,[2] skeptics, and researchers who openly criticize Tesla, Inc. and its CEO, Elon Musk.[3] The group primarily organizes on Twitter, often using the $TSLAQ cashtag,[4] and Reddit[5] to coordinate efforts and share news, opinions, and analysis about the company and its stock.
> Tesla Ticker Symbol + "Q" which is the NASDAQ notation for bankruptcy
First of all, Musk literally never said he would develop a Hyperloop. He literally said, here is an idea, I'm not gone do it, somebody else might want to try. Why don't people understand this?
Tesla had a battery swapping but realized it was a terrible idea and instead built the largest global chairing network in the world instead. Are you seriously gone shit on somebody because they changed strategy with more information and build something that is actually more impressive?
Yes, he was wrong about Robotaxi, or at least its delayed. Nobody bats 100%.
But if you are actually honest with yourself and start in 2000 and look what he said, compared to what he did. Literally nobody else on the planet predictions are closer to what actually happened in terms of large scale problems.
It's because of the former that he is able to command the latter.
Musk isn't super-human. He's a good engineer with great intuition for how to solve hard problems and a good "first principles" approach. He also has a great work ethic. There are tons of Elon Musks around.
What's rare is people as visionary and competent as him who are also rich and well-connected. Having someone who actually knows something in a position to make decisions about the allocation of large amounts of capitol yields results so amazing that they appear super-human, which buys him a hell of a reality distortion field and a lot of slack for being eccentric.
Musk got rich more or less by accident. The systems that advance people to those heights select for competence in the areas of social and financial advancement, not practical engineering or creativity, hence his rarity.
Of course today it's probably better than it ever has been in all of history. Historically the way to wealth and power was to brainwash a large number of people into killing (and risking their lives) for you and then sitting back and reaping the reward. That's still a thing, but at least now there are other routes.
Being able to steer the car effectively is only one part of an enormously complex engineering problem. A fully autonomous vehicle would need to:
1. Be robust to bad visibility, not an easy task in computer vision.
2. Have access to extremely detailed maps that include things like driveways and parking lots. These maps would need to be continuously updated. The labor required to make and maintain such maps will limit how many places cars can be fully autonomous.
3. Understand the etiquette of the road. For example, a person seeing a car stop in the street near a parking spot would assume the car is about to parallel park and give it room to pull off the maneuver. An autonomous vehicle wouldn't give it so much room, because it can't reason about context in that way.
4. Negotiate ambiguous situations requiring interaction with other drivers. There are frequently situations where it's not clear who has the right of way from the rules of the road. People resolve these by gesturing with their hands and with their cars. A fully autonomous vehicle would need to understand these signals.
Number 4 is just a deal breaker you can easier expect a car to reason about what other cars are doing, or trying to do, than respond to human gesticulation.
One might easier imagine cars communicating with one another over a local connection of some sort.
Personally I think he understands everything. The problem is the stock market and SEC which set up laws that encourage and incentivize bullshitting. Everything from letting a free market full of idiots to determine prices, to quarterly earnings reports, is set up to encourage every form of BS.
If we want to optimize for reducing bullshit, the stock market and rules around it need to be restructured.
Not really surprising if you've been following the beta progress. I doubt Tesla will move past level 2 until they are at a point where Level 4 or 5 is possible.
With that said, I have real doubts that they can get their with the current camera suite. IMO, they don't have enough redundancy. Further, I think elon's anti-lidar position is a mistake. Even if they can make it without lidar, I think it will set them behind the rest of the industry.
Likely, we will see something where they collect enough training data to hit level 4 + training suite and eventually they'll release cars with new sensor suites. Once that happens, my bet is that's when they'll actually release level 4 capabilities. I highly doubt any current car will ever allow hands off driving.
Something I’ve been curious about with computer-vision-only versus LIDAR: snow.
Now, I am naïve about the real intricacies of computer vision. But won’t snow be quite nasty for camera-only systems? It’s almost the physical definition of noise, as it’s random and reflective. Is there any research or experience with computer vision and dealing with a snowstorm? Because even human drivers aren’t great in that kind of low-vision scenario.
I think the argument is that if a human with two eyeballs can do it, a computer with a bunch of cameras should be able to as well. I can drive in the snow.
A human can drive in the snow, and has a sensing package which effectively consists of two cameras.
The hard part (when you only have two cameras) is that the processor needs to be exceedingly sophisticated, e.g. a human brain.
People like Waymo have decided the best way to decrease the brain requirement is to add more sophisticated sensing that humans are fundamentally incapable of.
Tesla's strategy is more to just build a better digital driving brain with machine learning + lots of miles, because arguably all sensing mechanisms have their shortcomings. That being said, they still have a lot more than two cameras in their FSD package – just not LIDAR. They still have radar and other things as well.
Only time will tell which approach is better. Probably in the end the best approach will be to combine the two: a brain that is sophisticated enough it can drive with just cameras, but you throw LIDAR and such in their too to go the extra mile.
> A human can drive in the snow, and has a sensing package which effectively consists of two cameras.
Two cameras, and a huge amount of mostly not understood wetware interpreting their signals. Our current approaches to ML may be a local maxima. The true path to AGI be somewhere as yet undiscovered. Lidar certainly seems like a good tool to use given our current ML capabilities.
it's a strange philosophy. Isn't the point of building machines that they can do things that humans can't? It's like building drones for warfare and turning the heat-seeking off because humans don't have that capacity.
A generally intelligent driving system is going to benefit from better sensory inputs just like a human would, it's not like that makes the system dumber. The whole dichotomy between sensory inputs and processing seems strange to begin with, there's not necessary a clear line, or a reason why one should be prioritized over the other.
We are not talking about generally intelligent driving systems though.
You have to realize that most SOTA self-driving systems are using a lot of NNs. NNs in turn want lots of training data. That training is not fully generic – in other words, if you train a NN using only cameras, giving LIDAR to that trained network (i.e. for inference) will not really do anything.
Having large LIDAR arrays on the top of vehicles is expensive and slow to deploy, which directly translates to being able to generate less training data.
So if your goal is to have the best possible self-driving vehicles 100 years from now, absolutely you should be using as much sophisticated sensing equipment you can get your hands on, and slowly train your nets for the next century. If your goal is to get out something that is better than a human driver as fast as possible, then it is not clear that requiring all vehicles to have LIDAR on them is the best way to go.
Not necessarily. Usually machines are built to do mundane tasks that humans have to do. In most cases they are built to replace what a human can do.
The argument for having a lidar vs not having a lidar is different. They chose to go without lidar because it should be cheaper and faster to get started, and the fact that people are capable of doing driving without lidar makes it seem possible that a machine could do it without lidar.
Saying that computer vision can view the world with snow present is hardly the same as claiming that computers can do literally anything a human can.
On the contrary, having computer vision capable of handling snow is a necessary precondition for self-driving, regardless of what other sensors are present.
> Just taking noisy camera data and turning it into a de-noised picture.
I think this is a huge simplification of the actual problem. Not only do you have to denoise the data, you have to use the data to generate models of 3D objects in space. Maybe you could argue that exact 3D models are not necessary, but IMO any neural network that can drive at Level 4/5 needs to approximate 3D space in some layer.
This is incredibly difficult; we have an entire region of the brain dedicated to for this exact purpose (the visual cortex). This region has the benefit of millions of years of evolution. We also have the advantage that we can predict 3D models based on everyday experience; we don't know how to encode that knowledge into a neural network efficiently. That's not even mentioning the latency requirements that driving in the real world would impose on such a network.
When it comes to things that don't require creation, I do believe that computers will be able to do anything humans can. The problem is simply that the software has yet to be written.
Keeping the context of a self-driving car, humans are perfectly capable of driving cars with just two eyes to determine surroundings. For 99.9% of your driving, you don't have to think about it. Stay between the lines. Stop for the red light. Make sure there's nobody next to you before changing lanes. There's no creativity here, just following a very concrete procedure.
Humans fail because humans do stupid human things. They text while driving and drift over the lines or don't see the red light. They don't always check before changing lanes. They drive drunk. Computers won't do any of these things.
And they'll do it better. Humans have only two eyes and can only look in one direction. A computer can check the blind spot while still looking ahead, so there's no chance it will rear-end someone because the person in front of them slammed their brakes while it was making sure the next lane was clear before making a lane change. A computer will notice a car in the blind spot veering into its lane without taking attention off the road in front.
Again though, it's all down to the software. We're not there yet. But that doesn't mean we can't be in the future.
Humans certainly do drive in the snow, but not well. I've driven without being able to see the road (very slowly). This is not a safe situation. This is actually a case where I think computers (with LIDAR or other advanced sensors) will be able to outperform humans and it won't be close. I agree with the posters that say the anti-LIDAR stance is not going to work out in the long run.
Musk is not only anti-LIDAR, he's anti-HD mapping. Imagine if you had a centimeter-accurate map overlaid on your windscreen. You'd probably be able to operate the vehicle right up to the limits of the conditions. It definitely won't be close in many situations.
More often than not, snow driving is CHALLENGING. And it's not for a lack of experience, as I'm doing multiple roadtrips through the BC Coquihalla Highway. More often than now, it's more of a matter of best attempt at driving in the path made by other car disregarding lines on the road than driving by the book. The utter HELL being driving at night with a blizzard, this is a level-up experience for sure.
1. Many people have no choice but to drive in bad snow conditions for any number of reasons including that they're emergency workers.
2.) Winter conditions catch you. Even though I don't commute and basically do everything to avoid bad winter weather conditions, I've still been caught in pretty bad conditions a couple times over the past few years.
Is that how much a pussy society we have become ? Avoid dangerous conditions because "it might be dangerous" ? If a freaking Tesla can't handle what my 06 SUV can handle without issue, and what people were handling ~forever, ... just as much as an Apple laptop can't handle what a then-IBM was handling, then the damn "car"/"laptop" is nothing but a piece of shit.
That's a ridiculous statement. Having a self driving car that works in only certain conditions can still be a benefit for the society. It won't benefit you, but it will benefit thousands of others.
They are not stuck, because they can drive the Tesla themselves and secondly they don't have to buy Tesla in the first place. It's still useful for places without snow.
How fast do humans drive in heavy snow? I think most people drive extremely slowly to allow themselves to filter out the snow noise by taking in more frames, this will be possible with AI too.
It's worth noting there are many accidents caused by people driving in snow and that AI could be hugely safer at some point.
And also having to drive manually when it is snowing is probably fine for v1.
I would worry more about bicycles, they can be extremely tricky if the rider doesn't understand how to explain their presence on the road to drivers.
I drive slower in the snow because of limited traction and the dangers of other drivers around me.
If I had some button that made spikes or another traction method pop out of my tires and gave me great grip I probably would drive full speed in the snow; because the main issue would have been solved.
Its a hard problem, but LIDAR has similar issues. Snow can cause noise for other sensors as well. In each case it seems like these issues are not insurmountable.
I don't expect that any self-driving systems will start out with handling all weather conditions. Even in the long term they may handle fewer conditions than humans.
I suspect it will be able to drive as well as the best human drivers in snow. Why would it be any different? We use our eyes which are basically cameras.
They need LIDAR or stereopsis. IMO stereopsis could probably do a better job than LIDAR much more redundantly, and for a tiny fraction of the cost.
But neither is insane to me. Structure-from-motion is all you could get with monocular vision, and I would not trust that as the primary means of depth estimation, especially with no redundancy.
Article is misleading. Tesla sent a letter to the CA DMV requesting approval for a feature which they referred to as "City Streets" which they say is level 2, and will continue to be level 2 as it is developed. They specifically say in this very document that they will develop higher levels of autonomy in the future, so it's pretty clear that the "final release of City Streets will continue to be level 2" quote is just referring to the particular subset of functionality that they're seeking approval for.
From the very same letter, there is also this quote:
Please note that Tesla’s development of true autonomous features (SAE Levels 3+) will follow our iterative process (development, validation, early release, etc.) and any such features will not be released to the general public until we have fully validated them and received any required regulatory permits or approvals.
"City Streets" is part of "Full Self Driving" $10,000 package though. They are calling it "Full Self Driving", but a subset of functionality will always be level 2? And the FSD Beta is also level 2 because it only works on city streets? That seems clearly misleading.
These are incremental releases of gradually improving functionality. Their plan is to get to level 3+.
People who buy it know what they are getting, and if they don't, it's entirely willful ignorance. I don't even own a Tesla yet I've read plenty about the actual capabilities and the fact that these cars can't truly drive themselves.
I don’t think we’ll see “full self driving” within the lifespan of several cars from now.
Yet, if I bought a Tesla I’d still need to sweat the decision of buying the FSD option because Musk insists the resale value of cars without it will be significantly smaller without it.
I always assumed by "full" they meant that it doesn't just do autopilot on highways, but is capable of handling local roads, traffic lights, stop signs, as well. Basically it can drive from local address A to local address B assuming good conditions.
I'm a robotics person though so it was always pretty clear to me that I couldn't possibly think of being able to deploy an L4 system with their set of sensors.
When you call it "full" and "self driving", I don't know how many qualifiers you can attach to it. I know it's not anywhere close to driving by itself, but between the naming and Elon Musk constantly talking about how level 5 (I mean, level 5 is the ultimate form of autonomous driving) is just around the corner, it's pretty easy to see that as misleading customers. All while they don't even bother to get a driverless testing permit from CA DMV.
The problem is (a) the stock market is full of idiots that won't listen to real engineers like us, and listen to some bullshitty CEOs and analysts instead (b) the SEC and stock market is perfectly optimized and incentivized to maximize the amount of marketing bullshit coming out of company PR.
If people wanted more honestly worded things out of companies, we need to stop nailing companies every time they don't "beat" earnings estimates every quarter, and begin to understand how real science and real technological advancement happens.
The way NASA is funded, for example, is the exact opposite, and minimizes bullshit, but NASA and Tesla share a lot in common in terms of what they're trying to accomplish for humanity.
EDIT: Okay maybe NASA wasn't a good example. But in any case, the current structure of public companies and letting a bunch of gullible <strike>sheep</strike> average citizens collectively determine the value of a company based on what a bunch of <strike>clowns</strike> non-engineers in suits at Wall Street try to make of a highly technical venture isn't a system optimized for minimizing bullshit, it's quite the opposite.
> The problem is (a) the stock market is full of idiots that won't listen to real engineers like us
So you are a real engineer and Musk is not?
> begin to understand how real science and real technological advancement happens
By pushing for it hard, investing your credibility and a lot of man power and resources into it? Like literally what Elon Musk does with all his projects.
> The way NASA is funded, for example, is the exact opposite, and minimizes bullshit
NASA claims nonsense literally all the time. The #JournyToMars under Obama for example.
Real engineers don't bullshit. Real engineers understand that there is a certain responsibility that comes with people trusting their lives to their engineering skills. Real engineers can give you confidences on everything in hard numbers with data. Real engineers can provide realistic estimates of timelines. Real engineers make mistakes sometimes, admit them candidly, and know perfectly well how to adjust those timelines based on past estimate errors and extrapolate adjusted project timeline estimates with basic math. Real engineers know the limits of systems and how to write spec sheets that describe them accurately. Real engineers don't always know the answer to every question, and know when to say "I don't know".
It's managers, investors, analysts, and sometimes customers who don't want to listen to those extrapolations and estimates and uncertainty. And so many real engineers gradually become increasingly fake, because the other (corrupt) parts of the system spread their toxic thinking into the engineering.
I have a lot of respect for Musk, who was a real engineer, probably REALLY wants to stay a real engineer, and is still mostly an engineer. But he is indeed forced to be a part-time <strike>bullshitter</strike> non-engineer by the system, which demands bullshit by its very incentive structure.
For Musk his bullshitting is still nowhere close to the competitors, who said for a long time that they can manufacture millions of electric cars if they really want to.
When I told my friend about self driving Tesla who's a BMW fan, he said that BMW will just sell self driving cars when it's ready.
Really? Without gathering any data? I'd say BMW is the bigger bullshitter. Elon needs to update the computers maybe, but everything else in the cars is ready for self driving im the near future.
> his bullshitting is still nowhere close to the competitors
I mean, as I said, the whole stock market is full of bullshit. My point is the system is designed to incentivize bullshitting, so it's expected given the incentive system created. Companies that don't bullshit don't win, and if that's the loss function you define, bullshit is what you're going to get.
> everything else in the cars is ready for self driving
I disagree with this. I think significant hardware upgrades to Tesla cars will be necessary to achieve L4. Any attempts at L4 without those upgrades will result in a lot of accidents. I won't be surprised if they come out 5 years from now asking former Tesla owners to get their cars retrofitted with some new sensors if they want L4.
L2 is achievable with their existing hardware and that's what they are doing.
They already evaluated retrofiring sensors. Doing it would cost about as much as a new car. They've made a bet that this sensor suite will be enough and I'm sure Musk got green light on it from his team. Time will tell if they were right.
I personally agree with them that technologies like PseudoLIDAR will be sufficient.
I believe that for a good human driver Tesla sensor suite would be sufficient to drive safely if you covered all windows. You could project the camera views on covered windows and keep the current bird eye visualization.
The issue really is with control of the car. Deep Reinforcement learning sucks. Training Deep RL is unstable, hyper-parameter sensitive, even random-seed sensitive, and requires ridiculous amount of data for even the simplest applications. And after training it does not generalize.
Yes. I know not to listen to CEOs when they say things like that. They're effectively puppets.
I've also probably coded way more lines of autonomous driving code than Elon Musk himself. That isn't to say I don't respect him and his vision, but I'm pretty certain I have a better idea of timelines than the BS he's forced to say to the public.
In my opinion Musk actually believes in his timelines. He does exactly the same mistakes without financial incentives e.g. Falcon Heavy. His "COVID panic is dumb" and "it's gone by April" are memes by now. Then he will also openly say that he believes shares are too high, so of course shares dropped. Not to mention that SEC is going after him and Tesla for his stupid tweets.
Either he is an alien playing 5 dimensional chess with time-travel, or he is pathologically overoptimistic. He is especially overoptimistic about AI, so it leads him to make stupid predictions about when we will have cars without steering wheels.
> They are calling it "Full Self Driving", but a subset of functionality will always be level 2?
Well, yes? That's how subsets work.
Thought experiment, pretend "City Streets" was called "Level 2 Mode". Would you be upset that "Level 2 Mode" will always be level 2? Or that "Level 2 Mode" is a subset of "Full Self Driving"?
> Thought experiment, pretend "City Streets" was called "Level 2 Mode".
That would be an accurate name for "city streets". Except currently, it's part of an umbrella feature called "full self driving" that's supposedly going to be "level 5" soon. This is it with most things with FSD. Qualifiers after qualifiers for a clearly misleading name.
If a subset of features that is part of "full self driving" cannot ever reach level 4/5, then "full self driving" is not possible.
"full" is a marketing terminology but the English meaning of it would be that it works comprehensively in the majority of cases. If it doesn't work on city streets, which is a major subset of use cases, how can it be "full"?
There is an enormous difference between "city streets" the location, and "City Streets" the name.
They're not saying the car will be stuck at level 2 on streets in cites.
They're saying that this particular feature which is named "City Streets" will always be level 2.
Eventually (they promise) the car will have level 4 driving on city streets. Level 4 driving inside cities will either have a different name or no name at all. At that point they will probably delete "City Streets", because it's a subset of level 4 and there's no reason to keep it after level 4 is finished.
Edit: especially because the full name is "Autosteer on City Streets". When they have level 4 support, they definitely won't be calling it "autosteer"
if people watched the fsd videos on youtube, they would understand that fsd beta is full self driving, but it requires driver attention and gives the user a quick way to take over control.
I think the only one that doesn't get this are journalists, regulators, and class action lawyers.
That is where there is a disconnect between the Tesla marketing language and technical language.
They publicly advertise and collect cash for pre-sales of "Full Self Driving". Most Hacker News technologists (a generalization for sure) would reserve a phrase that bold for Levels 4 and 5.
But here they are in writing stating that they are currently only at Level 2.
So a couple core questions around Tesla are:
- Is it acceptable to take pre-orders from the average consumer for a feature that might be 10 years away?
- Is it ethical to use exaggerated marketing speak for specific technical functions (to give a non-Tesla example, this is like when I see companies implementing rule-based decision trees or linear regression and calling their apps "Artificial Intelligence")?
It's not just technical and marketing, it's legal too. And this article references emails from tesla legal team. So legal is telling dmv that it's L2 so it doesn't have to get permits, but that doesn't mean the tech isn't destined to be L3 when it's fully developed and they are ready to submit themselves to regulatory scrutiny.
Your theory here is they're purposefully downplaying their level of self-driving technology to avoid regulatory processes?
Are they going to keep winking at customers and letting them think it's level 3 so they continue to take their hands off the steering wheel and do other things like countless Youtube videos already capture?
It's a little disturbing how people will jump through so mental hoops to give Tesla the benefit of the doubt.
I don't really see that as giving Tesla the benefit of the doubt. I read it as describing what may be a ploy to avoid the regulatory processes and that may be a valid point.
> this is like when I see companies implementing rule-based decision trees or linear regression and calling their apps "Artificial Intelligence"
AI means so many things that it has become as meaningful as "magic". Some don't consider anything less than human-level an AI (this is also called AGI or strong AI). On the other hand enemy behavior in video games is commonly called "AI" even when it is as simple as pacman ghosts. In the same way that magic can literally mean "impossible" or be applied to card tricks accessible to 5 year olds.
"AI" is ambiguous but "Full Self Driving" has a pretty clear plain meaning. It means that the car is self-driving. Not partially self-driving, not mostly self-driving, fully self-driving.
And I think that's the meaning car companies should use in public communications. I get why it's helpful for automotive engineers to talk in more detailed shades of grey, but for the rest of us, a car where I can't prop up my feet and start playing Minecraft is "assisted driving".
This disconnect seems to be growing globally, be it marketing, technical or legal.
It seems strange considering the wide availability of information and communication, but apparently there either is no feedback loop back to the people who write like that, or there is but it's mostly positive (be it more sales, less lawsuits or a greater following). At some point the meaning of those words will have hollowed out so much that we can't use them in normal conversation anymore.
> Tesla sent a letter to the CA DMV requesting approval
Can you cite that? It directly contradicts the linked article which claims that the CA DMV contacted Tesla to clarify if Tesla were testing a vehicle without proper permit.
If Tesla were seeking some kind of approval it makes their email very confusing as Tesla is insisting that they don't need additional permits/approval for the "City Streets" beta.
That seems to confirm what the article said. The first email appears to be dated December 20, 2019 from the CA DMV to Tesla.
Here's the original email sent to Tesla:
> Subject: Announcement of Full Self Driving Feature.
> Al we’ve seen press reports that Mr. Musk announced on Twitter that a “Tesla holiday software update has FSD sneak preview...” Many people generally translate “FSD” to be “full self-driving.” Is Tesla releasing a full self-driving software update to any California Tesla owners? As you are aware, the deployment of autonomously driven vehicles on public roads in California requires a permit to deploy. At this time Tesla does not have a permit to deploy. Please provide an update on what this announcement for the deployment of the feature means in terms of autonomous operation on public roads in California?
I don't see anywhere where Tesla are seeking approval, quite the opposite in fact. Tesla are saying they shouldn't be required to seek approval for "City Streets."
Oh I see what you're saying. I believe they are "seeking approval" for City Streets ... as a Level 2 product. Likely some discrepancy between what Musk says and what the rest of the company does.
They were already approved for Level 2 driving, the DMV is asking if City Streets is level 2 or per Tesla's own statements FSD which they were not permitted for testing.
If "Tesla sent a letter to the CA DMV requesting approval" per the above comment, that would have meant Tesla were trying to get an FSD permit in CA (the opposite of what happened).
Yeah, they just keep moving the carrot. By the time Tesla comes with FSD the current model lines will have reached their end of live and the $6k+ that people paid will be lost with the car.
An attorney general from a state will start to go after them if they don't come up FSD any time soon. My guess is the settlement will be that they can carry it forward to another Tesla. No of the features really work with the current FSD like smart summon. And there's probably tens of millions of money poured into a feature that doesn't limit up to expectations.
If they were performing as their CEO represented, the cars would be have driving themselves cross-country to delivery themselves to customers years ago. The time to have them up about lying what would be delivered and when was a looooong time ago.
Cynical take: Tesla GC says "We completely understand and agree that we won’t deploy any autonomous vehicle feature without a deployment permit" [ https://www.plainsite.org/documents/242a2g/california-dmv-te... ]. Absolutely certain that Tesla won't take liability for city driving unless they absolutely must. Especially when so many Tesla drivers are so open to accepting the liability. If Tesla's competitors (e.g. Waymo etc) launch taxis, then Tesla might change course and do the same. But if Tesla doesn't have to do that, they're going to play a game with California where they get to argue over if their cars are level 2 or actually self-driving or whatever .... all the while Tesla users continue to assume liability and help train Tesla's software.... which is what Tesla really wants to continue as long as possible.
These docs don't really shed light on Tesla's FSD plans so much as they illustrate the lengths to which Tesla will go to skirt regulation and liability. All the chatter and rumors about Tesla FSD just drive more attention and more potential users (trainers) into their pool.
If you think it trough, Tesla is misleading in making anyone velieve their curren tech will ever have full self-driving capabilities.
What's really required for full self-drivung capabilities is infrastructure support. This will come as large eatablished car manufacturers and governments start defining standards for that.
Here's a very simple example. Imagine a stop sign. Tesla's and Google's aproach is to detect the stoo sign via cameras. This can be problematic when the canera can't see the stoo sign (weather, tree branches blocking it partly, a truck in front of you). It would be much simpler if the stoo signbwas able to communicate to self driving cars that they have to stop. Now imagine an accident because your Tesla didn't recognize the stoo sign. Who is going to pay for the damages? Tesla for false advertising? The city for not cutting down the three? Your insurance? You?
As cynical as I'd like to be towards Musk, I'm sending this comment over a Starlink connection that has made me rethink what my future could be. Living in a rural area, I was always under the impression that I'd need to move to the city and get a fiber connection eventually. With Starlink, that's no longer a concern. My latency is rock-solid at <50ms and download speeds are pretty capable as well.
Obviously Elon isn't entirely responsible for it, but I'm constantly impressed by his initiative in tech world. If Starlink weren't available, I'd be waiting ~3 years for Amazon to build some sort of equivalent service.
Mine gets between 50-150mbps, sometimes 200-300 down. Up is between 20-40 from the tests I have done. Exactly as promised. My ping has gone from 40-60ms, and now is regularly around 20. This was just in a matter of months, and also me fucking with it for hours to try to find the best placement.
I still have the occasional drop out... I can't make it through a video conference, or wifi call without a freeze up or audio drop. My Starlink app says "No obstructions" but also says "Obstructed for 4 minutes" on a regular basis. Many people seem to echo this. I'm not honestly sure if this is what to expect or not.
I generally switch to my 1.5Mbps DSL line for phone calls and video conferences, which lowers the quality a ton, but at least I don't have to ask people to repeat the last 30 seconds. Web surfing and downloading large files is blazing fast, and multiple people sharing the link seems to have no impact on the quality. Video streaming is fantastic. Another downside is that you're behind a CGNAT, so you'll never get a public IP address. Then their router puts you behind another NAT...
StarLink is better than nothing for sure. But if Comcast didn't want $25,000 to run 550 feet of cable to my house, I would probably use that instead. I now have to call in an arborist or someone and top some trees that I am guessing are causing temporary obstructions, but I'm not really sure. I have a lot of trees on my property.
Curious if what you're seeing for data speeds matches what they say on their FAQ page. Your latency is just past what they are promising, which I suppose is not terrible.
"users can expect to see data speeds vary from 50Mb/s to 150Mb/s and latency from 20ms to 40ms"
They probably have a disclaimer somewhere in the contract saying the guarantee doesn't include latency added by the user's computer or network interface.
I get 50Mb/s on a bad day, and latency tends to be 40-60ms on my ethernet cable. I've found it to be just shy of what they promise, but certainly not far off.
Not to distract from original post, but! What download speeds are you getting? Do you suffer from throttling, for example of more people are connecting at the same time? Are there high and lows during the day?
Our invitation arrived and we are tempted, but it seems very expensive.
If you're getting less than 1mbps, I'd wholeheartedly recommend it. Download speeds are between 8 and 16mbps, and I haven't noticed any throttling with 6 people simultaneously connected. Latency is pretty mediocre right now, but it's a massive improvement over traditional satellite networks. I can play online games without issue, and my SSH connections are snappy and painless.
My only concern has been the stability of the connection, there are still intermittant disconnects happening on a daily basis, and it comes out to <3 minutes of downtime per day. If that's a dealbreaker, it might be best to hold off until they get it sorted out. Besides that, it's a great deal. The dish is dead simple to install, and even the router they provide is pretty damn good (sadly it only provides a single RJ45, so keep that in mind)
That's actually super unnerving. I've never seen something like that, it almost looks like it's automated in some way. Thank you for bringing this to my attention.
Having a strong reality distortion field is a necessary, but not sufficient condition to be a great innovator. You have to have some crazy ideas (landing a rocket!) and be able to convince people to invest with you or work for you. But no one bats 1000, and Musk's obsession with "autonomous" things, whether evil AI overlords or self-driving cars, tends to be his weak spot.
I wouldn't trade one CEO working on actual important problems, like automobile deaths, global warming, space travel and universal internet access for a hundred CEOs working on the next social media startup or crypto idea. No matter how many stupid jokes he makes about "$420.69" on Twitter or whatever, he's had so much more quantifiable impact on the world.
I think the "$420.69" jokes are a sort of biomarker.
Sort of like self-deprecating humor might not be the entire picture of mental health, but a sign that it is probably ok.
EDIT: I think if Elon Musk gets too serious, or gets shut down, I suspect some ingredient of entrepreneurial spirit might be lost. The "do the right thing poorly" vs "do the wrong thing correctly"
Watching the stark difference between the rover on mars and Musks starship experiments I have little faith that starship will ever be realized. Elon over promises and under delivers almost every time.
* Achieved dominance in the commercial launch market, arguably with an even more substantial technology lead.
* Performed the first landing and reuse of an orbital rocket booster.
* Developed a heavy-lift version of that rocket that has the largest payload to LEO out of any currently operational launch vehicle. Also with reusable boosters.
* Delivered a new human spaceflight system to NASA, the first such system since the Shuttle ~4 decades ago despite decades of promises and plans to the contrary. Starliner, the incumbents' entry for the commercial crew program, has not yet flown and is not expected to until later this year at the earliest.
* Developed the first full flow staged combustion cycle engine to leave the ground under its own power.
It seems really questionable to bet against them at this point, and minimizing their accomplishments to date is really, really silly. Not sure why you have to involve the rover in your comparison, though—it's also a monumental achievement, but it's barely comparable to the domains SpaceX is working in with their launch vehicles.
I'm dubious that Starlink will really be all that great once it really scales up to tons of users. It's not surprising that it's great for a limited beta run.
Driving on a highway is an amazing experience when it's mostly empty, and a nightmare in rush hour traffic.
Has there been any good write up about how well Starlink (or any other satellite internet) could feasibly scale? Even if SpaceX can start dumping tons of satellites for dirt cheap with Starship's economy of scale, surely there's a limit where the risk of a Kessler syndrome event makes further expansion impractical.
Space is really big. As a simple thought experiment, let's say each satellite is the size of a car (they are actually smaller but let's pretend).
There are currently about 1.5 billion cars on the planet.
They are confined to land which only makes up 29% of the earths surface.
If evenly distributed on that 29% (~57m sq miles) of earths surface (196m sq miles) there are about 25 cars per square mile.
The orbital shell in which SpaceX is flying satellites starts 550 miles up. If you confine yourself to a 2D spherical plane, this shell already has 30% more surface area than the earth, and you can use all of it, not just the 29% that is over land. If we "packed" orbit at the same density as cars across that 2D spherical plane we could fit 6.3 Billion satellites with a separation of 1000 feet minimum between any two satellites.
Of course we aren't limited to a 2D plane in space. You can fit millions of satellites in orbit with miles of separation between any 2 without issue. The bigger problem would be signal interference.
I don't think so, and I think level 3 is an error and should never be there. I think most will go from 2->4. L3 is just requiring human concentration that is impossible.
That being said I agree with the spirit of your comment.
Level 3 requires that a human be available to take control in a certain amount of time.
Whether that's a mistake depends on how long that amount of time it.
For example, 10 seconds is perfectly safe. It's fine to completely ignore the road until you hear the beep, because 10 seconds is plenty of time to look around and figure out the situation and take control.
I would argue that any system that requires constant human concentration is not actually level 3.
Level 2 is the one that's dangerous by design.
Edit: Though since the main difference between level 3 and level 4 is the ability to pull over, I wouldn't be surprised if level 3 gets quickly left behind or skipped entirely by many developers.
I'm pretty sure I'm not going to get that involved in a Netflix movie in a moving car. Is that a common thing?
But even in your scenario, if you hit vague awareness of the environment at 4 seconds, and you know you're taking over to drive... that 6 more seconds is more than enough to reach full environmental awareness.
> You were in a deep sleep?
You're not supposed to sleep for level 3. And "don't sleep" is not some ridiculous requirement. It doesn't even compare to the "be constantly vigilant" rule of level 2.
Full self-driving to me means that I can act as if there's an Uber driver at the controls. As far as I'm concerned, it's either that or I'm paying full attention even if I'm not fully controlling the vehicle.
It's one thing or the other. You're being fairly vigilant and looking ahead and ready to take control or you're a passenger.
I wasn't saying level 3 was "full self driving". Nobody had mentioned FSD in this comment thread. I'm just saying that level 3 isn't inherently dangerous.
> Full self-driving to me means that I can act as if there's an Uber driver at the controls. As far as I'm concerned, it's either that or I'm paying full attention even if I'm not fully controlling the vehicle.
> It's one thing or the other. You're being fairly vigilant and looking ahead and ready to take control or you're a passenger.
It's the latter option.
In level 3 there's an Uber driver but they might have you take over at some point. You're not in control at all, a passenger until the car tells you otherwise. You don't have to have any vigilance or looking ahead, you just have to be awake.
at 65 miles per hour 10 seconds is 950 feet. The appropriate space between vehicles to give yourself time to react to something ahead of you is 3 - 6 seconds worth of time.
If you need 10 seconds to react whatever was going to happen has already happened anyway.
It seems like Musk has decided that if humans can drive visually, then cars can. Why did he choose this line?
That is, if humans only need two eyes to drive, then why not just only use two cameras? Isn't using more than two cameras (and radar) admitting that cars need more "help" than humans? If using more than two cameras is okay, why not go further and use LIDAR or HD-mappings?
I believe that Musk chose this line because he knows that LIDAR (in its current state) is too expensive to sell to consumers. Same with HD-mapping. We know that LIDAR based systems have achieved "better" self-driving than camera based systems.
It's still possible that Musk is correct. But at present, it seems to me that Musk's position therefore seems like a mostly ideological stance, and not based on any real data. That's why people doubt Tesla's ability to execute on Level 4/5 self driving. The market will punish ideological purists when they are wrong.
Nobody has demonstrated L 4/5, so all positions are ideological ones.
You could also argue that because LIDAR is expensive, it's hard to get billions of miles of training from LIDAR-equipped vehicles. This is not the case with Tesla's sensing package, which is racking up the miles as we speak. So sure, expense makes it less appealing from a consumer perspective, but its also possible that the difficulty of deploying LIDAR fleets makes it less likely to achieve FSD in the short term due to less capacity to generate training data.
Long-term we'll probably want all our vehicles to have as many sensors as is feasible. But nobody trying to get to FSD today is worried about what it will look like in 100 years – they're all trying to get their first.
> Nobody has demonstrated L 4/5, so all positions are ideological ones.
Er, Waymo is doing a commercial L4 service. Cruise is doing pretty good. So no, they are not all "ideological positions". Also, no one except Tesla is targeting L5 which should tell you what the realistic goal is.
> You could also argue that because LIDAR is expensive, it's hard to get billions of miles of training from LIDAR-equipped vehicles.
LIDAR is expensive was a good argument in 2015. The prices are plummeting. Waymo reduced their top of line LIDAR cost from $75,000 to under $7500 and this was in 2017. They sell their short range LIDAR (Honeycomb) for under $1000. And their 5th gen hardware costs even less now (my guess is under $5000). Which way do you think the price will go as a number of companies mass produce it?
The fact is Elon made a bad bet a few years ago and didn't anticipate how much the cost would drop. Now Tesla is stuck with his ideological stance trying make a pipe dream work.
From what I've seen of Tesla's FSD beta, it could very probably be T4 if Tesla geofenced certain areas and worked out any particular kinks manually. T4 with geofencing is a far cry from T4 on any road anywhere.
That last sentence seems like a strange one. "Stuck" would be if you massively invested in LIDAR and it didn't help. If no-LIDAR FSD is not as feasible as Tesla hopes, they can just... use LIDAR. And they'll arguably be in a better position tahn early adopters because they weren't investing when the hardware was expensive.
Waymo is L4. L5 means it can take you to your favorite campground (anywhere) and handle any scenario. Waymo says it’s not a realistic goal. But yes, the normal levels are not very well defined.
> If no-LIDAR FSD is not as feasible as Tesla hopes, they can just... use LIDAR.
Tesla has promised the current fleet has all the necessary hardware to make it work (they couldn’t possibly know this at the time). So they can’t walk back on that without incurring huge costs for retrofitting existing cars and development cost for overhauling their tech stack.
Meanwhile, they would’ve sunk money and time pursuing a much harder problem. If they had used additional sensors from the beginning, they could’ve provided much more value to their paying customers by now.
Edit: Just saw your edit.
> From what I've seen of Tesla's FSD beta, it could very probably be T4 if Tesla geofenced certain areas and worked out any particular kinks manually.
Yes, that’s what they should have done. Make realistic promises. I mean, you’re talking about making it work on _any_ road, _any_ intersection and _any_ scenario. Without additional sensors and mapping, it’s seems pretty unlikely (if not impossible) to me. Certainly given the timeframe Elon Musk promises.
Part of the question is how useful the data sent back by random customer-owned vehicles is.
It's your own fleet that's collecting the really good detailed data. And adding LIDAR to your own fleet isn't tremendously expensive. You don't have to worry about an extra 10k per vehicle crippling your sales.
The reason they have less than a thousand cars isn't because LIDAR is expensive. They've spent billions of dollars and enough LIDAR units to double their fleet size would only cost 3-4 million.
Level 4+ using only cameras is risky and aspirational.
In truth, most of Elon Musk's high level goals are risky and aspirational. So far his record is pretty good, but eventually one or more are likely to fail.
The burden of proof for that statement is on Tesla, and thus far they have failed to deliver. 100% of vendors above 10k miles driven between disengagements have included LIDAR in their systems.
The fact is Tesla is not submitting test data to the California DMV, which means they are either not developing a self driving system and have no right to speak to what technology is or isn't mandatory - or they are doing illegal testing, which speaks volumes about the trust you should put in their opinion.
"You don't need LIDAR" is an extraordinary claim and counter to accepted best practice in the industry. The burden of proof is not on the rest of the world to disprove Tesla, but on Tesla to provide evidence that supports it. That is how science works.
Because at this moment, the only thing we know can be as good as a human is a human.
That means the only thing we know can work is eyes + brain.
The answer to your question is: we don't know what's needed, so people saying "sorry you need LIDAR for FSD" is much more "drank the coolaid" than my stance of "we don't know, maybe no LIDAR + lots of training data can work just fine". Or, turn the question on its head – how much better at driving do you think humans would be if we had LIDAR on top of our heads? My guess is... not very. Hence my belief that the Tesla approach (less SOTA sensing, more training data) is a reasonable one.
Your stance of "we don't know, maybe no LIDAR + lots of training data can work just fine" implies that a computer functions as well as a human brain, which as several other commenters have pointed out, is simply not true. I don't have a comment for you about LiDAR on top of humans heads, but I would bet that humans with one eye (and hence no depth perception) are worse at driving than humans with two eyes. There are also countless other sensors in your body that determine depth perception besides just your eyes - your entire nervous system is at play.
Since integrated LIDAR in humans is a bit hard to imagine, how about a human with 10 eyes? Will they be a 5x better driver?
You've still got things a bit mixed up. My stance does not imply that a computer functions as well as a human brain. I am not claiming that better sensors won't (eventually) give you better results. My stance merely implies that the brain (digital or wet) is a very important part of the equation, and just throwing sensors at the problem won't fix the fundamental need to have a highly sophisticated ML system running the show, and it's easier to get the training data you need for such a system if you're not using LIDAR.
Tesla is clearly gathering more data for their self-driving program than Waymo, due almost entirely to the fact that the Tesla sensing package does not require installing LIDAR on every car.
you are clearly out of your depth. Waymo gathers way more data than Tesla, mainly because they can. Each Waymo car is free to transfer Terra bytes of data through their hard drive after each shift.
Your average tesla car can only upload a couple hundred MB, if even that.
What about sound ? I drive with my eyes, but also with my ears. Do we know if any of those systems (Tesla, Waymo, Cruise, etc.) also use the sound for driving ?
While it is not sufficient on its own, it would yet another (small) piece of the puzzle.
We all have reasons to be cynical towards Elon, but what yo don’t realize is I'm sending this comment over a Starlink connection that has made me rethink what my future could be. Living in a rural area, I was always under the impression that I'd need to move to the city and get a fiber connection eventually. With Starlink, that's not a concern any longer. My latency is rock-solid at <50ms and download speeds are more than capable as well.
Obviously Elon isn't entirely responsible for it, but I'm constantly impressed by his initiative in tech world. If Starlink weren't available, I'd be waiting 3+ years for Apple to build some sort of similar service.
The incredible thing about the whole "FSD Beta" phenomenon is how many members of the public are willfully participating in the pump-and-dump. I just can't figure out if they are hypnotized, brainwashed, or just never had any critical thinking skills to begin with.
Take for example this well-known youtuber, AI DRIVR. This guy drives around in his Tesla talking about how amazing and mind-blowing it is, but if you mute him and just watch the video with a critical eye you can see that the software simply does not work. It should be outlawed immediately before it kills someone. He recently rampaged through my town of Berkeley, California, breaking five laws per minute (aside from the fact that he has the speed controls set 5 MPH above the limit at all times).
His explanation for the red light is "the front wheels were in the intersection before it turned red". I'm pretty sure it could have stopped on the amber, as required.
I have a 2021 model 3 and autopilot is a really bad driver. Teenager learning to drive with poor habits bad.
I went for a 500km+ drive on Sunday to visit new supercharger locations (it's been a long pandemic and I'm pretty bored). It was mostly rural highways without lane separation or limited access. These are some of the most dangerous roads to drive on since there is direct oncoming and cross traffic. Speed limits are 80 km/hr, but the prevailing traffic flows at just under 100 km/hr, so closing speeds are around 200 km/hr.
I hardly ever turn on autopilot since it's just bad at driving, but there were long stretches of completely straight roads with multiple minutes passing between oncoming vehicles. Even in that specific circumstances Autopilot made some truly weak decisions about its basic driving style.
It attached to the centreline between lanes which limited the gap with oncoming traffic. When I drive myself in the same circumstances I attach to the outside of the lane to maximize the space I can give oncoming cars. With the amount of room autopilot was giving even a small twitch or moment of distraction would be enough for disaster. If the event was timed poorly enough even a computer system would have no chance to react in time to physically manoeuvre the car out of the way.
Autopilot makes no accommodation for blind spots. When traffic is flowing at full speed on limited access or other multiple lane highways I adjust my relative position with other vehicles to ensure I'm never spending long periods of time in another driver's blind spot. If we're travelling the same speed I create a gap to ensure a quick lane change won't result in contact. Autopilot will continue along completely oblivious and no taking good defensive driving precautions.
There is a youtube channel featuring collision footage from Tesla onboard cameras. Nearly every collision when autopilot is engaged is a situation that never would have occurred in the hands of a skilled driver who anticipates issues and prevents dangerous situations from every occurring. Without knowing the specific collision that's about to happen in the video in most cases I can see a dangerous situation starting to develop I would have reacted to seconds before the collision actually occurs.
> Without knowing the specific collision that's about to happen in the video in most cases I can see a dangerous situation starting to develop I would have reacted to seconds before the collision actually occurs.
But you tend to know a collision is coming. Most accidents are due to people being comfortable and not being as alert as they should be; your example is the opposite.
AP is currently ~good on highway (probably safer than I am in good weather), but Meh elsewise. Hopefully it continues to improve; but it will be awhile still i think; and iffy to ever get there on our current Teslas.
fair enough on the anticipation, but when I have let autopilot take over it has often been only seconds before I'm already uncomfortable with the decisions it's making and spotting the really rookie driving mistakes it makes.
autopilot creates danger where none needed to exist by not using the full width of the lane and failing to create space through relative small speed adjustments. The cues it gives off while driving are some of the exact things I watch for when trying to identify beginning drivers so I can give them even more space.
I certainly would not describe autopilot as 'good' on highways. I'd barely call it adequate.
That is my conclusion as well. I have tried Autopilot and it's basically as good and useful as any modern lane steering and cruise control system out there.
You just wrote a long text about why it's important to use self-supervised learning to create a birds eyed view of the world, because object detection and planning from camera only view
doesn't work well enough. Yes, we knew it.
No, I don’t have FSD or the beta. I really do love my model 3, but charging for FSD that doesn’t yet exist is the closest thing to fraud I’ve ever seen so widely available.
The issues I have with AP have little to do with their object learning or need to gather more information (though I agree camera-only isn’t enough). The car knows where the lane boundaries are but doesn’t take advantage of the width. It knows accurately enough where other vehicles are that it could avoid staying in another driver’s blind spot indefinitely. Tesla is teaching their cars how to drive with terrible habits and not building defensive driving into their rule set.
The problem is not camera only autopilot. Your current software uses a projected view that can be seen from the camera for path planning. Dave made great interview about it on youtube.
If you are in the right region, you can try the FSD beta soon enough, which has showm amazing driving skills already from the Youtube videos I have seen.
I live in the EU though, so I guess I have to wait at least a year or 2 to try the beta :(
I did not pay for FSD, so I won't be getting any of that. My issues with autopilot are about the fact it fundamentally doesn't practice defensive driving. This has nothing to do with how tesla choses to input information as they already have more than enough detail to drive more defensively. They are fundamentally creating a very bad driver who cannot safely interact with other traffic and actively creates more risk on the road.
I see, if that’s true though, their safety statistics will not improve, which they can’t afford, so I’m sure it will improve over time, it’s just probably not yet ready for your safety profile.
We've had a model 3 for 2 years, autopilot has gotten worse over time. It was actually decent when we first had our car, but with every update it's slowly gotten worse and worse. I think it's intentional to either get people to pay for the full self driving package or to keep people from trusting autopilot too much.
Glad I'm not the only one who thinks this! I used to use it in my M3 all the time a few years ago. Doesn't make sense because the sign detection etc has improved...
I have [said this before](https://news.ycombinator.com/item?id=19446043 ); Tesla is not on track to deliver Level 5 with their current hardware and apparent software strategy. Piloting a car is a very hard problem, but I would say that Tesla is still struggling with the easy parts of it— they should have solved the "never hit any obstacle" problem to superhuman levels; I think that's pretty comfortably solvable with present technology and algorithms. Instead a Tesla feels about as good in that respect as a careful student driver.
The harder parts of machine-piloting involve interpreting the world (lane-finding in the absence of striping; distinguishing hard obstacles from ephemeral objects like paper, leaves, weather, reflections, and so on); and participating in a social environment (what driving postures indicate what intents? What do the local humans consider rude or polite? What are they expecting?)
(A truly advanced autopilot would seem almost magical to a human— it would be able to back-solve diffuse and specular reflections in the environment, and accurately deduce the presence of oncoming cars around corners that would be totally invisible to a human driver. It could predict accidents several seconds before they begin to play out, and the sudden braking/maneuvering in the absence of any visible threat would seem strange and confusing to the humans inside).
Tesla is nowhere near cracking these things; they're still on the ground floor. I still predict that current pre-buyers of so-called "full self-driving" are not going to get what they were promised, and going back on that will probably be bad both for Tesla's reputation and their wallet. If I had to guess, I'd say that if Tesla delivers on Level 5, it will be with a hardware platform that doesn't exist yet, and a software platform that is close to "complete re-write".
As it stands, their misleading marketing on self-driving capability is both dangerous and borderline fraudulent.
All that being said, I'm long on self-driving as a technology; just short on Tesla's current approach to it. It's a shame because I suspect both Waymo and Cruise are doing significantly better technologically, but I trust in their ability to execute a product significantly less. I'm not sure who will actually crack it.
Regardless of debates over the utility of "levels", it does seem misleading to name something "full self driving" that is not intended to fully drive itself.
193 comments
[ 1.9 ms ] story [ 209 ms ] threadPlease note that Tesla’s development of true autonomous features (SAE Levels 3+) will follow our iterative process (development, validation, early release, etc.) and any such features will not be released to the general public until we have fully validated them and received any required regulatory permits or approvals.
TL;DR: Tesla does not want their upcoming release of FSD Beta to be considered to be above Level 2, and their lawyers are submitting documents to California saying that it shouldn't be considered above Level 2 and thus they shouldn't have to apply for special permits in California to release it. They continue to work on autonomous driving features.
Tesla fans know that FSD isn't ready. They know that the $10,000 upgrade doesn't actually grant FSD, but the purchase of it indicates that they believe it will in the future. You can call it a $10,000 gamble if you wish, and you can express your belief that it will never happen, but to say "it's a $10,000 option that doesn't work" is only a half-truth, IMO.
https://www.roadandtrack.com/news/a35350331/checking-in-on-a...
(Still hoping for that super-tunnel to take me to O’Hare once flying is back up to full speed.)
None of which are honest about how things have progressed.
Ensuring said bad press is accurate/fair is certainly not the primary goal.
Disclaimer: I don't own any TSLA stock.
> Tesla Ticker Symbol + "Q" which is the NASDAQ notation for bankruptcy
https://en.wikipedia.org/wiki/TSLAQ
Tesla had a battery swapping but realized it was a terrible idea and instead built the largest global chairing network in the world instead. Are you seriously gone shit on somebody because they changed strategy with more information and build something that is actually more impressive?
Yes, he was wrong about Robotaxi, or at least its delayed. Nobody bats 100%.
But if you are actually honest with yourself and start in 2000 and look what he said, compared to what he did. Literally nobody else on the planet predictions are closer to what actually happened in terms of large scale problems.
Musk isn't super-human. He's a good engineer with great intuition for how to solve hard problems and a good "first principles" approach. He also has a great work ethic. There are tons of Elon Musks around.
What's rare is people as visionary and competent as him who are also rich and well-connected. Having someone who actually knows something in a position to make decisions about the allocation of large amounts of capitol yields results so amazing that they appear super-human, which buys him a hell of a reality distortion field and a lot of slack for being eccentric.
Musk got rich more or less by accident. The systems that advance people to those heights select for competence in the areas of social and financial advancement, not practical engineering or creativity, hence his rarity.
Of course today it's probably better than it ever has been in all of history. Historically the way to wealth and power was to brainwash a large number of people into killing (and risking their lives) for you and then sitting back and reaping the reward. That's still a thing, but at least now there are other routes.
I used to think like you, then I see these driverless demonstrations and negotiated left turns with zero user interaction.
I have this feeling that we will be surprised like we were when AlphaGO won...
1. Be robust to bad visibility, not an easy task in computer vision.
2. Have access to extremely detailed maps that include things like driveways and parking lots. These maps would need to be continuously updated. The labor required to make and maintain such maps will limit how many places cars can be fully autonomous.
3. Understand the etiquette of the road. For example, a person seeing a car stop in the street near a parking spot would assume the car is about to parallel park and give it room to pull off the maneuver. An autonomous vehicle wouldn't give it so much room, because it can't reason about context in that way.
4. Negotiate ambiguous situations requiring interaction with other drivers. There are frequently situations where it's not clear who has the right of way from the rules of the road. People resolve these by gesturing with their hands and with their cars. A fully autonomous vehicle would need to understand these signals.
One might easier imagine cars communicating with one another over a local connection of some sort.
If we want to optimize for reducing bullshit, the stock market and rules around it need to be restructured.
With that said, I have real doubts that they can get their with the current camera suite. IMO, they don't have enough redundancy. Further, I think elon's anti-lidar position is a mistake. Even if they can make it without lidar, I think it will set them behind the rest of the industry.
Likely, we will see something where they collect enough training data to hit level 4 + training suite and eventually they'll release cars with new sensor suites. Once that happens, my bet is that's when they'll actually release level 4 capabilities. I highly doubt any current car will ever allow hands off driving.
Now, I am naïve about the real intricacies of computer vision. But won’t snow be quite nasty for camera-only systems? It’s almost the physical definition of noise, as it’s random and reflective. Is there any research or experience with computer vision and dealing with a snowstorm? Because even human drivers aren’t great in that kind of low-vision scenario.
(Why yes, I did grow up in the Midwest.)
A human can drive in the snow, and has a sensing package which effectively consists of two cameras.
The hard part (when you only have two cameras) is that the processor needs to be exceedingly sophisticated, e.g. a human brain.
People like Waymo have decided the best way to decrease the brain requirement is to add more sophisticated sensing that humans are fundamentally incapable of.
Tesla's strategy is more to just build a better digital driving brain with machine learning + lots of miles, because arguably all sensing mechanisms have their shortcomings. That being said, they still have a lot more than two cameras in their FSD package – just not LIDAR. They still have radar and other things as well.
Only time will tell which approach is better. Probably in the end the best approach will be to combine the two: a brain that is sophisticated enough it can drive with just cameras, but you throw LIDAR and such in their too to go the extra mile.
Two cameras, and a huge amount of mostly not understood wetware interpreting their signals. Our current approaches to ML may be a local maxima. The true path to AGI be somewhere as yet undiscovered. Lidar certainly seems like a good tool to use given our current ML capabilities.
No LIDAR means more training data, as has been made clear by the amount of miles driven by Waymo vehicles vs Teslas in Autopilot.
A generally intelligent driving system is going to benefit from better sensory inputs just like a human would, it's not like that makes the system dumber. The whole dichotomy between sensory inputs and processing seems strange to begin with, there's not necessary a clear line, or a reason why one should be prioritized over the other.
You have to realize that most SOTA self-driving systems are using a lot of NNs. NNs in turn want lots of training data. That training is not fully generic – in other words, if you train a NN using only cameras, giving LIDAR to that trained network (i.e. for inference) will not really do anything.
Having large LIDAR arrays on the top of vehicles is expensive and slow to deploy, which directly translates to being able to generate less training data.
So if your goal is to have the best possible self-driving vehicles 100 years from now, absolutely you should be using as much sophisticated sensing equipment you can get your hands on, and slowly train your nets for the next century. If your goal is to get out something that is better than a human driver as fast as possible, then it is not clear that requiring all vehicles to have LIDAR on them is the best way to go.
The argument for having a lidar vs not having a lidar is different. They chose to go without lidar because it should be cheaper and faster to get started, and the fact that people are capable of doing driving without lidar makes it seem possible that a machine could do it without lidar.
On the contrary, having computer vision capable of handling snow is a necessary precondition for self-driving, regardless of what other sensors are present.
I think this is a huge simplification of the actual problem. Not only do you have to denoise the data, you have to use the data to generate models of 3D objects in space. Maybe you could argue that exact 3D models are not necessary, but IMO any neural network that can drive at Level 4/5 needs to approximate 3D space in some layer.
This is incredibly difficult; we have an entire region of the brain dedicated to for this exact purpose (the visual cortex). This region has the benefit of millions of years of evolution. We also have the advantage that we can predict 3D models based on everyday experience; we don't know how to encode that knowledge into a neural network efficiently. That's not even mentioning the latency requirements that driving in the real world would impose on such a network.
The difficulty imposed by snowfall itself is just de-noising.
Keeping the context of a self-driving car, humans are perfectly capable of driving cars with just two eyes to determine surroundings. For 99.9% of your driving, you don't have to think about it. Stay between the lines. Stop for the red light. Make sure there's nobody next to you before changing lanes. There's no creativity here, just following a very concrete procedure.
Humans fail because humans do stupid human things. They text while driving and drift over the lines or don't see the red light. They don't always check before changing lanes. They drive drunk. Computers won't do any of these things.
And they'll do it better. Humans have only two eyes and can only look in one direction. A computer can check the blind spot while still looking ahead, so there's no chance it will rear-end someone because the person in front of them slammed their brakes while it was making sure the next lane was clear before making a lane change. A computer will notice a car in the blind spot veering into its lane without taking attention off the road in front.
Again though, it's all down to the software. We're not there yet. But that doesn't mean we can't be in the future.
More often than not, snow driving is CHALLENGING. And it's not for a lack of experience, as I'm doing multiple roadtrips through the BC Coquihalla Highway. More often than now, it's more of a matter of best attempt at driving in the path made by other car disregarding lines on the road than driving by the book. The utter HELL being driving at night with a blizzard, this is a level-up experience for sure.
1. Many people have no choice but to drive in bad snow conditions for any number of reasons including that they're emergency workers.
2.) Winter conditions catch you. Even though I don't commute and basically do everything to avoid bad winter weather conditions, I've still been caught in pretty bad conditions a couple times over the past few years.
Fuck them...
Is that how much a pussy society we have become ? Avoid dangerous conditions because "it might be dangerous" ? If a freaking Tesla can't handle what my 06 SUV can handle without issue, and what people were handling ~forever, ... just as much as an Apple laptop can't handle what a then-IBM was handling, then the damn "car"/"laptop" is nothing but a piece of shit.
Urbanus homo sapiens sapiens is really a pussy...
It's worth noting there are many accidents caused by people driving in snow and that AI could be hugely safer at some point.
And also having to drive manually when it is snowing is probably fine for v1.
I would worry more about bicycles, they can be extremely tricky if the rider doesn't understand how to explain their presence on the road to drivers.
If I had some button that made spikes or another traction method pop out of my tires and gave me great grip I probably would drive full speed in the snow; because the main issue would have been solved.
I don't expect that any self-driving systems will start out with handling all weather conditions. Even in the long term they may handle fewer conditions than humans.
But neither is insane to me. Structure-from-motion is all you could get with monocular vision, and I would not trust that as the primary means of depth estimation, especially with no redundancy.
From the very same letter, there is also this quote:
Please note that Tesla’s development of true autonomous features (SAE Levels 3+) will follow our iterative process (development, validation, early release, etc.) and any such features will not be released to the general public until we have fully validated them and received any required regulatory permits or approvals.
Any bozo can make claims about the future.
And you're a bozo for believing them. Do you believe everything you see in writing?
What's next? "Really, truly full self driving" that's actually level 3+?
People who buy it know what they are getting, and if they don't, it's entirely willful ignorance. I don't even own a Tesla yet I've read plenty about the actual capabilities and the fact that these cars can't truly drive themselves.
Yet, if I bought a Tesla I’d still need to sweat the decision of buying the FSD option because Musk insists the resale value of cars without it will be significantly smaller without it.
I'm a robotics person though so it was always pretty clear to me that I couldn't possibly think of being able to deploy an L4 system with their set of sensors.
The problem is (a) the stock market is full of idiots that won't listen to real engineers like us, and listen to some bullshitty CEOs and analysts instead (b) the SEC and stock market is perfectly optimized and incentivized to maximize the amount of marketing bullshit coming out of company PR.
If people wanted more honestly worded things out of companies, we need to stop nailing companies every time they don't "beat" earnings estimates every quarter, and begin to understand how real science and real technological advancement happens.
The way NASA is funded, for example, is the exact opposite, and minimizes bullshit, but NASA and Tesla share a lot in common in terms of what they're trying to accomplish for humanity.
EDIT: Okay maybe NASA wasn't a good example. But in any case, the current structure of public companies and letting a bunch of gullible <strike>sheep</strike> average citizens collectively determine the value of a company based on what a bunch of <strike>clowns</strike> non-engineers in suits at Wall Street try to make of a highly technical venture isn't a system optimized for minimizing bullshit, it's quite the opposite.
So you are a real engineer and Musk is not?
> begin to understand how real science and real technological advancement happens
By pushing for it hard, investing your credibility and a lot of man power and resources into it? Like literally what Elon Musk does with all his projects.
> The way NASA is funded, for example, is the exact opposite, and minimizes bullshit
NASA claims nonsense literally all the time. The #JournyToMars under Obama for example.
Real engineers don't bullshit. Real engineers understand that there is a certain responsibility that comes with people trusting their lives to their engineering skills. Real engineers can give you confidences on everything in hard numbers with data. Real engineers can provide realistic estimates of timelines. Real engineers make mistakes sometimes, admit them candidly, and know perfectly well how to adjust those timelines based on past estimate errors and extrapolate adjusted project timeline estimates with basic math. Real engineers know the limits of systems and how to write spec sheets that describe them accurately. Real engineers don't always know the answer to every question, and know when to say "I don't know".
It's managers, investors, analysts, and sometimes customers who don't want to listen to those extrapolations and estimates and uncertainty. And so many real engineers gradually become increasingly fake, because the other (corrupt) parts of the system spread their toxic thinking into the engineering.
I have a lot of respect for Musk, who was a real engineer, probably REALLY wants to stay a real engineer, and is still mostly an engineer. But he is indeed forced to be a part-time <strike>bullshitter</strike> non-engineer by the system, which demands bullshit by its very incentive structure.
When I told my friend about self driving Tesla who's a BMW fan, he said that BMW will just sell self driving cars when it's ready.
Really? Without gathering any data? I'd say BMW is the bigger bullshitter. Elon needs to update the computers maybe, but everything else in the cars is ready for self driving im the near future.
I mean, as I said, the whole stock market is full of bullshit. My point is the system is designed to incentivize bullshitting, so it's expected given the incentive system created. Companies that don't bullshit don't win, and if that's the loss function you define, bullshit is what you're going to get.
> everything else in the cars is ready for self driving
I disagree with this. I think significant hardware upgrades to Tesla cars will be necessary to achieve L4. Any attempts at L4 without those upgrades will result in a lot of accidents. I won't be surprised if they come out 5 years from now asking former Tesla owners to get their cars retrofitted with some new sensors if they want L4.
L2 is achievable with their existing hardware and that's what they are doing.
I personally agree with them that technologies like PseudoLIDAR will be sufficient.
I believe that for a good human driver Tesla sensor suite would be sufficient to drive safely if you covered all windows. You could project the camera views on covered windows and keep the current bird eye visualization.
The issue really is with control of the car. Deep Reinforcement learning sucks. Training Deep RL is unstable, hyper-parameter sensitive, even random-seed sensitive, and requires ridiculous amount of data for even the simplest applications. And after training it does not generalize.
Completely false.
So when the CEO went on stage and said cars would be self-delivering via fully autonomous driving it was obvious to you that his statement was a lie?
I've also probably coded way more lines of autonomous driving code than Elon Musk himself. That isn't to say I don't respect him and his vision, but I'm pretty certain I have a better idea of timelines than the BS he's forced to say to the public.
Either he is an alien playing 5 dimensional chess with time-travel, or he is pathologically overoptimistic. He is especially overoptimistic about AI, so it leads him to make stupid predictions about when we will have cars without steering wheels.
Well, yes? That's how subsets work.
Thought experiment, pretend "City Streets" was called "Level 2 Mode". Would you be upset that "Level 2 Mode" will always be level 2? Or that "Level 2 Mode" is a subset of "Full Self Driving"?
That would be an accurate name for "city streets". Except currently, it's part of an umbrella feature called "full self driving" that's supposedly going to be "level 5" soon. This is it with most things with FSD. Qualifiers after qualifiers for a clearly misleading name.
Kind of like calling NNs a "subset" of General AI
If a subset of features that is part of "full self driving" cannot ever reach level 4/5, then "full self driving" is not possible.
"full" is a marketing terminology but the English meaning of it would be that it works comprehensively in the majority of cases. If it doesn't work on city streets, which is a major subset of use cases, how can it be "full"?
There is an enormous difference between "city streets" the location, and "City Streets" the name.
They're not saying the car will be stuck at level 2 on streets in cites.
They're saying that this particular feature which is named "City Streets" will always be level 2.
Eventually (they promise) the car will have level 4 driving on city streets. Level 4 driving inside cities will either have a different name or no name at all. At that point they will probably delete "City Streets", because it's a subset of level 4 and there's no reason to keep it after level 4 is finished.
Edit: especially because the full name is "Autosteer on City Streets". When they have level 4 support, they definitely won't be calling it "autosteer"
I think the only one that doesn't get this are journalists, regulators, and class action lawyers.
Prediction:
* Ridiculous Self Driving
* Ludicrous(+) Self Driving
* Plaid Self Driving
They publicly advertise and collect cash for pre-sales of "Full Self Driving". Most Hacker News technologists (a generalization for sure) would reserve a phrase that bold for Levels 4 and 5.
But here they are in writing stating that they are currently only at Level 2.
So a couple core questions around Tesla are:
- Is it acceptable to take pre-orders from the average consumer for a feature that might be 10 years away?
- Is it ethical to use exaggerated marketing speak for specific technical functions (to give a non-Tesla example, this is like when I see companies implementing rule-based decision trees or linear regression and calling their apps "Artificial Intelligence")?
Are they going to keep winking at customers and letting them think it's level 3 so they continue to take their hands off the steering wheel and do other things like countless Youtube videos already capture?
It's a little disturbing how people will jump through so mental hoops to give Tesla the benefit of the doubt.
AI means so many things that it has become as meaningful as "magic". Some don't consider anything less than human-level an AI (this is also called AGI or strong AI). On the other hand enemy behavior in video games is commonly called "AI" even when it is as simple as pacman ghosts. In the same way that magic can literally mean "impossible" or be applied to card tricks accessible to 5 year olds.
It seems strange considering the wide availability of information and communication, but apparently there either is no feedback loop back to the people who write like that, or there is but it's mostly positive (be it more sales, less lawsuits or a greater following). At some point the meaning of those words will have hollowed out so much that we can't use them in normal conversation anymore.
They’re basically long-term 0% loans like the Cybertruck deposits. They keep it in escrow and slowly recognize it feature by feature.
Can you cite that? It directly contradicts the linked article which claims that the CA DMV contacted Tesla to clarify if Tesla were testing a vehicle without proper permit.
If Tesla were seeking some kind of approval it makes their email very confusing as Tesla is insisting that they don't need additional permits/approval for the "City Streets" beta.
Here's the original email sent to Tesla:
> Subject: Announcement of Full Self Driving Feature.
> Al we’ve seen press reports that Mr. Musk announced on Twitter that a “Tesla holiday software update has FSD sneak preview...” Many people generally translate “FSD” to be “full self-driving.” Is Tesla releasing a full self-driving software update to any California Tesla owners? As you are aware, the deployment of autonomously driven vehicles on public roads in California requires a permit to deploy. At this time Tesla does not have a permit to deploy. Please provide an update on what this announcement for the deployment of the feature means in terms of autonomous operation on public roads in California?
I don't see anywhere where Tesla are seeking approval, quite the opposite in fact. Tesla are saying they shouldn't be required to seek approval for "City Streets."
If "Tesla sent a letter to the CA DMV requesting approval" per the above comment, that would have meant Tesla were trying to get an FSD permit in CA (the opposite of what happened).
These docs don't really shed light on Tesla's FSD plans so much as they illustrate the lengths to which Tesla will go to skirt regulation and liability. All the chatter and rumors about Tesla FSD just drive more attention and more potential users (trainers) into their pool.
What's really required for full self-drivung capabilities is infrastructure support. This will come as large eatablished car manufacturers and governments start defining standards for that.
Here's a very simple example. Imagine a stop sign. Tesla's and Google's aproach is to detect the stoo sign via cameras. This can be problematic when the canera can't see the stoo sign (weather, tree branches blocking it partly, a truck in front of you). It would be much simpler if the stoo signbwas able to communicate to self driving cars that they have to stop. Now imagine an accident because your Tesla didn't recognize the stoo sign. Who is going to pay for the damages? Tesla for false advertising? The city for not cutting down the three? Your insurance? You?
Obviously Elon isn't entirely responsible for it, but I'm constantly impressed by his initiative in tech world. If Starlink weren't available, I'd be waiting ~3 years for Amazon to build some sort of equivalent service.
I still have the occasional drop out... I can't make it through a video conference, or wifi call without a freeze up or audio drop. My Starlink app says "No obstructions" but also says "Obstructed for 4 minutes" on a regular basis. Many people seem to echo this. I'm not honestly sure if this is what to expect or not.
I generally switch to my 1.5Mbps DSL line for phone calls and video conferences, which lowers the quality a ton, but at least I don't have to ask people to repeat the last 30 seconds. Web surfing and downloading large files is blazing fast, and multiple people sharing the link seems to have no impact on the quality. Video streaming is fantastic. Another downside is that you're behind a CGNAT, so you'll never get a public IP address. Then their router puts you behind another NAT...
StarLink is better than nothing for sure. But if Comcast didn't want $25,000 to run 550 feet of cable to my house, I would probably use that instead. I now have to call in an arborist or someone and top some trees that I am guessing are causing temporary obstructions, but I'm not really sure. I have a lot of trees on my property.
"users can expect to see data speeds vary from 50Mb/s to 150Mb/s and latency from 20ms to 40ms"
https://www.starlink.com/faq
My only concern has been the stability of the connection, there are still intermittant disconnects happening on a daily basis, and it comes out to <3 minutes of downtime per day. If that's a dealbreaker, it might be best to hold off until they get it sorted out. Besides that, it's a great deal. The dish is dead simple to install, and even the router they provide is pretty damn good (sadly it only provides a single RJ45, so keep that in mind)
tesla_fan_88 is brand new so I think maybe they stole your msg..
Sort of like self-deprecating humor might not be the entire picture of mental health, but a sign that it is probably ok.
EDIT: I think if Elon Musk gets too serious, or gets shut down, I suspect some ingredient of entrepreneurial spirit might be lost. The "do the right thing poorly" vs "do the wrong thing correctly"
* Achieved dominance in the commercial launch market, arguably with an even more substantial technology lead.
* Performed the first landing and reuse of an orbital rocket booster.
* Developed a heavy-lift version of that rocket that has the largest payload to LEO out of any currently operational launch vehicle. Also with reusable boosters.
* Delivered a new human spaceflight system to NASA, the first such system since the Shuttle ~4 decades ago despite decades of promises and plans to the contrary. Starliner, the incumbents' entry for the commercial crew program, has not yet flown and is not expected to until later this year at the earliest.
* Developed the first full flow staged combustion cycle engine to leave the ground under its own power.
It seems really questionable to bet against them at this point, and minimizing their accomplishments to date is really, really silly. Not sure why you have to involve the rover in your comparison, though—it's also a monumental achievement, but it's barely comparable to the domains SpaceX is working in with their launch vehicles.
Driving on a highway is an amazing experience when it's mostly empty, and a nightmare in rush hour traffic.
Has there been any good write up about how well Starlink (or any other satellite internet) could feasibly scale? Even if SpaceX can start dumping tons of satellites for dirt cheap with Starship's economy of scale, surely there's a limit where the risk of a Kessler syndrome event makes further expansion impractical.
There are currently about 1.5 billion cars on the planet.
They are confined to land which only makes up 29% of the earths surface.
If evenly distributed on that 29% (~57m sq miles) of earths surface (196m sq miles) there are about 25 cars per square mile.
The orbital shell in which SpaceX is flying satellites starts 550 miles up. If you confine yourself to a 2D spherical plane, this shell already has 30% more surface area than the earth, and you can use all of it, not just the 29% that is over land. If we "packed" orbit at the same density as cars across that 2D spherical plane we could fit 6.3 Billion satellites with a separation of 1000 feet minimum between any two satellites.
Of course we aren't limited to a 2D plane in space. You can fit millions of satellites in orbit with miles of separation between any 2 without issue. The bigger problem would be signal interference.
Tesla intends on building a great L2 system, releasing it publicly, and then moving on to L3 afterwards.
That being said I agree with the spirit of your comment.
Whether that's a mistake depends on how long that amount of time it.
For example, 10 seconds is perfectly safe. It's fine to completely ignore the road until you hear the beep, because 10 seconds is plenty of time to look around and figure out the situation and take control.
I would argue that any system that requires constant human concentration is not actually level 3.
Level 2 is the one that's dangerous by design.
Edit: Though since the main difference between level 3 and level 4 is the ability to pull over, I wouldn't be surprised if level 3 gets quickly left behind or skipped entirely by many developers.
- You're deep into you Netflix movie
- +2 seconds. Oh wait, there's an alarm going off. WTF?
- +4 seconds. You've acquired some level of visual awareness of the environment but still don't know what's really going on.
- etc.
You were in a deep sleep? Maybe need a good minute and some indication of why the system has disengaged.
But even in your scenario, if you hit vague awareness of the environment at 4 seconds, and you know you're taking over to drive... that 6 more seconds is more than enough to reach full environmental awareness.
> You were in a deep sleep?
You're not supposed to sleep for level 3. And "don't sleep" is not some ridiculous requirement. It doesn't even compare to the "be constantly vigilant" rule of level 2.
It's one thing or the other. You're being fairly vigilant and looking ahead and ready to take control or you're a passenger.
> Full self-driving to me means that I can act as if there's an Uber driver at the controls. As far as I'm concerned, it's either that or I'm paying full attention even if I'm not fully controlling the vehicle.
> It's one thing or the other. You're being fairly vigilant and looking ahead and ready to take control or you're a passenger.
It's the latter option.
In level 3 there's an Uber driver but they might have you take over at some point. You're not in control at all, a passenger until the car tells you otherwise. You don't have to have any vigilance or looking ahead, you just have to be awake.
If you need 10 seconds to react whatever was going to happen has already happened anyway.
That is, if humans only need two eyes to drive, then why not just only use two cameras? Isn't using more than two cameras (and radar) admitting that cars need more "help" than humans? If using more than two cameras is okay, why not go further and use LIDAR or HD-mappings?
I believe that Musk chose this line because he knows that LIDAR (in its current state) is too expensive to sell to consumers. Same with HD-mapping. We know that LIDAR based systems have achieved "better" self-driving than camera based systems.
It's still possible that Musk is correct. But at present, it seems to me that Musk's position therefore seems like a mostly ideological stance, and not based on any real data. That's why people doubt Tesla's ability to execute on Level 4/5 self driving. The market will punish ideological purists when they are wrong.
You could also argue that because LIDAR is expensive, it's hard to get billions of miles of training from LIDAR-equipped vehicles. This is not the case with Tesla's sensing package, which is racking up the miles as we speak. So sure, expense makes it less appealing from a consumer perspective, but its also possible that the difficulty of deploying LIDAR fleets makes it less likely to achieve FSD in the short term due to less capacity to generate training data.
Long-term we'll probably want all our vehicles to have as many sensors as is feasible. But nobody trying to get to FSD today is worried about what it will look like in 100 years – they're all trying to get their first.
Er, Waymo is doing a commercial L4 service. Cruise is doing pretty good. So no, they are not all "ideological positions". Also, no one except Tesla is targeting L5 which should tell you what the realistic goal is.
> You could also argue that because LIDAR is expensive, it's hard to get billions of miles of training from LIDAR-equipped vehicles.
LIDAR is expensive was a good argument in 2015. The prices are plummeting. Waymo reduced their top of line LIDAR cost from $75,000 to under $7500 and this was in 2017. They sell their short range LIDAR (Honeycomb) for under $1000. And their 5th gen hardware costs even less now (my guess is under $5000). Which way do you think the price will go as a number of companies mass produce it?
The fact is Elon made a bad bet a few years ago and didn't anticipate how much the cost would drop. Now Tesla is stuck with his ideological stance trying make a pipe dream work.
That last sentence seems like a strange one. "Stuck" would be if you massively invested in LIDAR and it didn't help. If no-LIDAR FSD is not as feasible as Tesla hopes, they can just... use LIDAR. And they'll arguably be in a better position tahn early adopters because they weren't investing when the hardware was expensive.
Waymo is L4. L5 means it can take you to your favorite campground (anywhere) and handle any scenario. Waymo says it’s not a realistic goal. But yes, the normal levels are not very well defined.
> If no-LIDAR FSD is not as feasible as Tesla hopes, they can just... use LIDAR.
Tesla has promised the current fleet has all the necessary hardware to make it work (they couldn’t possibly know this at the time). So they can’t walk back on that without incurring huge costs for retrofitting existing cars and development cost for overhauling their tech stack.
Meanwhile, they would’ve sunk money and time pursuing a much harder problem. If they had used additional sensors from the beginning, they could’ve provided much more value to their paying customers by now.
Edit: Just saw your edit.
> From what I've seen of Tesla's FSD beta, it could very probably be T4 if Tesla geofenced certain areas and worked out any particular kinks manually.
Yes, that’s what they should have done. Make realistic promises. I mean, you’re talking about making it work on _any_ road, _any_ intersection and _any_ scenario. Without additional sensors and mapping, it’s seems pretty unlikely (if not impossible) to me. Certainly given the timeframe Elon Musk promises.
It's your own fleet that's collecting the really good detailed data. And adding LIDAR to your own fleet isn't tremendously expensive. You don't have to worry about an extra 10k per vehicle crippling your sales.
Level 4+ using only cameras is risky and aspirational.
In truth, most of Elon Musk's high level goals are risky and aspirational. So far his record is pretty good, but eventually one or more are likely to fail.
The track record of making things work is good, but it's a very different kind of problem when you're allowed to redesign.
Computers don't (yet) have the hardware nor software of a human brain. So they better come with extra tricks, like lidar.
LIDAR is not mandatory for better driving than humans.
The fact is Tesla is not submitting test data to the California DMV, which means they are either not developing a self driving system and have no right to speak to what technology is or isn't mandatory - or they are doing illegal testing, which speaks volumes about the trust you should put in their opinion.
Nobody has proven "better than human" capabilities, LIDAR or no.
Until you have something like the human brain powering the car you cannot just use "two eyes"
That means the only thing we know can work is eyes + brain.
The answer to your question is: we don't know what's needed, so people saying "sorry you need LIDAR for FSD" is much more "drank the coolaid" than my stance of "we don't know, maybe no LIDAR + lots of training data can work just fine". Or, turn the question on its head – how much better at driving do you think humans would be if we had LIDAR on top of our heads? My guess is... not very. Hence my belief that the Tesla approach (less SOTA sensing, more training data) is a reasonable one.
You've still got things a bit mixed up. My stance does not imply that a computer functions as well as a human brain. I am not claiming that better sensors won't (eventually) give you better results. My stance merely implies that the brain (digital or wet) is a very important part of the equation, and just throwing sensors at the problem won't fix the fundamental need to have a highly sophisticated ML system running the show, and it's easier to get the training data you need for such a system if you're not using LIDAR.
“Less data is better for self-driving cars” is not the take I expected to hear from the Tesla mob today.
Tesla is clearly gathering more data for their self-driving program than Waymo, due almost entirely to the fact that the Tesla sensing package does not require installing LIDAR on every car.
While it is not sufficient on its own, it would yet another (small) piece of the puzzle.
Obviously Elon isn't entirely responsible for it, but I'm constantly impressed by his initiative in tech world. If Starlink weren't available, I'd be waiting 3+ years for Apple to build some sort of similar service.
Take for example this well-known youtuber, AI DRIVR. This guy drives around in his Tesla talking about how amazing and mind-blowing it is, but if you mute him and just watch the video with a critical eye you can see that the software simply does not work. It should be outlawed immediately before it kills someone. He recently rampaged through my town of Berkeley, California, breaking five laws per minute (aside from the fact that he has the speed controls set 5 MPH above the limit at all times).
Here's his car running a stop sign: https://youtu.be/SNT5MzjAAms?t=219
Here it runs a red light, which the narrator takes pains to explain away. https://youtu.be/SNT5MzjAAms?t=238
Here it turns right from the center lane: https://youtu.be/SNT5MzjAAms?t=523
Here it violates the right-of-way of a car that it incorrectly classified as parked. https://youtu.be/SNT5MzjAAms?t=762
At numerous points in the video, it signals for a right turn while going straight. https://youtu.be/SNT5MzjAAms?t=806
Here it signals left while going straight: https://youtu.be/SNT5MzjAAms?t=898
Absolutely unacceptable. The government should act against this junk.
Teslas have a summon feature which is kinda level 4.
In summary: the driving autonomy level system kinda sucks.
They're clearly trying to get as far as possible (data flywheel) until they absolutely have to get authorisation from the regulators.
This is just lawyers doing lawyer stuff. Of course Tesla are aiming for complete automation.
I went for a 500km+ drive on Sunday to visit new supercharger locations (it's been a long pandemic and I'm pretty bored). It was mostly rural highways without lane separation or limited access. These are some of the most dangerous roads to drive on since there is direct oncoming and cross traffic. Speed limits are 80 km/hr, but the prevailing traffic flows at just under 100 km/hr, so closing speeds are around 200 km/hr.
I hardly ever turn on autopilot since it's just bad at driving, but there were long stretches of completely straight roads with multiple minutes passing between oncoming vehicles. Even in that specific circumstances Autopilot made some truly weak decisions about its basic driving style.
It attached to the centreline between lanes which limited the gap with oncoming traffic. When I drive myself in the same circumstances I attach to the outside of the lane to maximize the space I can give oncoming cars. With the amount of room autopilot was giving even a small twitch or moment of distraction would be enough for disaster. If the event was timed poorly enough even a computer system would have no chance to react in time to physically manoeuvre the car out of the way.
Autopilot makes no accommodation for blind spots. When traffic is flowing at full speed on limited access or other multiple lane highways I adjust my relative position with other vehicles to ensure I'm never spending long periods of time in another driver's blind spot. If we're travelling the same speed I create a gap to ensure a quick lane change won't result in contact. Autopilot will continue along completely oblivious and no taking good defensive driving precautions.
There is a youtube channel featuring collision footage from Tesla onboard cameras. Nearly every collision when autopilot is engaged is a situation that never would have occurred in the hands of a skilled driver who anticipates issues and prevents dangerous situations from every occurring. Without knowing the specific collision that's about to happen in the video in most cases I can see a dangerous situation starting to develop I would have reacted to seconds before the collision actually occurs.
But you tend to know a collision is coming. Most accidents are due to people being comfortable and not being as alert as they should be; your example is the opposite.
AP is currently ~good on highway (probably safer than I am in good weather), but Meh elsewise. Hopefully it continues to improve; but it will be awhile still i think; and iffy to ever get there on our current Teslas.
autopilot creates danger where none needed to exist by not using the full width of the lane and failing to create space through relative small speed adjustments. The cues it gives off while driving are some of the exact things I watch for when trying to identify beginning drivers so I can give them even more space.
I certainly would not describe autopilot as 'good' on highways. I'd barely call it adequate.
You just wrote a long text about why it's important to use self-supervised learning to create a birds eyed view of the world, because object detection and planning from camera only view doesn't work well enough. Yes, we knew it.
The issues I have with AP have little to do with their object learning or need to gather more information (though I agree camera-only isn’t enough). The car knows where the lane boundaries are but doesn’t take advantage of the width. It knows accurately enough where other vehicles are that it could avoid staying in another driver’s blind spot indefinitely. Tesla is teaching their cars how to drive with terrible habits and not building defensive driving into their rule set.
If you are in the right region, you can try the FSD beta soon enough, which has showm amazing driving skills already from the Youtube videos I have seen.
I live in the EU though, so I guess I have to wait at least a year or 2 to try the beta :(
How many of those beta software programs that you use on your PC would you trust enough that if they "crashed", you would be dead/horribly maimed?
"Beta" should not exist on public roads, except under very controlled conditions.
The harder parts of machine-piloting involve interpreting the world (lane-finding in the absence of striping; distinguishing hard obstacles from ephemeral objects like paper, leaves, weather, reflections, and so on); and participating in a social environment (what driving postures indicate what intents? What do the local humans consider rude or polite? What are they expecting?)
(A truly advanced autopilot would seem almost magical to a human— it would be able to back-solve diffuse and specular reflections in the environment, and accurately deduce the presence of oncoming cars around corners that would be totally invisible to a human driver. It could predict accidents several seconds before they begin to play out, and the sudden braking/maneuvering in the absence of any visible threat would seem strange and confusing to the humans inside).
Tesla is nowhere near cracking these things; they're still on the ground floor. I still predict that current pre-buyers of so-called "full self-driving" are not going to get what they were promised, and going back on that will probably be bad both for Tesla's reputation and their wallet. If I had to guess, I'd say that if Tesla delivers on Level 5, it will be with a hardware platform that doesn't exist yet, and a software platform that is close to "complete re-write".
As it stands, their misleading marketing on self-driving capability is both dangerous and borderline fraudulent.
All that being said, I'm long on self-driving as a technology; just short on Tesla's current approach to it. It's a shame because I suspect both Waymo and Cruise are doing significantly better technologically, but I trust in their ability to execute a product significantly less. I'm not sure who will actually crack it.