I feel Tesla livestreams are very much like their cars - they never launch on time, but when they eventually do, they're pretty good and worth the wait.
While you're waiting for the main event to start, here are some recent interviews with Elon about self-driving cars. He's very confident.
"To me right now, this seems 'game, set, and match,'" Musk said. "I could be wrong, but it appears to be the case that Tesla is vastly ahead of everyone."
I agree. The interview with MIT researcher Lex Fridman was difficult to watch because it didn't seem like they were on the same page at all - Lex asking thoughtful and pointed questions and Elon dismissing them as if the questions themselves are moot because self driving is right around the corner.
It was mind boggling. I am hoping Tesla can provide some specifics today because it seems Elon is living in a fantasy world (albeit one I'd like to live in if we can actually get safe self-driving cars).
I'm gonna say Elon is being extremely bold selling a technology that's current leader in deaths behind the autonomous wheels.
also I can't reconcile how the new hardware is this huge leap ahead beyond raw computing power if, by Tesla own claims, previous hardware was perfectly capable of autonomous driving.
seems people were getting fooled either now or before.
And Elon has a long history of making false claims about Tesla’s progress. For example in 2015 and 2016 he claimed that Teslas would be fully self-driving by 2018.
But Navigant curriculum was very unscientific. There no actual quantitive reason Tesla is worse. Is was based mainly on business factors like go-to market strategy and vision.
Yes absolutely saying Tesla who gets camera data from it's a half million car doesn't give it an advantage is crazy. That not even including the fact it's the only company who can do its strategy. Google would need to get constant data and GM and legacy automakers would need sensor suites on all it's cars yesterday.
No one knows if Tesla strategy will work because they don't have the data collection in place.
Based on their talk today and Andrew previous talk where he shows explicitly tools that do just that download data constantly is exactly what they do. https://vimeo.com/274274744
I mean saying a phone can upload videos to youtube but a can can't to tesla is a weird ledge to stand on. Even their windshield wipers work based on sending video data to tesla to be learned on.
The first article you link to sources another article as its source, which itself calls bullshit on the ranking.
Your link:
> According to Electrek, Tesla trails behind other companies in terms of autonomous driving tech based on a list created by Navigant Research, an independent research firm.
Electrek’s article:
> Electrek’s Take
> I think Navigant’s autonomous leaderboard is ridiculous. There are way too many brands that keep most of their development under wraps, which makes it hard to evaluate them and therefore, it gives very little value to a leaderboard like this in my opinion.
My guess is he means "on the highway". The scary bits of self-driving is person detection, crossing detection, roadwork detection, cyclist detection (e.g. coming up on the right when you are trying to make a right turn).
The Waymo end-game that I heard was "able to go through a drive-thru". I highly doubt Tesla is anywhere near that point.
There have been news reports about the model 3 autopilot getting its speed limits from maps, lacking any sort of sign recognition or manual override to adjust to local conditions. The maps seem to be outdated for germany (1). That’s an essential feature even on the autobahn. Given that test result I’d even be skeptical about any claims of being ahead of the game on the highway.
This is very strange though, is there any confirmation of this?
Basically, most other manufacturers like Opel, Audi, Mercedes, Hyundai, VW, Volvo, Ford, etc. has had for several years the feature to detect speed limit from computer vision recognizing the road signs. And it works reliably, as is pointed out in your link.
How can Tesla be a leader in using computer vision for cars, but not be able to read the road signs?
The kind of drive-thru that Tesla is currently associated with involves semis rather than fast food and it would be really nice to hear that they've at least licked that particular bug (and for good, this time).
> The scary bits of self-driving is person detection, crossing detection, roadwork detection...
Your point it very astute.
Among a few other ML/AI MOOCs, I completed Udacity's "Self-Driving Car Engineer" nanodegree - so when I'm out driving, I often come upon situations where I wonder "how would a self-driving car navigate this?"
Today, driving in to work (note: USA), I noticed one intersection I've been through many times before, and that question came to mind. The intersection is interesting, because on approaching it, the road curves to the right, and you can actually see one of the traffic lights on the left before you even see the intersection. By the time you see the intersection, you're already on top of it.
So as you round the curve, you see the lone traffic signal (red/yellow/green); if it is red, do you start to brake, or do you wait until you can "see" more traffic signals? If you wait - will you have time to slow down and/or stop? ...and so forth.
This and others are all kind of "edge cases" that will need to be trained on, and/or perhaps other cues for self-driving vehicles installed or set up so the vehicles can navigate such areas successfully. I know when I first went through the intersection it was a bit of a surprise; it's not a very safe intersection (going home in the opposite direction is not any better; in that direction, you're headed downhill, have to cross the intersection, and immediately start turning to the left after going through - the curve is really abrupt, and you have protected/unprotected left-hand turns both directions, etc).
Well they’re vastly ahead in one area: data collection. No other company is even close. You could argue about the quality of data but the platform is there and ever growing, and they can upgrade their hardware in the future and augment existing data.
Don't kid yourself, the car has no bandwidth storage or performance to send back anything other than a few raw frames from disengage events or other rare triggers.
It depends entirely on how they design the system. They don't necessarily need to send all the data from the cars back home when they can send test cases to cars, run the tests in a shadow mode to collect real world results, then send the test results back home.
I hope Tesla has strong governance controls over customer data, and a fierce inside counsel for pushing back against unnecessary or overly broad LEO requests.
Why would you say that? The car has LTE and connects to WIFI. It could easily send way more data than any care company at any time including over WIFI.
What they are sending way more data because from our knowledge GM and Ford are sending back 0 data and Waymo doesn't have half a million cars worth of data internally to pick from.
No, it's spot on. It's entirely what I said: the car can only deliver a few raw frames, and only in response to particular triggers.
Notice the cherry-picked examples in the presentation. There is a whole class of problems the field cars can never help with, since they lack the dead-reckoning sensor setup and precise odometry a development car would have.
> There is a whole class of problems the field cars can never help with, since they lack the dead-reckoning sensor setup and precise odometry a development car would have.
Can you give an example? I'm curious what kind of triggers strictly require lab-calibrated hardware.
Which presentation did you watch? Karpathy said specifically "it's not a massive amount of data, it's just very well picked data" when talking about how the cars only send data when one of the configured triggers fires.
There’s a large gap between ‘a few frames’ and a massive amount of data, and the amount sent lies somewhere in the middle. Clearly they can’t send all data (nor would they want to) but it seems it is sufficient for significant learning to take place and the examples shown were good quality over at least a few seconds, so hundreds of frames for each example.
how can you claim that? more than waymo? that would be extremely doubtful. google has been driving around cars with sensors and cameras for over a decade.
I really really don't want Tesla to die. I think it's an important company for a sustainable future. Failures in autonomous driving could easily turn into the straw that breaks the camels back.
If you call something "autopilot" and promote it as if it will drive for you and then it ends up killing dozens of people ... that's where successful class actions come from.
And I don't want them to succeed because they are actively hostile towards 3rd party repairs and car modifications, plus their current model 3 microtransaction bullshit scares me as a customer, I'm dreading the moment other companies catch on with that. I'm specifically talking about the fact that the base model 3 ships with heated seats out of the factory but you can pay to have them unlocked with a software update. I suspect that having them enabled with a simple software mod would get you accused of piracy since you're using something you haven't paid a licence for(even though you've obviously paid for the hardware)
> I'm specifically talking about the fact that the base model 3 ships with heated seats out of the factory but you can pay to have them unlocked with a software update.
If you think this is a new thing, you haven't followed the car industry much.
Just as an example, my 2002 el cheapo Peugot was bought without the option to show instantaneous MPG/average MPG/range/etc. Spend two hours soldering/gluing on the missing $2 toggle switch, ask a friend with the Peugeot Planet update tool to enable it in software, and Bob's your auntie.
It goes much further than that though, especially when you get into chip tuning. Software upgrades that add 50 horsepower to your engine output are commonplace.
But as for the "actively hostile to 3rd party repairs/mods", this part scares me to.
I can sort of understand the software differentiation - I have no issue with Windows Home and Pro editions having different price points even thought the code on the disc is identical for instance(maybe that's hypocritical of me?). But I do have an issue with paying for hardware(heated seats in this case) and that hardware being disabled until a payment is made. I don't know, it just feels different.
You're putting the cart before the horse on the heated seat argument though. It's cheaper to have one factory line for the seats. Ideally you would pay the same price for the car without heated seats whether or not the hardware was physically there.
It's the same thing they ran into with the software locked batteries. The goal is to have minimal overhead to try to drive costs down on Model 3, and making the SR just a software limited version of SR+ actually ends up pushing the overall cost down, rather than having to set up an entirely different line that does the cloth seats and the aluminum roof.
Think about it from a supply/configuration/logistics point of view. It might be cheaper to make 500k cars with 100% heated seats rather than 250k w/ heated and 250k w/o heated. Even if BOM cost is lower, the additional complexity due to the logistics and supply chain probably means it's more expensive to have both variants in production. That way people who want heated seats are subsidising the lower overall cost for those who don't. It also allows the next owner of the car to pay for that option at a later date if they want to. Almost every industry does this (there was a good case about oscilloscopes a while back, where the cheaper, lower bandwidth models just had a low-pass filter circuit installed).
It makes sense economically, but still feels paradoxical. A similar example would be building houses with 3 bedrooms, where 2 bedrooms remain locked and inaccessible unless you pay for them. So a significant proportion of the population ends up living in 1-bed houses even though there is no scarcity of resource.
Personally speaking I would not want to share a road with an autonomous car that had been tinkered on by some random Joe. Autonomous farm vehicles or something, sure, but public roads are way too risky. This is not the desktop environment of your Linux laptop, it’s monstrously complex software controlling thousands of pounds of metal and batteries.
See, I don't agree at all. Just like all cars have to pass a pretty strict annual inspection to be on the road, I imagine autonomous cars will have to go through regular assessment too.
More specifically, I'm bothered by the fact that if you get into a crash in a Tesla, Tesla can disable your car and stop it from activating unless you do repairs at their approved dealership. You can't just buy parts from a scrapyard and get it running again, Tesla is the gatekeeper and they don't let anyone else hold the keys(they don't even release service manuals unless absolutely required by law). In contrast I could buy a brand new Mercedes CLS with its very respectable smart cruise control and Mercedes can't do anything to stop me from replacing the engine, the head unit, or fixing the whole thing if it's totalled - they just can't. It's not a freedom I'm willing to lose by buying a Tesla.
I would expect autonomous cars to be have doubled down security against tinkering although I personally don't like that. The reason is purely public safety. If someone could re-program autonomous cars to slam into public crowds and places, imagine the havoc it can cause. A malicious actors can just walk in to parking garage, put in some wires and reprogram every vehicle they can get their hands on - imagine those scenarios.
You can already do that with existing cars(pretty much anything that has power steering, so last 30 years of cars) and yet you don't see people hijacking vehicles left and right this way(or any other way, you can cut someone's brakes in 5 seconds and yet it's extremely rare). I think this is a made up problem.
It's not about usual hijacking or rigging. It's about telling car to speed and slam into pedestrians if it sees large enough group of them, for example, and until that time it can just behave normally. Think of terrorists getting access to programmable robot that weighs over a ton, that can accelerate to 60 miles/hr under 5 seconds and it can identify objects in its surrounding.
So.....a new Mercedes with driver assist package then? And Mercedes will not stop you from modding their cars to your heart's content, you don't have to ask them for permission to work on your car like you have to with a Tesla.
What kind of event is this and who is it targeted to? Right now there’s a comparison of energy consumption for different instructions on processors, I’m very confused.
The technical angle seems like a very smart one for this talk. You can just smell the cowering insecurity of the analysts and their deference to Musk when the technical descriptions whoosh past them. They may be (rightly) skeptical, but they don't want to look stupid while being so.
...out of all the things I expected to see from a Tesla stream, I definitely did not expect a 25 minute discussion on the cost of 32bit additions, dot products, sram bandwidth, and chip design, at the level of a third year college hardware course.
If you’ve listened to their earnings calls this is par for the course. Say what you will about his tweets, but I find these discussions fantastically informative and transparent.
"you say you can put bananas in your smoothie. just a thought, but perhaps, might I ask, do you have the capability to put strawberries as well, or are we not there yet?"
That was my first instinct to that question as well, but I think he was just asking if ReLU was required by the hardware design - or if it was possible to use other activation functions as well. If the the ReLU was part of the hardware itself somehow, then it wouldn't be possible to use tanh or sigmoid (which may be better in certain situations); so I think he was just asking if ReLU was required, or if there was flexibility allowed in the activation function.
I'm not a NN expert, but based on what I have found, the point of using RELU instead of other activation functions is what is called the problem of "vanishing gradients".
Basically (IIRC), during backprop the error difference gets ever smaller the further back in layers you go, ultimately getting "lost in the noise", making learning in the earlier layers more difficult to impossible.
I'm not saying RELU is the only option to make this work, or that it's the only activation function that provides a "fix" for the issue; I'm sure there are other ways to deal with vanishing gradients that I don't know about.
I also lack the mathematical knowledge as to why RELU helps in this manner, but I suspect something having to do with the lack of "asymptotic structure" approaching the extremes (I don't know what the proper term would be). Or maybe it allows for some form of "forgetting", in the prevention of multiplying very small numbers (such values just go to zero ultimately)?
Maybe someone else here with the knowledge can explain it better, and we can both learn...?
ReLu is useful, but the question as asked was more about other (also essential) functions for other steps in RNNs -- in particular, sigmoid "squash" functions, as well as MaxPooling.
Which might explain why he's a crummy CEO. If your job is to strategize for the entire company and you're busy learning the minutiae that you spend a lot of money for other people to worry about, maybe you need to consider using your time a little better. If you want to build neural nets, keep your stock, fire yourself and go build neural nets.
But then you have the major car company problem and the phone companies before steve jobs. You have people who don't know what's possible, don't know who are the best people to hire, and can guide a vision. if you know capacitive touch screens are possible no one has ever tried it makes it easier to implement. Same with google and how their CEO are all engineers. So pitching the CEO on AI project becomes easy
> If your job is to strategize for the entire company and you're busy learning the minutiae that you spend a lot of money for other people to worry about, maybe you need to consider using your time a little better
This is what everyone that is not Tesla or SpaceX are doing. And have been doing for a long time. If the CEO is not an engineer at heart, what are they?
I seriously doubt Elon Musk has more engineering knowledge than the people he hires on their specific fields. However, he can make pretty well informed strategic decisions if he knows WTF the engineers are talking about without taking their word – not even that, as explanations have to be dumbed down.
This is not a new thing. Bill Gates was like that (1). Steve Jobs was no dummy and had an engineering background, but not at the same level – he did parter with a genius engineer, however.
I think Musk is doing the right thing.
> Which might explain why he's a crummy CEO
That's quite debatable, I'd say. Isn't he getting results?
> This is what everyone that is not Tesla or SpaceX are doing.
I think it's very important to point out here that Elon does not run SpaceX. It's well known that he's a front man for that company, and that all day to day operations are run by an actual executive who does the job of being an executive. This is why SpaceX is doing much better than Tesla.
SpaceX isn't public or open about their financials, but there's a lot of evidence that they are simply losing money. Ex. they were raising a lot of money last year, profitable companies don't do that. Also read https://www.theverge.com/2018/12/7/18129539/spacex-falcon-9-...
Yeah that article seems like a lot of tea leaf reading to me. All evidence points to SpaceX building orbital rockets cheaper than just about anyone else before reuse, but even if not reuse should drop their costs well below anyone else until blue origin comes online or until ULA decides to shake off the crust of old age. It's not at all surprising that investors are told not to expect to take a profit for 15 years because SpaceX is focused on growth and development. I wouldn't expect them to reinvest any less than 100% of their profits at this time. It makes just as much sense to assume they raised money to keep development moving at the pace they want which wasn't fast enough using profits alone.
Despite the fact that I own a Model 3 and own shares in Tesla I often find the critics make good points and just hope for the best anyways because I want electric cars to take over and I want self driving cars to develop sooner than later because driving is dangerous and tedious.
The SpaceX hate seems much less justified to me. SpaceX has been building launching, landing, and reusing rockets for years now. They've ushered in a new lower cost space age.
Actually being aware of the technical details of your core business is a very useful skill to have as a CEO. But it doesn't seem to actually result in any benefits with Tesla. He probably knows the limitations of neural networks but that doesn't stop him from using the tried and true startup strategy of promising the world and delivering very late. Building a car that can stop on red traffic lights is something they should have had by 2016 already. But it's 2019 and their self driving technology they have demonstrated in their demo blatantly ignores the speed limits on the highway by roughly 10mph.
On one hand there are hyper-bulls who claim Tesla is a $4000 stock and the future of transportation. On the other, hyper-bears claim the equity should trade around $0-$10. There seems to be no middle ground.
It seems like they are almost betting the company on FSD. I don’t think FSD is really even close to a possibility over the next 5-10yrs. I hope I’m wrong, but if I’m right, I don’t see how Tesla keeps going on like this.
> It seems like they are almost betting the company on FSD.
IMO they don't have any choice. The more and longer they operate like a car company shipping bigger and bigger volumes, the more their financials will be undeniably trend towards those of the existing car companies and the more their existing market valuation will be hard to justify.
They need something like this to not get traded at traditional car industry multiples.
If you just measure based on Tesla's last two quarters, their P/E is just 28. That's high for a car company but pretty low for tech. Self-driving is not necessary to maintain that valuation, only continued growth of the electric car market in general.
> That's high for a car company but pretty low for tech.
Yes! Totally agree.
> Self-driving is not necessary to maintain that valuation, only continued growth of the electric car market in general.
Disagree. The key properties of the car business are high capital costs and high variable costs and not huge margins. The key property of the tech business is low variable cost and often low capital costs (but not always) and high margins.
There's nothing "tech" about building electric cars vs. normal cars and 10 years from now the margins and capital expenses of electric car business will be like the existing ICE car business.
So this is Tesla saying: "Yeah, our financials are starting to look like a normal car company, but we've got this thing that you should keep value us even more like a tech company than you do today."
The market values Tesla like a car business that is growing ~50% per annum.
> 10 years from now the margins and capital expenses of electric car business will be like the existing ICE car business.
Tesla already has margins similar to existing car companies.
> So this is Tesla saying: "Yeah, our financials are starting to look like a normal car company, but we've got this thing that you should keep value us even more like a tech company than you do today."
Yes, they're saying that, but the market is obviously not buying it.
> On one hand there are hyper-bulls who claim Tesla is a $4000 stock and the future of transportation. On the other, hyper-bears claim the equity should trade around $0-$10. There seems to be no middle ground.
>On one hand there are hyper-bulls who claim Tesla is a $4000 stock and the future of transportation. On the other, hyper-bears claim the equity should trade around $0-$10. There seems to be no middle ground.
Well the middle ground is the actual stock market where Tesla is trading around $265.
I think FSD would be cherry on the top. If they can continue slashing down the prices and improve battery life it would be revolution in itself. Even just continuing with current pace if they can produce $18K electric car with 600 miles range in next 3-5 years, you could be golden as shareholder. Their advantage, like Apple, is attention to details and awesome design that other car manufacturers have failed to replicate.
What if they proved it by improving hardware from self-driving hardware 1 to 2. If they proved that and proved it made their NN better then it makes sense. Of course, it wouldn't to you cause you have no inside knowledge but it may make perfect sense to them based on their data.
So far, half an hour from the head of IC design. Nice special purpose IC and board. Dual everything for redundancy. Wide special purpose neural net evaluation. 100 watts for the compute system. Code signing. Shipping in new Model 3 cars since last 10 days.
"All we need to do is improve the software" - Musk [12:07 PDT]
LIDAR "unnecessary" - Musk [12:13 PDT]
Computer vision guy is now speaking.
Recognizes "driveable space", not just obstacles. Video shown, but just for a freeway. This is crucial to safety. Need to see this is a cluttered environment.
> "All we need to do is improve the software" - Musk [12:07 PDT]
That's an old claim, see this 2016 press release [1]:
> as of today, all Tesla vehicles produced in our factory – including Model 3 – will have the hardware needed for full self-driving capability at a safety level substantially greater than that of a human driver.
At this point I believe the claim once I see the self-driving functionality having been rolled out to the public and the accident number been reduced.
it snowed a few weekends ago in Chicago, my autopilot turned off because snow covered up the cameras. So I am not buying all this "self driving with no lidar" brouhaha
Depends on the unit. On some designs (e.g. Velodyne HDL-64) the unit is split into two sections. The whole upper section, along with all the optics and some of the electronics, spins at 10-15 Hz, which naturally tends to shed precipitation and other crud.
Sensors with non-moving optical windows can use other strategies, such as air knives or wipers.
Doesn't LIDAR has similar problems with precipitation? Regardless of what solution proves to be the best, there is going to be a decent amount of time between when a self driving car can handle most scenarios and when it can handle all possible scenarios.
For someone who lives in or close to that white part, "most scenarios" includes snow. Arguably for Level 4 autonomous cars in those areas, you either need to fully disable autonomy in September and enable it again in late April, or you'll need to handle snow.
You are comparing two different things. Old snow on the ground (your images) is not the same thing as active or recent snow accumulation on the vehicle (the original comment). It is completely reasonable for the car to disable autonomy due to snow accumulation until someone clears it off the vehicle. That says nothing about the autonomy of the vehicle with snow cover on the ground.
I'm assuming Level 4 is what Musk means by Full Self Driving. The bar to be passed is then
"No driver attention is ever required for safety (...) self-driving is only supported in limited circumstances (e.g. geofencing), and when these circumstances are no longer met the vehicle must be able to safely abort the trip, e.g. park the car, if the driver does not retake control."
How would that work if you're out driving on the highway, and it starts snowing hard so the car can't see anything? Just park on the highway?
If your car can't safely handle such a scenario, the automous feature would have to be "season-fenced" in addition to geofenced.
It isn’t like it all cameras simultaneously go from 0% obstruction to 100% obstruction instantaneously. The car should pull off the highway and park or at least pull to the shoulder and park once it has identified decreased visibility to an extent that might impact safety.
It kinda is though. Snow deposition is a function of surface temperature, which is uniform across the sensor. Hit the wrong initial temp when you enter a blizzard, and deposition takes your vision in seconds. It can be hard enough to see out the windshield which has heating, wipers and anti-freeze wiper fluid fighting for it.
Do you realize how far you have moved the goalposts during this conversation? You started out with the suggestion that they "fully disable autonomy in September and enable it again in late April" and now you are talking about situations with "the wrong initial temp when you enter a blizzard".
I will simply refer back to my initial comment in this thread: "there is going to be a decent amount of time between when a self driving car can handle most scenarios and when it can handle all possible scenarios". Blizzards are not in the "most scenarios" group. Barring emergencies, no one should be driving in a blizzard let alone an autonomous car.
I don't think OP has moved the goalposts at all. LIDAR can work in snow but cameras get covered. And yes you are right, there will be a decent time between testing in some basic scenarios and testing in all scenarios. But I do not understand why Mr. Musk is trying to reinvent the wheel and rely only on vision when LIDAR has already shown that it works better.
> But I do not understand why Mr. Musk is trying to reinvent the wheel and rely only on vision when LIDAR has already shown that it works better.
Because he's sold a bunch of cars (and a bunch of stock in the company selling the cars) without LIDAR with the explicit claim they are hardware-ready for full self-driving.
Freeway shoulders are not safe places to stop. You're likely to get hit by a drunk or distracted driver. Ask any police officer or tow truck driver. That's why stopping on the shoulder is only for emergencies.
Water affects certain wavelengths, and there's been good progress on LIDAR that uses wavelengths that aren't effected.
Most approaches want to use a wavelength that has as little ambient light as possible. That light is absent precisely because it is absorbed by water vapor in the atmosphere. The alternative approach must deal with lower signal:noise, but is not substantially affected by weather.
Yes, we implemented simple physical solutions such as wipers and window heaters in cars to maintain vision for drivers. What we didn't do was implement lidar to aid human drivers.
Simple physical solutions are available if necessary for the cameras too. I don't have a tesla so I don't know what measures they use to keep the cameras clean but I'm sure they've thought about it and designed for it.
Great-grandparent comment suggests that they did not, in fact, design for it. Assuming the best is nice when you have no information, but we have information.
I'm not assuming the best, I'm assuming basics. We've just watched a presentation where Tesla put immense resources and talent into designing a state of the art vision processing chip. Do you think they're going to let it be rendered useless by not implementing simple physical measures to keep vision coming in from the cameras? (What measures are already there plus easy opportunity for improvement). Of course not. Such a line of argumentation is transparently bogus. I don't know why the top level commenter's autopilot turned off in the snow but trying to argue that that shows that snow renders cameras useless and makes lidar necessary is stupid. Sorry for the strong language but really, I don't know how else to put it.
I have seen recent Teslas in person in the showrooms, and there doesn't appear to be any wiper for the cameras. Especially with a firsthand owner account just a few comments up I think it's actually only logical at this point to assume they didn't put wipers on the cameras.
Well they didn't do that either. They didn't do anything for it. But don't worry, FSD is just around the corner and the cars have all the hardware necessary, it's just a software problem now. And this time it's different than in 2016 when they said they had all the hardware they needed. And it's different from 2015 when they said the FSD was coming in 2016.
In our 2005 DARPA Grand Challenge vehicle, we had a system to clean the camera and LIDAR. It used a spray nozzle, and alternated spraying windshield cleaner and air. This is a commercial truck accessory used on mining trucks.
This talk of a "cut-in" detector is scaring me. It's like they lack any sort of higher level planning and decision making (which notoriously is not a neural network of any sort).
the narrative in this conference seems, minus the exaggeration, to follow exactly the flow for the presentations given by the MobilEye CTO for the last six years.
The thing I don't get about Tesla is why they keep heaping more on their plate rather than just (for large values of just) re-doing the motive power bit. That alone is a massive challenge and it is not something they've licked at economies of scale that allow a $10K vehicle to be brought to market, which is what it will take to really make this a success.
Every extra they tack on to the base product is one they will be expected to deliver on as well diverting attention from the main problem they are dealing with, which may in the longer term leave open enough room for the competition to wiggle through.
Sure, autonomy is a big deal, but it is also something that will once cracked be an instant commodity and there is a lot of money and talent focused on that particular problem, which is surprisingly hard to do well.
If Tesla ends up going under because of one of the side shows (Solar City, autonomy, Power Wall etc) that would be a serious loss.
I think it's related to who Musk is, and the kinds of things he wants to achieve.
SpaceX isn't trying to just do what has always been done and just make it 5 or 10% cheaper. That's boring. SpaceX is completely turning the launch industry on it's head.
Tesla is the same. Elon isn't interested in making "just" an electric car. He wants to change the world, and electric cars that drive themselves will do that.
Can he do it soon? I'm not an expert, I don't know. But damn it's exciting to watch.
The purpose of Tesla is to save the world from global warming, though. I'm not sure self-driving is that relevant compared to the electric propulsion part.
“diverting attention from the main problem they are dealing with, which may in the longer term leave open enough room for the competition to wiggle through.”
Perhaps this is the intent rather than a side-effect.
Indeed, Musk has encouraged competitors from the start. I believe they open sourced a lot of their patents in an effort to encourage competition in the electric vehicle space.
Apple's work on developing their own SOC (A13) have paid dividends for them. The same legendary chip designer (Jim Keller) designed the new Tesla chip after working for Apple.
For vertical integration to work you need to have funds to support it long term. Apple has it. Tesla - not that much (at least not in the current state)
Building one-off solution that will work good for your special use-case is hard, but in a doable territory. But cutting off fat that you don't need for your use-case is a one-off gain.
Building whole process that will keep on producing solutions that are better than what suppliers can offer, especially in highly competitive area, is way way harder. That's where the real money are spent.
From the video presentation, it did not seem to be a "one-off" effort. They mentioned that V2 was already 50% complete and will continue to be developed over time to avoid complacency and stay competitive while also keeping the cost at a manageable level.
I get what you're saying though. But if they manage to pull off the Robotaxi fleet, then $$$ won't be much of an issue.
How would power wall lead to them going under? They would just be selling excess batteries, which seems to be one of their advantages over competitors. Autonomy, I'm ambivalent on it, they can charge customers for it right now and many will pay.
If anything, the solar city/roof shingles side-show has seemed the worst to me personally, alongside the falcon-wing doors and the self-infliction of twitter harm by the CEO.
> Sure, autonomy is a big deal, but it is also something that will once cracked be an instant commodity
Not if the company that cracks it eats its competitors or, assuming it's like waymo, commoditizes car manufacturers into suppliers for autonomous taxi fleets.
First of all, a lot of these things are orthogonal. With that I mean, that they are independant from each other. The self-driving effort interacted only minimally with getting the Model 3 production up and running. As long as the money does not run out, it makes sense to do things in parallel, which are parallizeable. On the other side, Musk has a big picture in mind, which stretches quite a bit into the future. So these "side shows" actually fit together. You use solar cells to produce electricity, which you store in your power wall which can power your electric car.
If Musks ideas about self-driving actually become true, he holds the holy grail of the car industry. On one side, self-driving abilities become crucial in future car sales. On the other side, if he can make the Tesla car sharing network happen, the revenue of that could be a magnitude or more larger than the one of car selling. This would justify some of the crazier valuations of Tesla stock.
I wonder if he's going to explain this position at all? Or if it's just posturing against Waymo et al. Seems like more data is always better than less data for this application?
Because lidar is expensive, error-prone, and can't be created without flaws. So if you make a car LIdar will fail way before camera. The only problem is if camera only can work
Why not both? Infact, Waymo has Lidar, radar and cameras (Tesla have just the last 2 and are working overtime to demonize the first). More data (and more redundancy) is better - assuming you have the computing power to process all the sensor input. As far as I can tell, the rubbishing of Lidar is FUD.
Because if you're using lidar for depth and you try to mass produce the there a high chance the lidar will be broken and because it accurate people will assume it more correctly. Eyes and camera may sometime make mistake with distance but usually it will be software not hardware. Lidar opens the issue for both. So it's easier to use the camera and just try to make a camera as good as lidar.
All sensors and the systems they operate with have flaws. There is no way around this - it's just the real world. Even cameras will have flaws, or they can develop them over time (lens degradation, image sensor dropping pixels or gaining stuck pixels, etc).
That is point behind what is known as "probabilistic robotics" - the real world has noise, and you need to be able to deal with it. Don't expect perfect sensors, don't expect a perfect environment.
Self-driving vehicles are the application of probabilistic robotic principles to a real-world task - quite possibly one of the most difficult tasks for the field. To be quite honest, it's amazing how well it's worked in such a short period of time.
I think cameras and machine vision approaches will be needed for self-driving vehicles to be fully successful, but I wouldn't say that LIDAR should be counted out. It will probably be necessary - even required - to have all of the sensors currently being used, not just a subset.
One kind of sensor that hasn't been explored much that I think will be needed is some kind of audio input; some kind of wide-range microphone (probably binaural or stereo) to take in environmental audio and use that for driving cues. Some simple examples might be for vehicle horns, or the squeal of tires, or the revving up of an engine indicating someone is speeding up aggressively, or an emergency siren, etc.
There may even be other sensors needed or that could provide other data to fill in certain gaps of environment knowledge to help a self-driving vehicle navigate. I don't think any one or another should be discounted.
They explain it in the software talk. In short, I think his thesis is:
Lidar is expensive, and gives surprisingly limited data (just points of depth) and is like a crutch that will only get so far. It will not get you to full autonomy so why not spend the time and money on vision which will (as proven by humans). They also do use radar and ultrasonics (not visual spectrum).
At some point your machine must be good enough via machine learning to give a very good impression of understanding intention in other road users, pedestrians etc. One example they gave was a distracted pedestrian with a phone. Lidar tells you nothing save an obstacle is on the pavement, it won’t tell you they might step out without looking, but machine learning on a massive dataset can.
One thing LIDAR can do that Tesla's ML vision software cannot is identify freeway dividers and trucks turning in front of the car.
Once Tesla can actually handle those two extremely basic tasks maybe they can start talking about how vision is better than LIDAR. Until then, it's just more hot air, and that's not even taking into account Tesla's claims of using ML to predict the actions of uncontrolled independent agents in the field of view.
Yeah it's nonsense. It's an after the fact reframing of a decision that was made for other reasons. Lidar + camera is better for this kind of problem. It's because current lidar are too expensive and too fragile (1-2year life) to put in a production car, not because they aren't much better. If this wasn't the case he'd definitely be using them. Elon cannot use a lidar in a Tesla even if he wants too so he's coming up with some FUD to dismiss any completely reasonable questions in advance.
To summarize, it appears that their new chips can do 96x96 dot product in a single cycle (multiplying neural network weights by values), and have hardware for activations (ReLU, softmax, tanh) and pooling (as in convolutional neural networks). This results in a crazy 144 trillions ops/second.
How does this compare to TPUs and the Neural Engine in iPhone CPUs?
Can they actually do a 96x96 dot product in a single cycle? My understanding was that they can do 1 96x96 dot product per cycle - that is to say that their throughput is a 96x96 dot product but it'll be heavily pipelined.
Society has decided how to judge whether I am safe to drive, and I have passed the tests.
I worry that the laws on the books are insufficient to judge whether a computer is safe to drive. If Tesla already has plans to increase safety, I'd like to know how they judge it, where they think they are, where they'd like to improve, etc.
I think there is a lot of scope for an autonomous vehicle to be safer than a human driver. Consider:
1) computer has faster reaction times.
2) 360-degree awareness
3) better awareness of vehicle performance.
Imagine a scenario where you are at speed and get a tire blowout. The car can sense it instantly through tire pressure sensor, and counter the swerve appropriately - something that humans are very poorly prepared or practiced for.
Alternatively, a driver in a neighboring lane makes an unsafe lane change into your lane. The car sees him, honks the horn, moves/brakes/accelerates to avoid the other car, before you've even seen whats happening.
I don't see this line of reasoning. To me "safer than humans" or "in the 90th percentile among drivers" is a fair go to market line.
Optimally safe should still be the goal. That is, no human could have provided alternative input to the computer that would have created a safer outcome.
No, it just means the next-gen design had safety as a primary design objective. Secondary objectives could be additional modes of self driving like traffic light detection and managing intersections.
Volvo leadership states the exact same thing about their car designs. The primary motivation “in everything they do” is safety.
The statement can be made entirely distinct from the current safety level.
They are fundamentally different: to summarize, Waymo follows the HD maps approach when you have a very precise map of a given city, with both metric (actual 3D shape of the environment) and semantic (lanes, sidewalks, signs) information, in which you localize yourself with centimeter-level accuracy (with GPS, SLAM etc). When you're localized you can combine information that you see now such as cars around you and high quality information coming from the map.
Tesla, on the other hand, thinks this is too fragile to changing environments and works with regular (think google maps level of detail, probably a little better) maps, combined with local semantic information such as signs and lane markings. This also means that they put less pressure on localization, because they don't use the map to detect i.e. speed limit (I'm pretty sure Waymo can detect those signs, it's just that the existence and position of the sign is already known in a global world frame).
Feels like they're just throwing a bunch of hardware specs at investors to distract from any discussions about the current state of the software beyond "just need to improve it now".
That's fine. I don't care who's driving the car carrying the camera. I care about what the camera can see. If you've ever been sent an Amber alert for a local child abduction it often comes with a description of the car that the suspect is driving. If I'm in law enforcement I'd much rather get an emergency warrant to ask Tesla to query the network to find that kind of car in a given geographic area.
The future is here, and computers will see everything in excruciating detail via multiple planes of information. The power you consume, the data you send, the money you spend, the media you watch, the words you write, where you go.
It’s all out there in the world. There’s no taking it back, though we may be able to legislate about its use.
You can't say "You can query for anything!" and have it be anonymized.
In fact, if you're taking a camera shot, that's the most broadband sensor imaginable, and no thanks, I don't trust any company with profit motive not to crumble and pop out a software update to selectively disable picture blurrers, or to not have a Boolean flag in your data scrubber.
As a company, making that statement is just trying to slip one past and hope nobody asks inconvenient questions about it.
It's trivial to identify the car. You can see the reflections of the plates at times. If that was removed, you'd still have the ability to cross-check it with traffic cameras (or business cameras).
- First principles hardware design of focused self driving computer (many times better than any competing existing hardware). Already shipping in all newly produced cars. Currently working on next gen that will be 3x better to ship in a couple years.
- Lidar is an unnecessary mistake that competitors are making that won't succeed (too expensive, need too many, unnecessary).
- Real world fleet testing is critical to success, simulations are not good enough since there are too many unknown unknowns in the real world. Tesla uses simulations too, but nobody else comes close on real world fleet testing.
> Lidar is an unnecessary mistake that competitors are making
This is very controversial and rest of the industry thinks exact opposite, many even claiming that Tesla is being irresponsible and even delusional in trying to do autonomy without lidar. The main points I have heard in favor of lidar are that computer vision is very flaky not only in suboptimal weather but even in good weather. Cameras are simply no where close to in performance in dynamic range, rapid adaptation, focusing etc as human eye. Imagine car going under the shaded road with rapid changes between bright light and shade. The likelihood that your depth estimation will get messed up is very high. Of course, night driving becomes highly questionable as well. In additional the long range depth estimation is very flaky with stereo vision right now and a topic of research for mono-vision. If you want to retreat to level-4 only and that too with conservative speed, weather etc then may be vision+radar more doable?
They pair it with a forward facing radar which is inexpensive and good at depth perception.
Elon predicts all competitors will eventually drop lidar. He mentions it's expensive, but also not as good in a lot of cases (and all roads/signs are designed for vision).
He argues that getting vision to work is a prerequisite for getting self-driving to work and once you have it working, lidar is worthless (and unnecessary).
Elon Musk is noted for his premature optimization.
Case in point: both the Gigafactory (vastly overbuilt for the quantity of batteries actually produced) and the Alien Dreadnought (vastly overbuilt for the number of cars Tesla current produces...assuming that Tesla is ever able to get the fancy automation working). Boring Co digging a two-mile tunnel in West LA without bothering to learn how to pour concrete smoothly, or to make the "rails" the proper width, or learning about ventilation, or access points....
It's not easier to optimize a working system for consumers if "optimization" == "not dying". What will happen is you will deoptimize your brand's safety and find yourself regulated.
Often reverse is true. Trying to optimize system with million lines of code may require serious change in architecture and much harder because of backward compatibility and all the legacy baggage. As many people would describe it, it much harder to fix an airplane that must also continue to stay in the air.
“In my view, it’s a crutch that will drive companies to a local maximum that they will find very hard to get out of,” Musk said. He added, “Perhaps I am wrong, and I will look like a fool. But I am quite certain that I am not.”
Their forward facing radar has failed to avoid crashes into big, stationary objects (at least 2 semi trailers, a fire truck, and more). The explanation I got was that radar is rather noisy, so it's handy to filter-out stationary objects (like road signs, broken down vehicles on the shoulder, and fatally, any any semi trailer crossing your lane)
I don't think anyone wants to use Lidar - the claim is that they need it to achieve full, safe, autonomy. If it turns out that it isn't needed than Tesla will be in a great position.
I think it also fundamentally depends on the business model being pursued:
- If you are trying to sell the Model 3 directly to the end consumer with autonomous mode, the extra $10K for Lidar and bulk (which will greatly effect the exterior design) are definitely non-starters.
- If you are going to robo-taxi route (such as Waymo and Uber), the extra one time $10K cost to add lidar for the 5 year life of the car is probably a blip on the income statement of the operator as compared to a full time human driver which probably costs $10K PER MONTH for the lifetime of the service. For the robo-taxi business model - its a bit of a no brainer - they could stick every sensor known to man on the car and still make out like a warlord by getting rid of the human driver but maintained a 100% safety record. Plus making the car stand-out with a unique Lidar inclusive shape is a great marketing differentiator. Also reduces your liability if a taxi rider ever sues since you can claim you have redundancy in the system.
Eh, I'm pretty fine with that. Imagine the current gen slightly beats human drivers - also known as "revolutionary, but not perfect" - of course the next gen would focus on improving safety further.
Not that I'm confident about either side of this, but the TOPS isn't the only important factor here. It's also the efficiency. If Tesla's CPU can do more ops/watt, then it's more ideal for the conditions that Tesla needs the CPU to fulfill. On top of that, I have no idea what "320 TOPS" means in the context of the specific workload Tesla has. Will they _actually_ get 320 TOPS?
> Currently working on next gen that will be 3x better to ship in a couple years.
That's ... not great. If that was CPU type of unit, it'd be great. But TPU-type accelerators are growing at massive speed (as it's still pretty new and simple tech), where you're more looking for 10x type of gains.
549 comments
[ 2.9 ms ] story [ 396 ms ] thread"To me right now, this seems 'game, set, and match,'" Musk said. "I could be wrong, but it appears to be the case that Tesla is vastly ahead of everyone."
I am eager to see what they unveil today.
https://www.youtube.com/watch?v=dEv99vxKjVI
https://ark-invest.com/research/podcast/elon-musk-podcast
It was mind boggling. I am hoping Tesla can provide some specifics today because it seems Elon is living in a fantasy world (albeit one I'd like to live in if we can actually get safe self-driving cars).
also I can't reconcile how the new hardware is this huge leap ahead beyond raw computing power if, by Tesla own claims, previous hardware was perfectly capable of autonomous driving.
seems people were getting fooled either now or before.
https://www.google.com/url?sa=i&source=web&cd=&ved=2ahUKEwiR...
And Elon has a long history of making false claims about Tesla’s progress. For example in 2015 and 2016 he claimed that Teslas would be fully self-driving by 2018.
So why shouldn’t we be skeptical?
https://arstechnica.com/cars/2019/03/teslas-self-driving-str...
Your "third party research" is obviously bullshit, they go as far as including Apple in their ranking.
This right here is just typical worthless marketing press release spam from a management consultancy firm.
No one knows if Tesla strategy will work because they don't have the data collection in place.
They have no way to store or transmit the massive data you are describing off the platform do they?
My understanding is that they have very limited storage and transmit onboard.
I mean saying a phone can upload videos to youtube but a can can't to tesla is a weird ledge to stand on. Even their windshield wipers work based on sending video data to tesla to be learned on.
Your link:
> According to Electrek, Tesla trails behind other companies in terms of autonomous driving tech based on a list created by Navigant Research, an independent research firm.
Electrek’s article:
> Electrek’s Take
> I think Navigant’s autonomous leaderboard is ridiculous. There are way too many brands that keep most of their development under wraps, which makes it hard to evaluate them and therefore, it gives very little value to a leaderboard like this in my opinion.
Electrek is not exactly unbiased. It's literally called EV and Tesla news. Fanboi site's opinion should probably be taken with a grain of salt.
Here's the Navigant executive summary directly: https://www.navigantresearch.com/reports/navigant-research-l...
The Waymo end-game that I heard was "able to go through a drive-thru". I highly doubt Tesla is anywhere near that point.
(1) https://m.heise.de/autos/artikel/Test-Tesla-Model-3-4400919....
Basically, most other manufacturers like Opel, Audi, Mercedes, Hyundai, VW, Volvo, Ford, etc. has had for several years the feature to detect speed limit from computer vision recognizing the road signs. And it works reliably, as is pointed out in your link.
How can Tesla be a leader in using computer vision for cars, but not be able to read the road signs?
The kind of drive-thru that Tesla is currently associated with involves semis rather than fast food and it would be really nice to hear that they've at least licked that particular bug (and for good, this time).
Your point it very astute.
Among a few other ML/AI MOOCs, I completed Udacity's "Self-Driving Car Engineer" nanodegree - so when I'm out driving, I often come upon situations where I wonder "how would a self-driving car navigate this?"
Today, driving in to work (note: USA), I noticed one intersection I've been through many times before, and that question came to mind. The intersection is interesting, because on approaching it, the road curves to the right, and you can actually see one of the traffic lights on the left before you even see the intersection. By the time you see the intersection, you're already on top of it.
So as you round the curve, you see the lone traffic signal (red/yellow/green); if it is red, do you start to brake, or do you wait until you can "see" more traffic signals? If you wait - will you have time to slow down and/or stop? ...and so forth.
This and others are all kind of "edge cases" that will need to be trained on, and/or perhaps other cues for self-driving vehicles installed or set up so the vehicles can navigate such areas successfully. I know when I first went through the intersection it was a bit of a surprise; it's not a very safe intersection (going home in the opposite direction is not any better; in that direction, you're headed downhill, have to cross the intersection, and immediately start turning to the left after going through - the curve is really abrupt, and you have protected/unprotected left-hand turns both directions, etc).
Other than for accidents, the SEC investigation, etc.
I can imagine that police could mine this just like they're doing with Google geolocation data.
Except that it isn't, and even Karpathy said the quantity doesn't matter, it's the data quality.
Yes, two of the autonomous vehicle leaders are not using any data whatsoever.
I thought this was the smartest forum on the internet?
Notice the cherry-picked examples in the presentation. There is a whole class of problems the field cars can never help with, since they lack the dead-reckoning sensor setup and precise odometry a development car would have.
Can you give an example? I'm curious what kind of triggers strictly require lab-calibrated hardware.
If you call something "autopilot" and promote it as if it will drive for you and then it ends up killing dozens of people ... that's where successful class actions come from.
(I obviously don't want people to die either.)
If you think this is a new thing, you haven't followed the car industry much.
Just as an example, my 2002 el cheapo Peugot was bought without the option to show instantaneous MPG/average MPG/range/etc. Spend two hours soldering/gluing on the missing $2 toggle switch, ask a friend with the Peugeot Planet update tool to enable it in software, and Bob's your auntie.
It goes much further than that though, especially when you get into chip tuning. Software upgrades that add 50 horsepower to your engine output are commonplace.
But as for the "actively hostile to 3rd party repairs/mods", this part scares me to.
It's the same thing they ran into with the software locked batteries. The goal is to have minimal overhead to try to drive costs down on Model 3, and making the SR just a software limited version of SR+ actually ends up pushing the overall cost down, rather than having to set up an entirely different line that does the cloth seats and the aluminum roof.
More specifically, I'm bothered by the fact that if you get into a crash in a Tesla, Tesla can disable your car and stop it from activating unless you do repairs at their approved dealership. You can't just buy parts from a scrapyard and get it running again, Tesla is the gatekeeper and they don't let anyone else hold the keys(they don't even release service manuals unless absolutely required by law). In contrast I could buy a brand new Mercedes CLS with its very respectable smart cruise control and Mercedes can't do anything to stop me from replacing the engine, the head unit, or fixing the whole thing if it's totalled - they just can't. It's not a freedom I'm willing to lose by buying a Tesla.
I knew I'd find someone here who hates when John Deere locks down their vehicles, but is fine when Elon Musk does.
How about 2 thus far and Tesla was not at fault.
this is wonderful
I don't know the context here but it feels like they're trying to ML-splain Karpathy?
"you say you can put bananas in your smoothie. just a thought, but perhaps, might I ask, do you have the capability to put strawberries as well, or are we not there yet?"
Basically (IIRC), during backprop the error difference gets ever smaller the further back in layers you go, ultimately getting "lost in the noise", making learning in the earlier layers more difficult to impossible.
I'm not saying RELU is the only option to make this work, or that it's the only activation function that provides a "fix" for the issue; I'm sure there are other ways to deal with vanishing gradients that I don't know about.
I also lack the mathematical knowledge as to why RELU helps in this manner, but I suspect something having to do with the lack of "asymptotic structure" approaching the extremes (I don't know what the proper term would be). Or maybe it allows for some form of "forgetting", in the prevention of multiplying very small numbers (such values just go to zero ultimately)?
Maybe someone else here with the knowledge can explain it better, and we can both learn...?
This is what everyone that is not Tesla or SpaceX are doing. And have been doing for a long time. If the CEO is not an engineer at heart, what are they?
I seriously doubt Elon Musk has more engineering knowledge than the people he hires on their specific fields. However, he can make pretty well informed strategic decisions if he knows WTF the engineers are talking about without taking their word – not even that, as explanations have to be dumbed down.
This is not a new thing. Bill Gates was like that (1). Steve Jobs was no dummy and had an engineering background, but not at the same level – he did parter with a genius engineer, however.
I think Musk is doing the right thing.
> Which might explain why he's a crummy CEO
That's quite debatable, I'd say. Isn't he getting results?
(1) https://www.joelonsoftware.com/2006/06/16/my-first-billg-rev...
I think it's very important to point out here that Elon does not run SpaceX. It's well known that he's a front man for that company, and that all day to day operations are run by an actual executive who does the job of being an executive. This is why SpaceX is doing much better than Tesla.
Sure, if you call burning billions of dollars a result.
If your goal is to make money, I'd follow anyone but Tesla or SpaceX's example.
Despite the fact that I own a Model 3 and own shares in Tesla I often find the critics make good points and just hope for the best anyways because I want electric cars to take over and I want self driving cars to develop sooner than later because driving is dangerous and tedious.
The SpaceX hate seems much less justified to me. SpaceX has been building launching, landing, and reusing rockets for years now. They've ushered in a new lower cost space age.
On one hand there are hyper-bulls who claim Tesla is a $4000 stock and the future of transportation. On the other, hyper-bears claim the equity should trade around $0-$10. There seems to be no middle ground.
It seems like they are almost betting the company on FSD. I don’t think FSD is really even close to a possibility over the next 5-10yrs. I hope I’m wrong, but if I’m right, I don’t see how Tesla keeps going on like this.
IMO they don't have any choice. The more and longer they operate like a car company shipping bigger and bigger volumes, the more their financials will be undeniably trend towards those of the existing car companies and the more their existing market valuation will be hard to justify.
They need something like this to not get traded at traditional car industry multiples.
Yes! Totally agree.
> Self-driving is not necessary to maintain that valuation, only continued growth of the electric car market in general.
Disagree. The key properties of the car business are high capital costs and high variable costs and not huge margins. The key property of the tech business is low variable cost and often low capital costs (but not always) and high margins.
There's nothing "tech" about building electric cars vs. normal cars and 10 years from now the margins and capital expenses of electric car business will be like the existing ICE car business.
So this is Tesla saying: "Yeah, our financials are starting to look like a normal car company, but we've got this thing that you should keep value us even more like a tech company than you do today."
> 10 years from now the margins and capital expenses of electric car business will be like the existing ICE car business.
Tesla already has margins similar to existing car companies.
> So this is Tesla saying: "Yeah, our financials are starting to look like a normal car company, but we've got this thing that you should keep value us even more like a tech company than you do today."
Yes, they're saying that, but the market is obviously not buying it.
Well, other than actual investors.
Well the middle ground is the actual stock market where Tesla is trading around $265.
That does not mean that their cars can self drive today.
That does not mean that their cars can self drive three years from now.
It's 100% not proven or obvious how car self driving skill and car self driving error rates scale with compute -- but it's surely not linear.
"All we need to do is improve the software" - Musk [12:07 PDT]
LIDAR "unnecessary" - Musk [12:13 PDT]
Computer vision guy is now speaking.
Recognizes "driveable space", not just obstacles. Video shown, but just for a freeway. This is crucial to safety. Need to see this is a cluttered environment.
That's an old claim, see this 2016 press release [1]:
> as of today, all Tesla vehicles produced in our factory – including Model 3 – will have the hardware needed for full self-driving capability at a safety level substantially greater than that of a human driver.
At this point I believe the claim once I see the self-driving functionality having been rolled out to the public and the accident number been reduced.
[1]: https://www.tesla.com/blog/all-tesla-cars-being-produced-now...
Sensors with non-moving optical windows can use other strategies, such as air knives or wipers.
https://www.natice.noaa.gov/pub/ims/ims_v3/ims_gif/ARCHIVE/U...
Here is Europe and Asia on the same day:
https://www.natice.noaa.gov/pub/ims/ims_v3/ims_gif/ARCHIVE/E...
For someone who lives in or close to that white part, "most scenarios" includes snow. Arguably for Level 4 autonomous cars in those areas, you either need to fully disable autonomy in September and enable it again in late April, or you'll need to handle snow.
"No driver attention is ever required for safety (...) self-driving is only supported in limited circumstances (e.g. geofencing), and when these circumstances are no longer met the vehicle must be able to safely abort the trip, e.g. park the car, if the driver does not retake control."
How would that work if you're out driving on the highway, and it starts snowing hard so the car can't see anything? Just park on the highway?
If your car can't safely handle such a scenario, the automous feature would have to be "season-fenced" in addition to geofenced.
I will simply refer back to my initial comment in this thread: "there is going to be a decent amount of time between when a self driving car can handle most scenarios and when it can handle all possible scenarios". Blizzards are not in the "most scenarios" group. Barring emergencies, no one should be driving in a blizzard let alone an autonomous car.
https://www.cambridge.org/core/journals/journal-of-glaciolog...
Because he's sold a bunch of cars (and a bunch of stock in the company selling the cars) without LIDAR with the explicit claim they are hardware-ready for full self-driving.
Freeway shoulders are not safe places to stop. You're likely to get hit by a drunk or distracted driver. Ask any police officer or tow truck driver. That's why stopping on the shoulder is only for emergencies.
Most approaches want to use a wavelength that has as little ambient light as possible. That light is absent precisely because it is absorbed by water vapor in the atmosphere. The alternative approach must deal with lower signal:noise, but is not substantially affected by weather.
Simple physical solutions are available if necessary for the cameras too. I don't have a tesla so I don't know what measures they use to keep the cameras clean but I'm sure they've thought about it and designed for it.
https://teslamotorsclub.com/tmc/threads/water-dripping-into-...
Tesla is new to the game, give them a few more cars to catch small stuff like that.
But ask yourself, what's more likely, the other car companies catch up to their AI and battery tech, or Tesla fixes the trunk in the next model?
In our 2005 DARPA Grand Challenge vehicle, we had a system to clean the camera and LIDAR. It used a spray nozzle, and alternated spraying windshield cleaner and air. This is a commercial truck accessory used on mining trucks.
Every extra they tack on to the base product is one they will be expected to deliver on as well diverting attention from the main problem they are dealing with, which may in the longer term leave open enough room for the competition to wiggle through.
Sure, autonomy is a big deal, but it is also something that will once cracked be an instant commodity and there is a lot of money and talent focused on that particular problem, which is surprisingly hard to do well.
If Tesla ends up going under because of one of the side shows (Solar City, autonomy, Power Wall etc) that would be a serious loss.
SpaceX isn't trying to just do what has always been done and just make it 5 or 10% cheaper. That's boring. SpaceX is completely turning the launch industry on it's head.
Tesla is the same. Elon isn't interested in making "just" an electric car. He wants to change the world, and electric cars that drive themselves will do that.
Can he do it soon? I'm not an expert, I don't know. But damn it's exciting to watch.
Perhaps this is the intent rather than a side-effect.
Intentional misdirection to distract investors.
Musk's ego plays a lot into this for better or worse.
Apple's work on developing their own SOC (A13) have paid dividends for them. The same legendary chip designer (Jim Keller) designed the new Tesla chip after working for Apple.
Elaborate.
Building whole process that will keep on producing solutions that are better than what suppliers can offer, especially in highly competitive area, is way way harder. That's where the real money are spent.
I get what you're saying though. But if they manage to pull off the Robotaxi fleet, then $$$ won't be much of an issue.
If anything, the solar city/roof shingles side-show has seemed the worst to me personally, alongside the falcon-wing doors and the self-infliction of twitter harm by the CEO.
Can't they patent their breakthroughs? How is this different than pharma breakthroughs?
Definitely not and I'm not sure why you think this.
The data pipeline advantage is a big part of this talk.
Not if the company that cracks it eats its competitors or, assuming it's like waymo, commoditizes car manufacturers into suppliers for autonomous taxi fleets.
If Musks ideas about self-driving actually become true, he holds the holy grail of the car industry. On one side, self-driving abilities become crucial in future car sales. On the other side, if he can make the Tesla car sharing network happen, the revenue of that could be a magnitude or more larger than the one of car selling. This would justify some of the crazier valuations of Tesla stock.
Let's hope he is right.
Strong words from Musk about sticking with video only.
Later on in the software talk:
“Lidar is really a shortcut which sidesteps the fundamental problems...and gives us a false sense of progress”
Doubt it.
That is point behind what is known as "probabilistic robotics" - the real world has noise, and you need to be able to deal with it. Don't expect perfect sensors, don't expect a perfect environment.
Self-driving vehicles are the application of probabilistic robotic principles to a real-world task - quite possibly one of the most difficult tasks for the field. To be quite honest, it's amazing how well it's worked in such a short period of time.
I think cameras and machine vision approaches will be needed for self-driving vehicles to be fully successful, but I wouldn't say that LIDAR should be counted out. It will probably be necessary - even required - to have all of the sensors currently being used, not just a subset.
One kind of sensor that hasn't been explored much that I think will be needed is some kind of audio input; some kind of wide-range microphone (probably binaural or stereo) to take in environmental audio and use that for driving cues. Some simple examples might be for vehicle horns, or the squeal of tires, or the revving up of an engine indicating someone is speeding up aggressively, or an emergency siren, etc.
There may even be other sensors needed or that could provide other data to fill in certain gaps of environment knowledge to help a self-driving vehicle navigate. I don't think any one or another should be discounted.
Lidar is expensive, and gives surprisingly limited data (just points of depth) and is like a crutch that will only get so far. It will not get you to full autonomy so why not spend the time and money on vision which will (as proven by humans). They also do use radar and ultrasonics (not visual spectrum).
At some point your machine must be good enough via machine learning to give a very good impression of understanding intention in other road users, pedestrians etc. One example they gave was a distracted pedestrian with a phone. Lidar tells you nothing save an obstacle is on the pavement, it won’t tell you they might step out without looking, but machine learning on a massive dataset can.
Once Tesla can actually handle those two extremely basic tasks maybe they can start talking about how vision is better than LIDAR. Until then, it's just more hot air, and that's not even taking into account Tesla's claims of using ML to predict the actions of uncontrolled independent agents in the field of view.
How does this compare to TPUs and the Neural Engine in iPhone CPUs?
Mumbled Answer: "Safety."
...doesn't that mean the current-gen chip... isn't as safe as you want?
I worry that the laws on the books are insufficient to judge whether a computer is safe to drive. If Tesla already has plans to increase safety, I'd like to know how they judge it, where they think they are, where they'd like to improve, etc.
As a voter.
1) computer has faster reaction times.
2) 360-degree awareness
3) better awareness of vehicle performance.
Imagine a scenario where you are at speed and get a tire blowout. The car can sense it instantly through tire pressure sensor, and counter the swerve appropriately - something that humans are very poorly prepared or practiced for.
Alternatively, a driver in a neighboring lane makes an unsafe lane change into your lane. The car sees him, honks the horn, moves/brakes/accelerates to avoid the other car, before you've even seen whats happening.
Optimally safe should still be the goal. That is, no human could have provided alternative input to the computer that would have created a safer outcome.
Volvo leadership states the exact same thing about their car designs. The primary motivation “in everything they do” is safety.
The statement can be made entirely distinct from the current safety level.
Tesla, on the other hand, thinks this is too fragile to changing environments and works with regular (think google maps level of detail, probably a little better) maps, combined with local semantic information such as signs and lane markings. This also means that they put less pressure on localization, because they don't use the map to detect i.e. speed limit (I'm pretty sure Waymo can detect those signs, it's just that the existence and position of the sign is already known in a global world frame).
Every law enforcement entity on the planet: drool
The future is here, and computers will see everything in excruciating detail via multiple planes of information. The power you consume, the data you send, the money you spend, the media you watch, the words you write, where you go.
It’s all out there in the world. There’s no taking it back, though we may be able to legislate about its use.
You can't say "You can query for anything!" and have it be anonymized.
In fact, if you're taking a camera shot, that's the most broadband sensor imaginable, and no thanks, I don't trust any company with profit motive not to crumble and pop out a software update to selectively disable picture blurrers, or to not have a Boolean flag in your data scrubber.
As a company, making that statement is just trying to slip one past and hope nobody asks inconvenient questions about it.
- First principles hardware design of focused self driving computer (many times better than any competing existing hardware). Already shipping in all newly produced cars. Currently working on next gen that will be 3x better to ship in a couple years.
- Lidar is an unnecessary mistake that competitors are making that won't succeed (too expensive, need too many, unnecessary).
- Real world fleet testing is critical to success, simulations are not good enough since there are too many unknown unknowns in the real world. Tesla uses simulations too, but nobody else comes close on real world fleet testing.
"next gen that will be 3x better to ship in a couple years"
applying psudomathematical varnish to marketing-speak
This is very controversial and rest of the industry thinks exact opposite, many even claiming that Tesla is being irresponsible and even delusional in trying to do autonomy without lidar. The main points I have heard in favor of lidar are that computer vision is very flaky not only in suboptimal weather but even in good weather. Cameras are simply no where close to in performance in dynamic range, rapid adaptation, focusing etc as human eye. Imagine car going under the shaded road with rapid changes between bright light and shade. The likelihood that your depth estimation will get messed up is very high. Of course, night driving becomes highly questionable as well. In additional the long range depth estimation is very flaky with stereo vision right now and a topic of research for mono-vision. If you want to retreat to level-4 only and that too with conservative speed, weather etc then may be vision+radar more doable?
Elon predicts all competitors will eventually drop lidar. He mentions it's expensive, but also not as good in a lot of cases (and all roads/signs are designed for vision).
He argues that getting vision to work is a prerequisite for getting self-driving to work and once you have it working, lidar is worthless (and unnecessary).
Case in point: both the Gigafactory (vastly overbuilt for the quantity of batteries actually produced) and the Alien Dreadnought (vastly overbuilt for the number of cars Tesla current produces...assuming that Tesla is ever able to get the fancy automation working). Boring Co digging a two-mile tunnel in West LA without bothering to learn how to pour concrete smoothly, or to make the "rails" the proper width, or learning about ventilation, or access points....
https://www.theverge.com/2018/2/7/16988628/elon-musk-lidar-s...
But only time will tell.
- If you are trying to sell the Model 3 directly to the end consumer with autonomous mode, the extra $10K for Lidar and bulk (which will greatly effect the exterior design) are definitely non-starters.
- If you are going to robo-taxi route (such as Waymo and Uber), the extra one time $10K cost to add lidar for the 5 year life of the car is probably a blip on the income statement of the operator as compared to a full time human driver which probably costs $10K PER MONTH for the lifetime of the service. For the robo-taxi business model - its a bit of a no brainer - they could stick every sensor known to man on the car and still make out like a warlord by getting rid of the human driver but maintained a 100% safety record. Plus making the car stand-out with a unique Lidar inclusive shape is a great marketing differentiator. Also reduces your liability if a taxi rider ever sues since you can claim you have redundancy in the system.
- Also only a matter of time before lidar cheap enough that even Tesla will add for redundancy and edge cases.
>Currently working on next gen that will be 3x better to ship in a couple years.
Someone asked what the primary design objective of the next generation chip would be and the engineer muttered 'safety' before Musk made the 3x claim.
Except that it isn't. NVIDA's Drive AGX Pegasus delivers 320 TOPS:
https://www.marketwatch.com/story/nvidia-says-tesla-inaccura...
Of course, nobody inside the Musk Reality Distortion Field actually cares about this.
That's ... not great. If that was CPU type of unit, it'd be great. But TPU-type accelerators are growing at massive speed (as it's still pretty new and simple tech), where you're more looking for 10x type of gains.