Tesla Autonomy Day | Full Live Event | Self Driving Demo

Hello, what's up progress people, what

event

and what kind of text rights can we witness? It is truly a joy. I really don't want to stop you from watching the

full

Tesla

autonomy

event

, so if you want to see my analysis and hear my thoughts, don't forget to subscribe. I'll also link it in the description below as soon as it's ready. If you want to support my activism for sustainable energy and transportation, please like this video, share it with those who should see it, and if you want to contribute a dollar or more to help me, check out the Patreon link in the description below. below and have fun with that and I will see peace in the next one.

Hello everyone. I'm sorry I'm late. Welcome to our First day of analysts for

autonomy

. I really hope that this is something we can do a little more regularly now to keep you informed about the development we are doing regarding autonomous

driving

. About three months ago we were preparing for our fourth quarter. Earnings call with Elon and a lot of other executives and one of the things I told the group is that of all the conversations I continue to have with investors on a regular basis, the biggest gap I see with what I see inside the company. and what the external perception is is our autonomous

driving

capability and it makes sense because for the last few years we've been talking about the model three ramp and you know a lot of the debate has been around the model three, but in reality, a lot of things have been happening in the background, we have been working on the new force of you have a chip, we have had a complete overhaul of our neural network for vision recognition etc., so now that we finally start producing our be complete. -Driving the computer we thought it was a good idea to just open the veil, invite everyone in and talk about everything we've been doing for the last two years, so about three years ago we wanted to use it and we wanted to find the best chip possible for

full

autonomy and discovered that there is no chip designed from scratch for neural networks, so we invited my colleague Pete Bannon, vice president of silicon engineering, to design such a chip for us.

He has about 35 years of experience in chip construction and design. About 12 of those years were for a company called PA semi that was later acquired by Apple, so he worked on dozens of different architectures and designs and was the lead designer, I think for Apple's iPhone 5, right before joining Tesla and will be Elon Musk joined the stage. Thanks, in fact, I was going to introduce Pete, but that's it, he's the best traveler and systems architect I know in the world and it's an honor to have you, your team at Tesla and We'll take him, just tell him.

I don't think I ever tried to work in Unity. Thanks Eli. It's a pleasure to be here this morning and a real pleasure to tell you about all the work that my colleagues and I have been doing here. Tesla for the last three years, I think I'll tell you a little bit about how it all started and then I'll introduce you to the fully autonomous computer and tell you a little bit about how it works, which we'll go into further. the chip it

self

and I'll go over some of those details. I'll describe how the custom neural network accelerator we designed works and then show you some results and hopefully you'll still be awake by then.

I was hired in February 2016, I asked Elon if he was willing to spend all the money necessary to do a completely custom system design and he said, "Well, we're going to win." We hired a group of people and started thinking about what a custom-designed chip for full autonomy would look like. We spent eighteen months doing the design and in August 2017 we released the design for manufacturing. We got it back in December. and it actually worked very very well on the first try we made some changes and released a Rev B zero in April 2018 in July 2018 the chip was qualified and we started full production of production quality parts in December 2018 we had the autonomous driving stack runs on the new hardware and we were able to start retrofitting employees' cars and testing the hardware and software in the real world last March, we started shipping the new computer in the Model S and beginning of April I started production and model three, so this whole program, from hiring the first employees to having it in full production on our three cars, lasts a little over three years and is probably the systems development program fastest I have ever been associated with. and it really speaks volumes to the benefits of having a tremendous amount of vertical integration to allow you to do simultaneous engineering and accelerate implementation.

In terms of goals, we were totally focused exclusively on Tesla's requirements and that makes life a lot easier if you have just one customer, you don't have to worry about anything else. One of those goals was to keep the power below 100 watts so we could fit the new machine to existing cars. We also wanted a lower part cost to be able to allow for redundancy for safety at the time we had our thumb in the wind, I presented that it would take at least 50 trillion operations per second of neural network performance to drive a car , so we wanted to get at least that and really as much as We could possibly batch measure how many things it operates at the same time, for example Google's TPU has a batch size of 256 and should wait until it has 256 things to process before be able to start.

We didn't want to do that, we designed our machine with a batch size of one, so that as soon as an image appears, we process it immediately to minimize latency, which maximizes security. We needed a GPU to do some post-processing at the time we were doing quite a bit. A lot of that, but we speculated that over time the amount of post-processing on the GPU would decrease as neural networks got better and better and that actually happened, so we took a risk by putting a fairly modest GPU in the design. You'll see, and it turned out to be a good bet.

Security is very important. If you don't have a safe car, you can't have a safe car, so there's a lot of emphasis on safety and then of course on safety in terms of actually doing the chip design, as Elon alluded to earlier, in There was actually no basic neural network accelerator. In 2016, everyone was adding instructions to their CPU, GPU, or DSP to improve inference, but no one was actually doing it natively. so we set out to do it ourselves and then for other components we bought industry standard IP for cpus and gpus which allowed us to minimize design time and also risk to the program, another thing that was a bit unexpected when I did it for the first time.

What came was our ability to leverage existing teams at Tesla. Tesla had wonderful power supply design teams. Signal integrity analysis. Package design. Systems software. Firmware board designs. And a really good systems validation program that we were able to leverage to accelerate this program. This is what you see. As there on the right you can see all the connectors for the video that comes in from the cameras that are in the car, you can see the two autonomous computers in the middle of the dashboard and then on the left is the power supply. and some control connections, so I really love when a solution is stripped down to its most basic elements, has video and power computing, and is straightforward and simple.

Here is the original 2.5 hardware case that the computer was installed in and that we have been shipping for the last two years here is the new design for the fsd computer, it is basically the same and that is of course driven by limitations to have a modernization program for automobiles. I would like to point out that this is actually a pretty small computer to fit on. Behind the glove box, between the glove box and the firewall of the car, it doesn't take up half of the trunk, as I said before, there are two completely independent computers on the dashboard.

You can see them, they are highlighted in blue and green on each side. In the large SOC you can see the DRAM chips that we use for storage and then on the bottom left you see the flash chips that represent the file system, so these are two independent computers that boot and run their own operating system . Yes, if I can add something that The general principle here is that if any part of this could fail and the call will still work, so the cameras could fail, the power circuits could fail, one of the entire strata of Tesla could fail. for autonomous computer chips, the car keeps driving.

The probability of this computer failing is substantially less than that of someone losing consciousness; that's the key metric, at least by an order of magnitude, yes, so one of the additional things we do to keep the machine running is to have redundant power supplies in the car, so a machine runs on one power supply. power and the others on the other, the cameras are the same, so half of the cameras work with the blue power supply, the other half works with the green power supply and both chips receive all the video and process it in a way independent, so in terms of driving the car, the basic sequence is to collect a lot of information from the world around it, not only do we have cameras, we also have radar GPS maps, the IM uses ultrasonic sensors around the car, we have tear angles on the wheels.

We know what the acceleration and deceleration of the car is supposed to be, all of that is integrated to form a plan, once we have a plan the two machines exchange their independent version of the plan to make sure it is the same and assuming we are okay then act and drive the car now once you have driven the car with some new control you have the costs you want to validate it so we validate that what we transmitted was what we intend to transmit to the other actuators in the car and then you can use the sensor array to make sure this happens, so if you ask the car to accelerate, brake, or turn left or right, you can look at the accelerometers and make sure you're actually doing it. , so there is a tremendous amount of redundancy and overlap in both our data acquisition and our data monitoring capabilities.

Here we move on to talk a little about the complete autonomous driving chip. It is packaged in a thirty seven point five millimeter BGA with 1600 balls, most of which are used to feed ground but many for signals as well, if you remove the cover it looks like this, you can see the substrate of the package and you can see the dye in the center, if you remove the dye and turn it over, it looks like this, there are 13,000 C, four bumps spread out. On top of the tint and then under the mesh underneath there are 12 layers of metal and if that obscures all the details of the design, if you remove it, it looks like this: This is a 14 nanometer FinFET CMOS process, it's 260 millimeters diameter.

The size, which is a modest size ISO for comparison, the typical cell phone chip is about a hundred square millimeters, so we're quite a bit larger than that, but a high-end GPU would be more like six hundred eight hundred square millimeters, so we're about in the middle, I'd call it the sweet spot, it's a comfortable size to build, there's 250 million logic gates in there and a total of six billion transistors, which, although I work on this all the time, it's mind-blowing to me. The chip is manufactured and tested to ECQ 100 standards, which is a standard automotive criteria.

Next, I'd like to just walk through the chip and explain all the different pieces, and I'll do it in the order that a pixel is displayed. coming from the camera would visit all the different pieces, so above at the top left you can see the Euler interface of the camera. We can ingest 2.5 billion pixels per second, which is more than enough to cover all the sensors we know of. On-chip network that distributes data from the memory system so that pixels travel through the network to memory controllers on the left and right edges of the chip. We use industry standard ddr4 LPD memory that runs at 400 4266 gigabits per second, giving us a maximum bandwidth of sixty-eight gigabytes per second, which is a pretty healthy bandwidth, but again, this is not ridiculous , so we are trying to stay in a sweet spot and comfortable for cost reasons.

The image signal processor has an internal 24-bit processor. pipeline that allows us to take full advantage of the HDR sensors that we have around the car, it does advanced tone mapping that helps bring out details and shadows and then it has advanced noise reduction that just improves the overall quality of the images that we have. We are using in the neural network the neural network accelerator it

self

, there istwo of them on the chip, each has 32 megabytes of SRAM to maintain temporary results and minimize the amount of data we have to transmit on and off the chip, which helps reduce power each.

The matrix has a 96 by 96 multiple sum matrix with in-place accumulation that allows us to do almost 10,000 multiplied ads per cycle as dedicated clustering hardware from riilu hardware and each of these de

live

rs 306, excuse me, each de

live

rs 36 billion of operations per second and operate. at 2 gigahertz, the two together and delivered 72 billion operations per second, so we exceeded our goal of 50 rates by quite a bit. There is also a video encoder, we encode video and use it in a variety of places in the car, including the rear view camera screen there is optionally a user function for the dash cam and also for a clip that records data in the cloud, which Stuart and Andre will talk about later there is an on-chip GPU its modest performance it has support for 32 and 16 bit floating point and then we have 12 to 72 64 bit CPUs for general purpose processing that operate at 2.2 gigahertz and this represents approximately two and a half times the performance available in the current solution.

There is a safety system that contains two CPUs running at the same time, this system is the final arbiter of whether it is safe to drive the actuators in the car, so this is where the two plans come together and we decide whether or not it is safe to do so. move forward and lastly there is a security system and then basically the job of the security system is to ensure that this chip only runs software that has been cryptographically signed by Tesla, if it has not been signed by Tesla then the chip will not It works now, I already said it.

There are a lot of different performance numbers and I thought maybe it would be helpful to put that into perspective a little bit, so throughout this talk I'm going to talk about a neural network from our narrow camera that uses 35 Giga 35 billion operations 35 Giga Applications and if we used all 12 CPUs to process that network, we could do one and a half frames per second, which is super slow. I'm not suitable for driving the car if we use the 600 gigaflop GPU on the same network we would use. get 17 frames per second which is still not good enough to drive the car with cameras, the neural network accelerators on the chip can deliver 2100 frames per second and you can see on the scale as we go that the amount of computing and CPU GPUs are basically insignificant for what is available in the neural network accelerator.

It's actually night and day, so moving on to talk about the neural network accelerator, we'll just stop for some water. On the left is a cartoon of a neural network. To give you an idea of ​​what is happening, the data comes at the top and visits each of the frames and the data flows along the arrows to the different frames. The frames are typically convolutions or d-convolutions with actual exudates. green boxes are grouping layers and The important thing about this is that the data produced by one box is then consumed by the next box and then you no longer need it, you can discard it so all that temporary data is created and destroyed as you go. in the network there is no need to store that chip and DRAM, so we keep all that data in SRAM and I will explain why that is super important in a few minutes.

If you look on the right side of this, you'll be able to see that in this network. Of the 35 billion operations, almost all of them are convolution which is based on dot products, the rest are deconvolution also based on dot product and then riilu and pooling which are relatively simple operations, so if you were designing some hardware, clearly you would aim to knit. products that rely on multiplied ads and really kill that, but imagine you sped it up by a factor of 10,000 so that 100% suddenly becomes 0.1% 0.01 percent and suddenly riilu and pooling operations will be quite significant, for what our hardware not our hardware design includes dedicated resources for processing riilu and pooling, now this chip is operating in a thermally restricted environment, so we had to be very careful with how we burn that energy, we want to maximize the amount of rhythmicity that we use. we can do it, we choose an integer ad, it has 9 times less energy than a corresponding floating point addition and we choose an 8-bit by 8-bit integer multiplication, which has significantly less energy than the multiply operations and probably has enough precision to get good results in terms of memory we choose to use SRAM as much as possible and you can see there that going from chip to DRAM is about a hundred times more expensive in terms of power consumption than using local SRAM, so clearly we want to use local SRAM as much as may be possible. possible in terms of control, this is data that was published in a paper by Mark Horowitz in SCC, where he sort of criticized how much power it takes to execute a single instruction on a normal integer CPU and you can see that the add operation is only 0.15 cents percent of the total energy, the rest of the energy is general control and accounting, so in our design we basically reset to get rid of all that as much as possible because what we are really interested in is the arithmetic, so here's the layout that When you're done you can see it's dominated by the 32 megabytes of SRAM, there are big banks on the left and right and bottom center and then all the computing is done in the top middle.

At each clock we read 256 bytes of activation data from the SRAM array 128 bytes of weight data from the SRAM array and combine them into a 96 by 96 mol array that performs 9000 multiplied ads per clock at 2 gigahertz, i.e. a total of 3.6 336 air points in 8 Tare operations now that we are done with a dot product, we download the engine to transfer the data through the dedicated riilu unit, optionally, through a pooling unit and then finally, to a write buffer where all the results are added and then we write 128 bytes. per cycle back to the SRAM and all of this is spinning all the time continuously, so we're doing dot products while downloading the previous results, grouping and writing back to memory, if you add it all up to the year in Hertz, you need one. terabyte per second of SRAM bandwidth to support all that work and therefore the hardware supplies it, so one terabyte per second of bandwidth per engine there are two on the chip two terabytes per second the chip has the accelerator has a relatively small instruction set we have a DMA read operation to fetch data from memory, we have a DMA write operation to send the results back to memory, we have three instructions based on dot products, convolution, deconvolution, inner product and then two relatively simple, a scale is a one input, one output operation and L The wise thing is two inputs and one output and then of course stopping when you're done.

We had to develop a neural network compiler for this, so we took the neural network that has been trained by our vision team as it would be implemented in older cars. and when you take it and compile it for use in the new accelerator, the compiler performs a layer fusion that allows us to maximize the computation every time we read data from the SRAM and put it back, it also softens the demands of the system a bit. of memory is not too uneven and then we also do channel padding to reduce burst conflicts and we do bank aware SRAM allocation and this is a case where we could have put more hardware in the design to handle the conflicts of the bench, but pressing it. in the software we save hardware on power at the cost of some software complexity, we also automatically insert DMA into the graph so that the data arrives just in time for computing without having to stop the machine and then at the end we generate all the code that we generate. we compress the weight data and add a CRC checksum for reliability to run a program.

All descriptions of neural networks. Our programs are loaded into SRAM at startup and then stay there ready to go all the time to run a network. you have to program the address of the input buffer, which is presumably a new image that just came in from a camera, set the address of the output buffer, set the pointer to the network weights and then set go and then the machine runs turn off and it will do the sequence the whole neural network on its own usually runs for a million or two million cycles and then when it's done an interrupt occurs and the results can be post-processed so moving on to the results we had like We aim to stay below 100 watts, these are measured data. of cars driving around running full autopilot and we are dissipating 72 watts, which is a little more energy than the previous design, but with the dramatic improvement in performance it is still a pretty good response of those 72 watts, they are consumed around 15 watts. run the neural networks in terms of cost, the silicon cost of this solution is about 80% of what we paid before, so we are saving money by switching to this solution and in terms of performance we take the narrow chamber neural network which I've been talking about that has 35 billion operations, we run it on the old hardware in a loop as fast as possible and deliver 110 frames per second, we take the same data, the same network compiled it for the new FST computer hardware and we use with the four accelerators we can process 2300 frames per second, so a factor of 21.

I think this is perhaps the most significant slide. It is night and day. I've never worked on a project where the performance increase was more than three, so that was nice. It's funny if you compare it with what is said in the videos: Driving Xavier's solution, a single chip offers 21b operations. Our fully autonomous two-chip computer has 144 ter operations. So, in conclusion, I believe we have created a design that offers exceptional performance. 144 tears. for neural network processing it has exceptional energy performance, we managed to include all that performance in the thermal budget we had, it allows for a fully redundant computing solution, it has a modest cost and really the important thing is that this FSD computer will allow a new level of security and autonomy in Tesla vehicles without affecting their cost or range, something that I think we are all waiting for, yes, I think when we do quality control after each segment, if people have questions about the hardware, they can ask right now.

The reason I asked Pete to do a much more detailed analysis and perhaps most people would appreciate diving into Tesla's fully autonomous computer is because at first it seems unlikely how it could be that Tesla, which had never before designed one chip, out designed the best ride in the world, but that's objectively what happened not the best by a small margin better by a big Roger is in cars right now all the Teslas being produced right now have this computer that we switched from the Nvidia solution to SMX about a month ago and switched to model three about ten days ago, all the cars that are produced have all the necessary hardware and otherwise for fully autonomous driving I will say that again all the Tesla cars being produced right now have everything needed for fully autonomous driving, all you need to do is improve the software and later today you will drive the cars with the development version of the improved software and see for yourself that the questions repeat a trip to three very, very impressive global stock investigations in all their forms. and I was wondering, I made some notes, you are using the activation function with respect to Lu, the linear unit of rectification, but if we think about deep neural networks, it has multiple layers and some algorithms may use different activation functions for different hidden layers. , like soft Max or tan H, do you have the flexibility to incorporate different trigger functions instead of Lu into your platform?

So I have a follow-up. Yes, we have tan H and sigmoid information, for example, beautiful. One last question like on the nanometers you mentioned 14. nanometers, as I was wondering, wouldn't it make sense to go a little bit lower, maybe 10 nanometers two years down or maybe seven at the time we start the design, not all intellectual property that we wanted to buy was available in 10 nanometers, we had to finish the design. at 14 maybe it's worth noting that we finished this design maybe a year and a half ago and started designing if the next generation we're not talking about the next generation today, but we're halfway there, that'll be it.

Things that are obvious for a next generation chip that we are doing. Hello, you talked about the software. Now you did a great job. II was impressed. I understood ten percent of what you said, but I trust he's in good hands. Thank you. It feels like you've already finished the hardware pieces and that was very difficult to do and now you have to do the software piece, maybe that's outside of your expertise, but how should we think about that software piece? What could you ask for a better introduction? Talk to Andre. and Stewart, I think so, are there any questions about funding for the chip part before the next part of the presentation is neural networks and software, so maybe I'm on the chip side, the last slide was 144 trillion of operations per second versus Nvidia 21?

Okay, and maybe you could contextualize that for a finance person, why that gap is so significant. Thanks, well, I mean, it's a factor of seven and Delta performance, which means you can do seven times as many frames as you can run neural networks which is seven. times bigger and more sophisticated, so it is a very big coin that you can spend on many interesting things to improve the car. I think Savior's energy usage is higher than ours. Xavier has powers higher than anything else. I don't know what he is. As I understand it, the D power requirements would increase by at least the same factor of seven and the costs would also increase by a factor of seven, yeah, cool, sure, yeah.

I mean the power is a real problem because it also reduces the range so you have the auxiliary power is very high and then you have to get rid of that power because the thermal problem becomes really significant because you had to go get rid of all that can. Thank you so much. I think you already know a lot of things. a little bit, this app asked the questions, so if you guys don't mind the day getting a little bit longer, we'll just do the driving

demo

s afterwards, so if you have, if you have, if anyone needs to pop. go out and

demo

a little early, you can do it.

I want to make sure we answer your questions. Yes, Pradeep Romani from UBS. Intel and AMD, to some extent, have begun to move toward a chip lab-based architecture. I didn't notice anything. garland-based design here, do you think, going forward, that would be something that might be of interest to you from an architectural point of view? a chip-based architecture. Yes, we are currently not considering any of that. I think it is most useful when you need to use. different styles of technology, so if you want to integrate silicon, germanium or DRAM technology on the same silicon substrate, that becomes quite interesting, but until the die size becomes unpleasant, I wouldn't go there, okay, to be Sure, there's the strategy, this is what started, basically three a little over three years ago, where you design and build a computer that is completely optimized and aims for fully autonomous driving, then you write software that is designed to run specifically on that computer and you get the most out of that computer, so you've adapted to the hardware, meaning you're a master of a craft in autonomous driving.

Nvidia is a big company, but they have a lot of customers, so when they apply their resources, they need to make a widespread solution. We care about one thing, autonomous driving, so it was designed to do it incredibly well, the software is also designed to run on that hardware incredibly well and the combination of software and hardware I think is unbeatable. The chip is designed to process video input in case you use say lidar, could it process that too or is it mainly for video? What was written today explained to you that lidar is nonsense and anyone who gets lucky relying on orb lidar is damned damned expensive expensive sensors that are unnecessary it's like having a bunch of expensive pens just fantasies like a pet an appendix is ​​bad well no put a bunch of them that's ridiculous you'll see hey so just two questions about just As far as power consumption is concerned, is there a way to give us a general rule of thumb?

You know that each watt reduces the range by a certain percentage or a certain amount, just so we can get an idea of ​​how much a target model three consumes: 250 watts. per mile, it depends on the nature of the driving as to how many miles that effect in the city would have a much larger effect than on the highway, so you know if you're driving for an hour in a city and hypothetically you have a solution What do you know, if it was a kilowatt, you would lose four miles on a model three, so if you're only going to say 12 miles per hour, then it's like there's a 25 cent impact on the range in the city, it's basically increases the power.

That system power has a massive impact on city range, which is where we think the majority of the Robo taxi market will be its own power, is extremely important. I am sorry, thanks. What is the main objective of the design? of the next generation ship, we don't want to talk too much about the next generation ship, but it will be at least, say, three times better than the current system, within two years, if the chip is the main one. It does not mean that you manufacture the chip, contract it and how much cost reduction that savings represents in the total cost of the vehicle.

The 20% cost reduction I cited was part cost reduction per vehicle, not that, it was not a development cost. It was just the stock, yes, I say that, but if I mass produce them, it's a money saver to do it yourself, yes, a little bit. I mean, most chips are made so most people don't just make them with what's out there. Babbitt's are quite unusual. I think there is no supply problem seen without the chip being mass produced. Cost savings pay for development. I mean Elon's basic strategy was to build this chip and reduce the cost. Anil said that the deal for a million cars a year is correct, yes, sorry, if they are really chip-specific questions, we can answer them.

Other times there will be a Q&A opportunity after Andre's talks and after Stuart's talks, so there will be two more Q&A opportunities. this is very difficult specific, so I will also be here all afternoon, yes, and exactly, and P will also be here at the end, very good, your prisoner. Thanks, that dead photo you had, the neural processor takes up quite a bit of the dye. I'm curious if it's your own design or is there some external IP. Yeah, that was Tesla's custom design and then I guess next there's probably a good amount of opportunity to reduce that footprint as you modify the design.

It's actually pretty dense, so in terms of reducing it, I don't think it gets any better. greatly functional capabilities in the next generation. Well, and then the last question, can you share where you are, are you vaping this part, what? Where are we still? Oh, it's essentially Samsung, yes, honors in Texas, thank you. I'm Graham Tanaka Tanaka Apple, I'm just curious how defensible your chip technologies and design are from an IP point of view and I hope you don't. I will be offering much of the intellectual property from abroad for free. Thank you, we have filed around a dozen patents on this technology, fundamentally it is linear algebra, which I don't think you can patent.

Oh, I'm not sure. Okay, I think if anyone. They started today and they were really good, they could have something like what we have now in three years, but in two years we will sometimes have something three times better talking about intellectual property protection, you have the best intellectual property and some people just steal it for fun. I was wondering if we looked at some interactions with Aurora where companies and industry believe their intellectual property was stolen. I think the key ingredient you need to protect is the weights that are associated with various parameters. your chip can do something to prevent someone from encrypting all the weights so that not even you know what the weights are at the chip level, so your intellectual property stays inside of it and no one knows and no one can just steal it, man .

I wish I was the person who could do that because they were. I would hire them in a heartbeat. Yeah, so it's a really difficult problem. Yes, join. I mean, we encrypt the code. It's a tough journey to figure out, so if you can figure it out. Alright, so give it a try and then also discover the software and the neural network system and everything else. They can design it from scratch. That's all. Our intention is to prevent people from stealing all that stuff. I mean, if they do. I hope it at least takes a long time, it definitely takes a long time, yeah, I mean, I feel like if we were, if our goal was to do it, how would we do it?

You're very difficult, but I think that's how it is. A very powerful sustainable advantage for us is the fleet, no one has the fleet. Those weights are constantly updated and improved based on billions of miles driven. Tesla has a hundred times more cars with fully autonomous hardware than everyone else combined. You know, we have. At the end of this quarter we will have 500,000 cars with eight full cameras and twelve ultrasounds, someone will still be on hardware two, but we still have the ability to collect data and within a year we will have over a million. Because with completely autonomous computer hardware, everything, yeah, we should have fun, it's just a huge data advantage, it's similar to, you know, Google's search engine has a huge advantage because people use it and people who people are programming Google program effectively. the queries and the results, yes, I just insist on that and please rephrase the questions.

I'm a man if it's appropriate, but you know, when we talk to weigh Moe or Nvidia, they speak with equal conviction about their leadership because of their competence. When simulating miles driven, can you talk about the advantage of having real miles versus simulated miles because I think they expressed that you know when you get a million miles, they can simulate a billion and no Formula One racing driver, for example, could ever do it? Successfully complete a real world track without driving in a simulator. Can you talk to us about the advantages that you seem to perceive are associated with ingesting data from real-world miles versus simulated miles?

Absolutely the simulator we have a good experience. Simulation too, but it doesn't capture the long tail of weird things that happen in the real world. If the simulation completely captured the real world, well, we'll never be proof that we live in a simulation. I don't think so. Hopefully. but simulations don't capture the real world, they don't capture the real world, it's really weird and messy, you need the driest cars on the road and we actually get that, get into that in Andre's presentation at Stewart, yeah, okay When Continue with Andre, thank you. The last question was actually a very good transition because one thing to remember about our F is that the computer can run much more complex neural networks for much more accurate image recognition and talk to you about how.

We actually get that image data and how we analyze it. We have our Senior Director of AI, Andre Party, who will walk you through all of that. Andre has a PhD from Stanford University, where he studied computer science focusing on arousal recognition and deep learning. Andre, why not? Don't just talk, make your own introduction, there are a lot of Stanford PhDs, that's not important, yeah, okay, be careful, come on, thanks. Andre started the computer vision class at Stanford, that's much more important, that's what matters, so please, he talks about. your experience in a way that's not shy I'm just telling you that you're really talking, what's up with the SEC redundancy, yeah, I'm sure, yeah, so yeah.

I think I've been training neural networks for basically what's been a decade now and these neural networks weren't actually used in the industry until maybe five or six years ago, so it's been a while since I've been training these neural networks and that included, you know, institutions at Stanford at the opening of Google and I really trained a lot. of neural networks not only for images but also for natural language and designing architectures that combine those two modalities for my PhD, so every computer science class, oh yeah, and at Stanford I actually taught the convolutional neural norc class, so which I was the main instructor of In that class I started the course and designed the entire curriculum so at first it was about a hundred and fifty students and then it grew to seven hundred students or two or three years so it is a very popular class as one of the biggest losses at Stanford right now, so that was also very successful.

I mean, I wonder if he's really one of the best computer vision people in the world, possibly the best. Okay, thanks, yes. Hi everyone, Pete told you all about the chip we designed. runs neural networks in the car, my team is responsible for training these neural networks and that includes all data collection from thefleet neural network training and then part of the deployment on that chip, so what do you know then what exactly works? in the car, so what we're seeing here is a sequence of videos from throughout the vehicle, on the other side of the car, there are eight cameras that send us videos and then these neural networks look at those videos, process them and make predictions about what will happen.

We're looking at and so some of the things that we're interested in there are some of the things that you're seeing in this visualization here our lane line marks other objects the distances to those objects what we call drawable space shown in blue which is where the car can go and many other predictions such as traffic lights, traffic signs, etc. Now for my talk, I'm going to talk in roughly three stages, so first I'll give you a brief introduction to neural networks and how they work. work and how they train and I need to do this because I need to explain in the second part why it is so important that we have the fleet and why it is so important and why it is a key factor to actually train these neural networks and make them work properly. effectively on roads and in the third stage I will talk about vision and lidar and how we can estimate depth from vision alone, so the central problem that these networks are solving in the car is that an image recognition, so, For United, this is a very simple problem.

You can look at these four images and you can see that they contain a cello on an iguana or scissors, so this is very simple and effortless for us, this is not the case. for computers and the reason for this is that these images are for a computer really just a massive grid of pixels and on each pixel you have the brightness value at that point and so instead of just seeing an image, a computer actually comes up with a million numbers. a grid that tells you the brightness values ​​at all positions, the main nose so to speak, really is the matrix, yes, so we have to move from pixels and brightness values ​​to high level concepts like iguana , etc., and like you.

We might imagine that this iguana has a certain pattern of brightness values, but iguanas can actually take on many appearances, so they can have many different appearances, different poses, and different brightness conditions against different backgrounds, you can have different crops of that. iguana, so we have to be robust in all those conditions and we have to understand that all those different brightness palette patterns actually correspond to iguanas. The reason you and I are very good at this is because we have a massive neural network inside our heads that processes those images, so the light hits the retina and travels to the back of the brain, to the cortex. visual, and the visual cortex consists of many neurons that are connected to each other and that do all the pattern recognition on top of those images and actually in the last, I would say about five For years, the most advanced approaches to the image processing using computers have also started to use neural networks, but in this case, artificial neural networks, but these artificial neural networks, and this is just a cartoon diagram, are a very rough mathematical approximation to your visual cortex.

We will actually have neurons and they are connected to each other and here I only show three or four neurons in three or four in four layers, but a typical neural network will have tens to hundreds of millions of neurons and each neuron will have a thousand connections, so these they're really big chunks of almost simulated tissue and then what we can do is take those neural networks and show them images so that, for example, I can feed my iguana into this neural network and the network. will make predictions about what is seen now at the beginning, these neural networks are initialized completely randomly, so the connection strengths between all those different neurons are completely random and therefore the predictions of that network will also be completely random, so you might think that you We're actually looking at a boat right now and it's very unlikely that it's actually an iguana and during training during a training process really what we're doing is we know that that's actually what it is. an iguana, we have a label, so what we are What we do is basically say we would like the probability of the iguana to be higher for this image and the probability of all other things to go down and then there is a mathematical process called backpropagation stochastic gradient descent that allows us to propagate backwards. that signal through those connections and update each of those connections sir and update each of those connections just a small amount and once the update is complete the iguana probability for this image will increase a little bit so you could become 14% and a property of the other things will decrease and of course we don't just do this for this single image, we actually have large entire data sets that are labeled, so we have many images, usually , you might have millions of images, thousands of tags or something and you're doing steps back and forth over and over again, so you show the computer here's an image, it has an opinion and then you say this is the correct answer and it tunes a little bit, you repeat this millions of times and sometimes you show images, the same image to the computer, which you also meet hundreds of times, so training the network will usually take a few hours or a few days depending the size of the network you are training. and that is the process of training a neural network.

Now there's something very unintuitive about the way neural networks work that I really have to dig into and that is that they really require a lot of these examples and they really start from scratch, they don't know anything. and it is very difficult to understand this, so as an example, here is a cute dog and you probably don't know the breed of this dog, but the correct answer is that this is a Japanese spaniel. Now we are all watching this. and we're looking at a Japanese spaniel, we're fine, I get it. I understand what this Japanese spaniel looks like and if I show you some more pictures of other dogs, you can probably make out other Japanese spaniels here, so those in particular. three look like a Japanese Spaniel on the others and they don't so you can do this very quickly and you need an example but computers don't work like that they actually need a ton of Japanese Spaniel data so this is an grid of japanese spaniels. to show them you need a source of examples, show them in different poses, different brightness conditions, different backgrounds, different crops, you really need to teach the computer from all different angles what this Japanese spaniel looks like and it really takes all that data to so it works. otherwise the computer can't detect that pattern automatically, so for us this all involves setting up the auto grab, of course we don't worry too much about dog breeds, maybe we will at some point, but for now we really care about the line.

It marks objects where they are, where we can drive, etc., so the way we do it is we don't have tags like iguana for the images, but we do have fleet images like this and we are interested in, for example, marks of lines. So we, a human being, typically go into an image and, using a mouse, annotate the lane line markings. Here's an example of an annotation where a human could create a label for this image and it says that's what you should see in this image. the lane line markings and then what we can do is go to the fleet and we can ask for more images from the fleet and if you ask the fleet, if you do a good job with this and just ask for random images. the fleet could respond with images like this, usually moving along some road, this is what you could get as a random collection like this and we would write down all that data now if you're not careful and just write down a random distribution of this. data, your network will pick up this random distribution of data and work only in that regime, so if you show you a slightly different example, for example, here is a picture that actually the road has curves and is a bit more of a residential neighborhood so if you show the neural network this image, that network might make a wrong prediction, it might say okay, I've seen it many times on highways, the lanes just move forward, so here's a possible prediction and so Of course, this is very wrong, but you can't really blame the neural network, it doesn't know that the train in the tree on the left matters or not, it doesn't know if the car on the right matters or not towards the lane line. you don't know if the buildings in the background matter or not, you really start completely from scratch and you and I know the truth is that none of those things matter, what really matters is that there are some white lane markings on top. there and at a vanishing point and the fact that they curve a little bit should alter the prediction, except that there is no mechanism by which we can tell the neural network, hey, those ley line markings really matter, the only tool in the toolbox we have it is labeled. data, so what we do is we take images like this when the network fails and we need to label them correctly, so in this case we will turn the lane to the right and then we need to send a lot of images of this to the neural system. net and neural that over time will accumulate will basically pick up this pattern that those things don't matter, but the lane markings do, and we learn to predict the correct lane, so what's really critical is not just the scale of the data set.

Not only do we want the millions of images we really need to do a very good job of covering the possible space of things the car might encounter on the roads, so we need to teach the computer how to handle scenarios where there is low light and humidity. . you have all these different specular highlights and as you can imagine the brightness patterns and these images are going to look very different. We have to teach a computer how to deal with shadows, how to deal with forks in the road, how to deal with large objects that they might be grabbing. we collect most of that picture how to deal with tunnels or how to deal with construction sites and in all of these cases there is again no explicit mechanism to tell the network what to do, we just have massive amounts of data that we want to get all those things from. images and If you want to annotate the right lines, the network will pick up the patterns from those now large and varied data sets, basically it makes these networks work very well.

This is not just a finding for us here at Tesla, this is a ubiquitous finding across the industry. So experiments and research from Facebook's Google's Baidu's Alphabets Deepmind show similar graphs where neural networks really love data and love scale and variety. As you add more data, these neural networks start to perform better and get higher accuracies for free, so there's just more data. makes them work better now several companies have pointed out that we could potentially use simulation to achieve scale of the data sets and we are in charge of many of the conditions here, maybe we can achieve some variety in a simulator now at Tesla and that It was also mentioned in the questions just before this now at Tesla this is actually a screenshot of our own simulator we use simulation extensively we use it to develop and evaluate software I have even used it for training quite successfully but when It's about training neural network data, there's really no substitute for real data.

Simulator simulations have many problems with modeling the physics of the appearance and behavior of all the agents around. So there are some examples to try really, that point in the real world really throws a lot of crazy things at you, so in this case, for example, we have very complicated environments with snow, trees and wind, we have several visual artifacts that are difficult to simulate potentially we have complicated construction sites, bushes and plastic bags that can come in and that can move with the wind, complicated construction sites that can feature many people, children, animals, all mixed together and simulate how those things interact and flow through of this construction.

The zone is being completed in a completely intractable way, it's not about the movement of any pedestrians there, it's about how they respond to each other and how those cars respond to each other and how they respond to you driving in that environment and all of that is really complicated. To simulate it's almost like you have to solve the autonomous driving problem to just simulate other cars in your simulation, so it's really complicated, we have dogs, exotic animals and in some cases it's not even that you can't simulate, the thing iscan. It didn't even occur to me, yeah, so, for example, I didn't know that you can have truck after truck like that, but in the real world you find this and you find a lot of other things that are very difficult to achieve.

Really, the variety I'm seeing in the data coming from the fleet is crazy compared to what we have in a simulator. We have a really good simulator, yeah, it's like a simulation that you are fundamentally a pimple. You're grading your own homework so you know that if you know you're going to simulate it, that's fine, you can definitely solve it, but like Andre says, you don't know what you don't know, the world is very strange and has millions of edge cases and if someone can produce a self-driving simulation that precisely matches reality which in itself would be a monumental achievement of human capability, they can't, there's no way, yeah, so I think the three points that I've really tried so far are to To get neural networks to work well, you need these three essential elements, you need a large data set, a very large data set and a real data set, and if you have those capabilities, you can train your networks and make them work very well.

Well, then why is Tesla such a unique and interesting position to really get these three essential elements right and the answer to that, of course, is the fleet. We can actually get data from it and make our neural network systems work extremely well, so let me. We'll show you a concrete example of, for example, how to make the object detector work better to give you an idea of ​​how we develop them into everything that works, how we iterate on them, and how we get them to work overtime so that object detection be something that matters to us.

There are a lot of things we would like to put bounding boxes on, let's say the cars and the objects here because we need to track them and understand how they might move, so again we could ask the human annotators to give us some annotations for these and the humans. You could go in and tell it okay, those patterns are cars and bikes, etc., and you can train your neural network on this, but if you're not careful, the neural network hole will fail predictions in some cases, so like For example, if we ran into a car like this that has a bike on the back, then the neural network actually activated when I joined, it would actually create two deductions, it would create a car deduction and a bike deduction and that's actually correct because I assume that both objects really exist, but for the purposes of the controller and the post planner, you don't really want to deal with the fact that this bike can go with the car, the truth is that that bike is attached to that car, so in terms of like just objects on the road, there's a single object, a single car, so what you'd like to do now is potentially annotate a lot of those images, since this is just a single car, so the process that we go through internally on the team is that we take this image or some images that show this pattern and we have a machine learning mechanism by which we can ask the fleet to provide us with examples that look like this and the fleet could respond with images containing those patterns, so as an example these six images could come from fleet, they all contain bikes on the back of cars and we would go in and annotate all of them as a single car and then the performance That detector actually gets better and the network internally understands that, hey, when the bike is just connected to the car, it's actually just a car and it can learn that, given enough examples and that's how we solve that problem.

I will mention that I talked quite a bit about getting fleet data. I just want to briefly point out that we have designed this from the beginning with privacy in mind and that all the data we use for training is anonymous now that the fleet not only responds with bikes on the back of cars we search for everything we search for many things everything time so for example we search for boats and the fleet can respond with boats we search for construction sites and the fleet can send us many construction sites from All over the world we search for even slightly rarer cases, so for example finding debris on the road is very important to us.

These are examples of images that have come to us from the fleet that show tires, cones, plastic bags and things like that. If we can get them to scale we can annotate them correctly and the neural network will learn how to deal with them in the world. Here is another example of animals, of course it is also a very rarely occurring event, but we wanted the neural network to really understand what is happening here, that these are animals and we are. we want to deal with that properly, so to summarize the process by which we iterate on neural network predictions looks like this: we start with a seed data set that was potentially randomly obtained, annotate that data set, and then train its networks with that data. set and the pen in the car and then we have mechanisms by which we notice inaccuracies in the car when this detector may be misbehaving, for example, if we detect that the neural network could be uncertain or if we detect that or if there is driver intervention . or any of those configurations, we can create this trigger infrastructure that sends us data of those inaccuracies and, for example, if we don't do very well in detecting lane lines in tunnels, then we can notice that there is a problem in the tunnels in those that the image would enter. our unit tests so we can verify that we have actually fixed the problem over time, but now what you need to do is fix this inaccuracy, you need to get many more examples that look like this, so we ask fleet to send us many more tunnels. and then we label all those tunnels correctly, add it to the training set and retrain the network, redistribute and iterate the loop over and over again, so we refer to this iterative process by which we improve these predictions as the data that is implemented iteratively. something potentially in shadow mode that generates inaccuracies and incorporates the training set over and over again and we do this for basically all the predictions of these neural networks.

So far I've talked a lot about explicit labeling, so as I mentioned, we ask people to write down data, this is a time-consuming process and we also respect that, oh yeah, it's just an expensive process and therefore, These annotations, of course, can be very expensive to achieve, so what I want to talk about is also using the power of the fleet. I don't want to go through this human annotation bottleneck, I just want to stream data and automate it automatically and we have multiple mechanisms by which we can do this, for example an example from a project we worked on recently is stream detection. so you are driving on the highway, someone is on the left or right and cut in front of you to enter your lane, so here is a video showing the autopilot detecting that this car is invading our lane, of course we would do it.

I like to detect a current as quickly as possible, so the way we approach this problem is that we don't write explicit code to know if the left turn signal is on, the right turn signal follows the keypad over time and we see if moves horizontally. a fleet learning approach, so the way this works is we ask the fleet to send us data every time they see a car transition from the right lane to the center lane or from the left to the center and then we do it. What we do is we rewind time backwards and we can automatically note that, hey, that car will turn in 1.3 seconds, it will get in front of the unfair view and then we can use it to train instead of your bib, so the neural network will automatically detect many of these patterns, for example. cars are usually strange and then they move this way, maybe the blinker is on.

All of that happens internally within the neural network just from these examples, so we ask the fleet to automatically send us all this data, we can get about half a million images and everything. of these would be annotated for the currents and then we train the network and then we take this outage in the network and we deploy it to the fleet but we don't turn it on yet, we run it in shadow mode and in shadow mode the network is always making predictions, hey, I think this vehicle is going to intervene because of the way it looks, this vehicle is going to intervene and then we look for wrong predictions, so as an example, this is a clip that we had of the shadow mode of the Cut Network and It's a little bit hard to see but the grid thought the vehicle just in front of us and to the right was going to come in and you can see it is flirting slightly where the lane line is trying to encroach a little. and the network got excited and they thought that was going to cut off on that vehicle, it would actually end up on our center line, which turns out to be incorrect and the vehicle didn't actually do that, so what we do now is we just turn the El data engine that we get and that was run in shadow mode is making predictions, it generates some false positives and there are some false negative detections, so we get overexcited and sometimes we missed a stream when it actually happened, all of that creates a trigger that reaches us. and that is now incorporated for free, there are no humans harmed in the process of labeling this data incorporated for free into our training set, we retrain the network and we reimplement shadow mode, so we can spin this a few times and we always look at the false. positives and negatives coming from the fleet and we are happy with the ratio of false positives or negatives, we actually changed a little bit and let the car control that network, so you may have noticed that we actually sent one of our first versions of a copy . detector architecture about, I think, three months ago, so if you've noticed that the car is much better at detecting currents, that's fleet learning operating at scale, yeah, it actually works pretty well, so that's the plate learning, no human beings were harmed in the process, it's just too much. from training neural networks based on data and a lot of shadow mode and looking at those results, another thing very similar to everyone training the network all the time is what it means if the order of ordering pilots in or out of the network is being trained every The mile traveled by the car that is more difficult or superior is to train the network.

Yeah, another interesting way we use this in a fleet learning scheme in the other project I'll talk about is a route correction while driving a car. What you're really doing is writing down the data because you're steering, you're telling us how to get through different environments, so what we're looking at here is a fleet person who turned left at an intersection. and what we do here is we have the full video from all the cameras and we know the path that this person took because of the GPS, the inertial measurement unit, the wheel angle, the wheel marks, so we put all that together and we understand the path that this person took through this environment and then of course this, we can use that to monitor the network, so we get a lot of this from the fleet, we train a neural network on those trajectories and then the neural network predicts paths only from that data, so what this refers to is generally called imitation learning.

We are taking human trajectories from the real world. I'm just trying to imitate how people drive in real worlds and we can also apply the same data engine crank. to all of this and make this work over time, so here is an example of route prediction going through a complicated environment, so what you are seeing here is a video and we are overlaying the network predictions, so this it's a route that the network would follow in green and some yes, I mean the crazy thing is that the network is predicting paths that it can't even see with incredibly high accuracy, they can't see around the corner, but it would, but he says the probability of that curve is extremely high, so that's the way and he nails it, you'll see it in the cars today, but we'll turn on augmented vision so you can see the lane lines and the trajectory predictions of the cars. cars superimposed on the video.

Yes, there are actually more. going under the hood that you might even tell yourself is a little scary, you know, of course, there are a lot of details that I'm skipping, you might not want to write down all the drivers that you could write down, you just might want to just imitate the ones. top. drivers and there are many technical ways that we actually slice and dice that data, but the interesting thing here is that this prediction is actually a 3D predictionwhich we project back onto the image here, so the road ahead is a three-dimensional thing that we're only rendering in 2D, but we know the slope of the terrain from all of this and that's actually extremely valuable for driving, so, by the way, mathematical prediction is actually available in the fleet today, so if you're driving Clover Leafs if you're on a cloverleaf on the road until about five months ago or so, your car couldn't do cloverleafs, but now you can.

That's a backward prediction that runs live on your cars. We shipped this a while ago and today you're going to experience this when traversing intersections, a big component of how we traverse intersections in your drives today comes from an automatic label prediction, what I talked about so far are actually the three key components How we iterate on network predictions and how we make it work over time requires a large, varied, and real-world data set. We can actually achieve this here at Tesla and we do it through scaling to float the data engine by sending things in shadow mode, iterating that loop and potentially even using float. learning where no human annotators are harmed in the process and just using it automatically and we can really do it at scale, so in the next section of my talk I'm going to talk especially about depth perception using just vision so you're familiar with that. there are at least two sensors in the car one is vision cameras that just get pixels and the other is lidar which a lot of companies also use and lidar gives you these point measurements of the distance around you now one thing I would do I would like Point out, first of all, that all of you came here, many of you drove here and used your neural network and your vision, you weren't shooting lasers from your eyes and still ended up here, we could clearly be humans too.

The neural network derives the distance and all measurements and 3D understanding of the world are adjusted from the vision. It actually uses multiple cues to do this, so I'll briefly go over a few of them to give you a rough idea of ​​what's going on inside as a For example, we have two ices marked, so you get two independent measurements at each time step. of the role in front of you and your brain puts this information together to come up with a depth estimate because you can triangulate any point across those two points of view many times.

Instead, animals have eyes placed on the sides, so they have very little overlap in their visual fields, so they typically use the structure of movement and the idea is that they move their head and because of the movement, they actually get multiple observations of the world. and you can triangulate the depths again and even with one eye closed and completely still you can still have some sense of depth perception if you did this. I don't think you would notice me coming two meters towards you or 100 miles back and that's because there are a lot of very strong monocular signals that your brain also takes into account.

This is an example of a fairly common visual illusion where you know these two blue bars are identical, but your brain, because of the way it puts the scene together, just waits. one of them must be larger than the other because of the vanishing lines in this image, so your brain does a lot of this automatically and a neural network also plugs an artificial neural network into the scan, so let me give you three examples how it can reach depth. perception from vision alone is a classical approach and two that are based on neural networks, so here is a video.

I think this is San Francisco from a Tesla, so these are our cameras, our sensors, and we're looking at everything. I'm only showing the main camera, but all the cameras are on, all eight autopilot cameras and if you only have this six second clip, what you can do is stitch this environment together in 3D using multi-view stereo techniques, so! oops!, it's supposed to be a video, isn't it a video? Even though I know it's underneath, here we go, so this is the 3D reconstruction of those six seconds of that car driving down that road and you can see that this information is purely recoverable. videos and, more or less, that's through the process of triangulation and, as I mentioned, multiple view in Syria and we've applied similar techniques a little bit sparser and rougher in the car as well, so it's notable that all of that information is really there on the sensor and it's just a question.

To pull it out, the other project I want to talk about briefly is, as I mentioned, there's nothing about the neural network. Neos are very powerful visual recognition engines and if you want them to predict depth then you need to, for example, look for depth tags and then they can actually do that extremely well, so there is nothing that limits the networks when it comes to predict this monocular depth except the tag data, so an example project that we have looked at internally is we use the forward facing radar shown in blue and that radar is looking and measuring the depths of objects and we use that radar to annotate what vision sees, the bounding boxes that come out of neural networks, so instead of human annotators telling you okay, this car and this bounding box are about 25 meters apart. .

You can annotate that data much better using sensors, so use sensor annotation, so as an example, radar is pretty good at that distance, you can annotate it and then you can train your lab work on it and if you have enough data, this neural network is very good at predicting those patterns, so here's an example of predictions of that, in circles I show radar objects and the keyboards coming out of here are purely from vision. So the keyboards here just coming out of the view and depth of those cuboids is learned through an annotation of the radar sensor.

If this works very well, you will see that the circles in the top-down view will match the cuboids. and they do and that's because you know the letters are very proficient at predicting depths, they can learn the different sizes of vehicles internally and they know how big those vehicles are and you can actually derive the depth from that quite accurately, the last mechanism which I'll talk about very briefly, it's a little bit more sophisticated and it gets a little bit more technical, but it's a mechanism that has recently come up in some articles basically over the last year or in this approach it's called self-monitoring, so what you do in a lot of these articles it's just put raw videos into neural networks without labels of any kind and you can still learn, you can still get neural networks to learn in depth and it's a bit technical so I can't go into all the details, but the idea is that neural network. predicts the depth at each frame of that video and then there are no explicit goals that the neural network is supposed to return with labels, rather the goal of the network is to be consistent over time, so any Predicting depth should be consistent over time. the length of that video and the only way to be consistent is to be right as the neural network automatically predicts the correct depth for all pixels and we have reproduced some of these results internally so this works quite well too so which, in short, people drive. vision only no, there are no lasers involved, this seems to work pretty well.

The point I would like to make is that visual recognition and very powerful visual cognition are absolutely necessary for autonomy, it is not a good thing to have them, since we must have neural networks. that really understand the environment around it and the lidar points are a much less information rich environment, so the vision really understands all the details, only a few points around it have much less information, so as an example to the left, here is a plastic. bag or is that a tire well lidar might give you some points on that, but vision can tell you which of those two is true and that affects your control.

Is it that person who looks back slightly? Are you trying to merge into your lane? on the bike or are just moving around construction sites, what do those signs say? How should I behave in this world? All the infrastructure we've built for roads is designed for human visual consumption, so all sizes. all the traffic lights everything is designed for vision and that is where all that information is and that is why you need that skill if that person is distracted and on their phone they are going to work walk towards your lane those answers to all these questions are only found in vision and they are necessary for level 4 level 5 autonomy and that is the capability that we are developing at Tesla and through this it is done through a combination of training large scale neural hours through the data engine and getting it to work with the time and use the power of the fleet and in this sense lidar is really a shortcut, it avoids the fundamental problems, the important problem of visual recognition that is necessary for autonomy, so it gives a false sense of progress and ultimately Ultimately, it's a crutch that offers really quick demonstrations.

So if I had to summarize my entire talk in one slide, it would be this autonomy thing because you want level 4 and level 5 systems that can handle all possible situations in 99.9% of the cases and go after some of the last ones. some nice ones are going to be very complicated and very difficult and require a very powerful visual system, so I'm going to show you some images of what you might find on any segment of that line, so at first you just have something very simple. The cars move forward, then those cars start to look a little weird, then maybe you have some cars, then maybe if there are cars and cars, maybe you start getting into really weird events, like cars overturned or even cars in the air, we see many things. they come from the fleet and we see them in some ways as a really good pace compared to all of our competitors and so the rate of progress at which you can actually address these issues iterates in the software and really feeds the neural hours with the correct data. that rate of progress is really proportional to how often you encounter these situations in the wild and we encounter them much more often than anyone else, so we'll do great, thank you, it's all super impressive, thank you very much .

How much data, how many images are you collecting on average of each car per time period and then it looks like the new hardware with the dual computers active gives you some really interesting opportunities to run in full simulation a copy of the neural network? while you're running the other one, learn that the other one drives the car and compare the results to ensure quality and then I was also wondering if there are other opportunities to use the computers for training when they're parked in the garage for the 90 percent of the time. I don't drive my Tesla.

Thank you so much. Yes, for the first question, how much data do we get from the fleet? That's why it's very important to note that it's not just a scale of the data set. It's really the variety of that data set that matters, if you just have a lot of images of something moving down the road, at some point a neurologist gets it, you don't need that data, so we're really strategic and how we can choose and the The activation infrastructure we've built is quite sophisticated and allows us to get only the data we need at the moment, so it's not a massive amount of data, it's just very well curated data for the second question regarding redundancy.

Absolutely you Basically I can run the network copy on both and that's how it's designed to achieve a little bit for the entire flock system that is redundant so that's absolutely the case and your last question. Sorry, I didn't train the car. It's a computer optimized for inference, we have a major program at Tesla that we don't have enough time to talk about today called dojo, which is a super powerful training computer, the golf dojo will be able to absorb large amounts of data and train. at the video level and I do massive unsupervised training of large amounts of video with the dojo program dojo computer, but that's for another day as a test pilot in a way because I drive the 405 10 and all this really complicated and long tail happen every day, but the One challenge I'm curious how you're going to solve is changing lanes because every time I try to get into a lane with traffic, everyone cuts you off, so human behavior is very irrational when you're driving in Los Angeles and the car just wants to do it safely and you almost have to do it unsafely, so I was wondering how you're going to solve that problem.

Yes, one thing I will point out is thatI talked about the data engine as iteration in neural networks, but we do exactly the same thing. Same thing at the software level and all the hyperparameters that go into the choices of when we actually link change how aggressive we are, we're always changing those that potentially run in shadow mode and seeing how well they perform and adjusting our heuristics. when it's okay to change lanes, we would also potentially use the data engine and a shadow mode, etc. Ultimately, designing all the different heuristics for when it's okay to change lanes is actually a bit intractable, I think in the general case and therefore ideally.

You actually want to use fleet learning to guide those decisions, so when do humans change lanes, in what scenarios, and when do they feel it's unsafe to change lanes? Let's look at a lot of the data and train machine learning classifiers to distinguish when. It's too safe to do it and those machine learning classifiers can write much better code than humans because they have maximum data backup, so they can actually set all the right thresholds and agree with humans and make something safe. Well, we probably will. a mode that goes beyond Mad Max mode to L.A. traffic mode, yeah, well you know, Mad Max would have a hard time in L.A. traffic, I think so, so it really required a trade-off, since you don't want create unsafe situations, but you want to be. assertive, but that little dance of how to make that work as a human is actually very complicated, it's very difficult to write code, but I think we really do it, it really seems like machine learning approaches are the right way to do it. where we just look at a lot of ways that people do this and try to imitate that we are more conservative right now and then as we gain more confidence, it will allow users to select a more aggressive mode that will be active. for the user, but in the most aggressive modes when trying to merge into traffic there is a slight.

I mean, you know, no matter how many new ones there's a small chance of a fender bender or a non-serious accident, but basically you'll have a choice. If you want to have a non-zero chance of getting into an accident in freeway traffic, which unfortunately is the only way to navigate Los Angeles traffic, yes, yes, yes, yes, I mean, yes, yes, and it was very good, like the story, there was a great movie against everyone because this is a game of chicken that is happening yes, it will offer more aggressive options over time that will be specified by the user yes Mad Max Plus exactly oh yes hi hi Jed o Hummer from Canaccord Genuity thank you and congratulations for everything you have developed When we look at the alpha 0 project, it was a very defined and limited variable in terms of the parameters that allowed the learning curve to be so fast that the risk or desire, what What you're trying to do here is almost developed consciousness in cars through the neural network, so I guess the challenge is how not to create a circular reference in terms of extracting the centralized model of the fleet to that handover where the car has enough information, where is that line?

I guess in terms The point of the learning process is to deliver it when there is enough information in the car and not have to take it out of the fleet. See, the car can work if it is completely disconnected from the fleet, it just charges. the training is that you know it better and better as the free food gets better and better, so simply, if you can really reform the fleet from then on, it would stop improving, but it worked very well, share the heart with the former. version and I talked about a lot of the energy benefits of not storing a lot of images, so in this part you're talking about the learning that occurs when pulling from the fleet.

I guess I'm having a hard time reconciling how if there was a situation where I was driving uphill like you showed and I'm predicting where the road is going to go, which comes from all the other fleet variables that led to that intelligence of how I'm not How am I. I'm getting the benefit of the low power using the cameras with the neural network, that's where I'm losing the...maybe it's just me, but I guess what I mean is that the computing power and the fully autonomous computer are amazing and such Maybe we should mention that if I had never seen that path before, I would still have made those predictions as long as it was a path in the United States in the case of lidar, the march of the nines, isn't there an example that it won't reach? your hit on lidar because it's pretty clear that you don't like the light is in this last flame the lighter is a name that's not there as a case where at some point nine nine nine nine nine in the future we're actually lidar can be useful and why not have it as some sort of redundancy or backup sets up my first question and the second so you can still focus on computer vision but have it as redundant.

My second question is if that's true, what about the rest? of the industry that is building their autonomy solutions on lidar, everyone is going to get rid of lidar, that is my prediction, mark my words. I should point out that I don't actually hate lidar as much as it might seem, but at SpaceX, the basics, dragon uses lidar to navigate to the space station or dock not only that we, SpaceX developed its own light from scratch to do that and I I spearheaded that effort personally because in that scenario lidar makes sense and size, it's fucking stupid, it's expensive and unnecessary and like doing laundry was saying once you solve vision it's worthless so you have expensive hardware that It has no value in the car.

We have a forward radar that is low cost and is useful especially for occlusion situations, so if there is fog, dust or snow. radar can see through that, if you are going to use active photon generation, don't use the visible wavelength because once you use passive optics you have taken care of all the visible wavelengths you want if you want to use a wavelength that is penetrating occlusion like a radar so certain Lana is just individual active photon generation spectrum if you're going to do active photon generation do it outside the visual spectrum on the radars in the radio spectrum like three point eight millimeters versus 400 to 700 nanometers that you're going to There's going to be much better occlusion penetration and that's why we have a forward radar and then we also have just twelve ultrasounds for near field information in addition to the eight cameras and the Ford radar, you just need the radar at four. direction because that's the only direction you're going very fast so I mean we've gone over this several times like we're always sure that we have the right sensor set if we add something else not high so right here, so you hadI mentioned that you asked the fleet for the information you are looking for for part of the vision.

I have two questions about this. Well, it looks like the cars are doing some calculations to determine what kind of information to send you. That's a correct assumption and are they doing it in real time or are they doing it based on stored information, so they absolutely do calculations in real time on the car there and we would wait to basically specify the condition that we are interested in and then those cars they do that competition there, if they didn't then we would have to send all the data and do it offline on our backend, we don't want to do that so all those calculations have us in the car so it's based on that question, it sounds like you guys are in a very good position to currently have half a million cars in the future, potentially millions of cars that are essentially computers that represent free, almost free, data centers for you to do calculations on.

It is a great future opportunity. for the Tesla car, it is a current opportunity and that has not been taken into account yet, that is incredible, thank you, we have four hundred and twenty-five thousand cars with Hardware two and more, which means they have the eight cameras to the right of the radar on ultrasounds and they have at least one nvidia computer which is enough to essentially figure out what information is important and what is not, compress the information that is important into the most prominent elements and upload it to the network for training, so it is a massive data compression In the real world, you have this kind of network of millions of computers that are like massive data centers, essentially distributed data centers for computing capacity.

Do you think it will be used for things other than autonomous driving in the future? Suppose it could be used for something besides autonomous driving, we're going to focus on autonomous driving, so as we understand that, maybe there's some other use for, you know, millions and then tens of millions of computers with hardware , three or four slow running computers, yeah, maybe there would be, could be, maybe this kind of AWS angle here is possible. Hello, hello at Mat Choice Loop Ventures. I have a model 3 in Minnesota where it snows a lot as the camera and radar cannot see the road markings through the snow.

What is your technical strategy to solve this challenge? Is it a high precision GPS? Yes, actually, like today, Auto Pal will do a decent job in the snow, even when Layla's marks are covered even when the marks are faded or when it's raining a lot, we still seem to drive relatively well, we still don't specifically look for snow and with our data engine, but I actually think this is completely manageable because in a lot of those images, even when there's snow, when you ask a human annotator where the lane lines are, they might actually tell you that actually they like relatively consistent in training those ley lines, as long as the annotators are consistent in your data, then I have there is the neural network that will pick up those patterns and work well, so it's really about the signal being there even for the human annotator.

If that's the answer is yes, then the neural network can do it well, yes, actually there are There are several important signals that say that lane lines are one of those things, but one of them is the most important signal is space. to drive, so what is a driving space and what is not a driving space and what really matters most? This is walkable space more than landlines and the prediction of walkable space is extremely good and I think especially after next winter is going to be incredible, it's going to be like it's so good, that's crazy, the other thing.

What I want to point out is that maybe it's not even just human annotators, as long as you as a human can get through that disabled fleet by learning, we actually know the path you took and obviously use vision to guide you to through that road, you don't just use the lane line markings, you use all the geometry of the whole scene, so you see like you know, you see how the world curves roughly, you see how the cars are placed around you, you know that the job will detect all those patterns automatically within it if you just have enough data on people going through those environments, yes, it's actually extremely important that things are not rigidly tied to GPS because GPS error can vary quite a bit and if the actual situation of a road can vary quite a lot, the degree of construction that could be a detour and if the car uses GPS as main this is a really bad situation as looking for problems it is ok to use GPS to get similar tips and tricks , so that you can drive in your home neighborhood better than in a sufficient neighborhood. you like some other country or some other part of the country, so you know your own neighborhood well and you use the knowledge of your neighborhood to drive more confidently and maybe have counterintuitive shortcuts and that kind of thing, but you, it's GPS overlay data should only be helpful but never primary, if ever this issue is primary so ask here in the back corner.

I just wanted to partially follow up on that because several of your competitors in the space in recent years have done that. I know I've talked about how they're augmenting all of their route planning and perception capabilities that are in the car platform with high-definition maps of the areas they're driving. Does that play a role in your system? Sees it? Adding any, are there areas where you would like more data that is not collected from the fleet, but is more mapping style types of data? I think high precision altitude type high precision GPS maps and lanes are a very bad idea.

The system becomes extremely fragile, so any changes like this could make any changes to the system unable to accommodate, so if it gets stuck on GPS and lines ofHigh precision lane and does not allow vision override, in fact your vision should be what does everything it is as if the lane lines are a guide, but they are not the main thing. We briefly barked up the high-precision lane line tree and then realized it was a big mistake and reversed it. It's not good, so it's very useful for understanding the annotations, where the objects are and how the car is driven, but what about the negotiation aspect for parking lots, roundabouts and other things where there are other cars on the road being driven by humans, where is it more artistic? that the science, it's actually pretty good, with things cut out, it works very well, yeah, so I'll mandate that we're using a lot of machine learning right now in terms of prediction, creating an explicit representation of what the roll looks like and then there's an explicit scheduler and a controller on top of that representation and there are a lot of heuristics on how to traverse and negotiate etc., there is a long tail, like in the visual environment aspect, there is a long tail just in those negotiations. and a little game of chicken that you play with other people and so on, so I think we're very confident that eventually there must be some kind of fleet learning component on how to actually do it, because writing all those rules by hand goes to plot quickly oh I think so, we've fixed this issue with cuts and it's like it would allow gradually more aggressive behavior from the user, they can just check the setting and say be more aggressive.

Less aggressive, you know. Drive easy, relaxed mode, aggressive, yes, incredible progress, phenomenal. Two questions first in terms of platoon. Do you think the system is adapted because someone asked when there is snow on the road, but if you have a big winning feature from Platte, you can just keep going? the car in front makes your system your system capable of doing that and I have two follow ups so you're asking about the peloton so I think we could build those features but again if you just use them if you just train yours . networks, for example, about imitating humans, humans already like to follow the car in front and that neural network actually incorporates those patterns internally it's just that it discovers that there is a correlation between the way you look at the car that is before you and the path you will follow. but all of that is done internally in the network, so you're just worried about getting enough data and the complicated data and the neural training process is actually pretty magical.

It does everything else automatically to turn all the different problems into one problem. just collect your data set and use your own top training yes there are three steps to autonomous driving you know this is a complete future so there is a complete future as far as we think the person in the car does not need to pay. attention and then there is the level of reliability, we have also convinced the regulators that that is true, so there are like three levels that we hope to have full features in autonomous driving this year and we hope to be confident enough from our point of view to say we think people don't need to touch the steering wheel, look out the window at some point probably around the second quarter of next year, and then we start hoping to get regulatory approval, at least in some jurisdictions, for that towards the end of the year.

Next year, what is it? That's about the timeline I expect things to go on and probably four, four trucks, regulators will approve platooning before anything else and you could have done it, maybe if it's a long haul trip doing cargo hauling. long distance. you can have one driver in front and then have four semis behind in platoon form and I think regulators will probably be quicker to approve that than other things, of course, you don't have to convince us that lidar is a technology. in my opinion, which one has an answer looking for a probably dead question.

I mean, this is very impressive what we saw today and probably the demo could show something else. I'm just wondering what is the maximum dimension an array can have. your training or in your deep learning process rough figure out what is matrix information so you're doing you know what matrix multiplication operations inside you know that work that you're asking about there's a lot of different ways to answer that question, but I'm not. sure if they are useful, they are useful answers, these neural hours generally have, as I mentioned, between tens and hundreds of millions of neurons, each of them has on average about a thousand connections with the following neurons, so these are the typical scales that are used in T on this train and that will also be reduced, yes.

In fact, I was very impressed with the rate of improvement on autopilot last year on my model three, the two scenarios I wanted feedback on last year. Last week, the first scenario was that I was in the rightmost lane of the freeway and there was an on-ramp to the freeway and then my model 3a was able to detect two cars on the side, slow down and let the car passed in front of me. Me and a car were behind me and I thought oh my god this is crazy, I didn't think my model tree could do that so it was super impressive but the same week there was another scenario where I was on the right .

Manual lane again, but my right lane was merging into the left lane and it wasn't an on-ramp, it's just a normal highway lane and my Model T couldn't really detect that situation and I couldn't slow down or accelerate and I had We need to step in, so from your perspective, can you share the background on what a neural network would look like, how Tesla could adapt to that and do you know how that could be improved in the EU? Over time, yes, as I mentioned, we have a very sophisticated activation infrastructure. If it has intervened, it is potentially likely that we received that clip and we can analyze it and see what happened and adjust the system so that it probably inputs some statistics.

Well, at what rate are we correctly merging traffic and we look at those numbers and we look at the clips and we see what's wrong and we try to fix those clips and make progress against those benchmarks, so yeah, we would potentially miss. a categorization phase and then we looked at some of the most important types of categories that actually seemed to be semantically related to a simple, simple problem and then we looked at some of them and then we tried to develop software based on that. I have one more presentation, which is that software is essentially like Autopilot Hardware with Stewart, there's the neural network kind of vision with Andre and then there's software engineering at scale which is a computing presenter from Stewart, so Thank you and I will have the opportunity. later to ask questions, so yes, thank you.

I just wanted to tell you very briefly that if you have an early flight and would like to take a test ride with our latest development software, you could talk to my colleague and/or send him an email and he can take you for a test ride and Stewart will come over to you, so it's actually a clip of an uninterrupted 30+ minute ride with no interventions. Now forget about autopilot on the road system that's in production today in hundreds of thousands of cars, so I'm Stewart and I'm here to talk about how we build some of these systems that scale like a very short induction.

That's pretty much where I come from in what I do, so I've been in a couple of companies or less. I've been writing about software for about twelve years. What I'm most excited about and what I'm really passionate about. It's about taking the cutting edge of machine learning and actually connecting it to customers through failure and scaling, so at Facebook I initially worked within our infrastructure. of ads to build some machine learning. Some really very smart people and why she tried to build. that to a single platform that Zdenek we could scale to every other aspect of the business, from how we rank the newsfeed to how we deliver search results to how we make each recommendation across the platform and that became the machine learning cluster applied which is something I was incredibly proud of and a lot of that was not just the core algorithms, I'm the really important improvement that happened there, which matters a lot, actually, the engineering practices of building these systems at scale, the same thing happened at the instant, where I went to where we were very excited to help monetize this product, but the hardest part was using Google at the time and they were effectively running us on a pretty small scale and we wanted to build that same infrastructure.

We understand that these users connect to a cutting-edge machine learning build at massive scale and generate billions and then trillions of predictions and auctions every day in a system that is really robust, so when the opportunity came to come to Tesla, that's something I'm really incredibly excited to do, which is specifically taking the incredible things that are happening on both the hardware side and the computer vision and artificial intelligence side and actually packaging it together with all the schedule that controls the testing of the operating system kernel patch. All of our continuous integration, our simulation, and actually turning it into a product that we put into people's cars in production today, so I want to talk about the timeline of how we did that with autopilot navigation and how We are going to make it to measure.

We get into a Vinod cabin off the highway and into the city streets, so we're already 770 million miles away from navigating on autopilot, it's a really cool thing and I think one thing worth noting about This is that we continue to accelerate and we continue to learn from this data, as Andre talked about this data engine. As this speeds up, we actually make more and more assertive lane changes. We're learning from these cases where we'll intervene because they don't detect pop up correctly or because they wanted the car to be a little more dynamic in different environments and we just want to continue to make that progress, so to start all of this, we start by trying to understand the world that surrounds us and we talk about the different sensors in the vehicle, but I want to go a little deeper here, we have eight cameras, but we also have 12 ultrasonic or radar sensors, an inertial measurement unit, GPS and one thing we forget is also the actions of the pedal and steering, so not only can we look at what's going on around the vehicle, we can see how humans chose to interact with that environment, so I'll talk to this clip right now.

This basically shows what is happening in the car today and we will continue to push this forward. We start with a single neural network, look at the detections around it, then build all of that from multiple neural networks in multiple directions, bring in the other sensors and turn it into an Elan called vector space, an understanding of the world around us . And this is something where as we continue to get better at this, we're moving more and more of this logic to the neural networks themselves and the obvious endgame here is that the neural network looks at all the cars and brings back all the information. together and ultimately generates a source of truth for the world around us and actually this is not like a more difficult representation in many ways, it is actually the result of one of the debugging tools that we use in the team every day to understand what the world is like. you see it all around us, so another thing that I think is really exciting to me.

I think when I hear about sensors like lidar, a common question is about having additional sensory modalities like why not have some redundancy in the vehicle and I want to dig deeper. In one thing that's not always obvious with neural networks themselves, so we have a neural network running on our wide fisheye camera, that neural network doesn't make one prediction about the world, it makes many separate predictions, some of which actually audit each one. another, then it is a real example: we have the ability to detect a pedestrian, something we train very carefully and work a lot on, but we also have the ability to detect obstacles on the road and a pedestrian is an obstacle and it shows.

Unlike the neural network, it says, "Oh, there's something I can't get through," and these elements together combine to give us a better idea of ​​what we can and can't do in front of the vehicle and how to plan for it and then do it. this through multiple cameras because we have overlapping fields of view and many around the vehicle in front we have a particularly large number of fields of viewoverlapping. Lastly, we can combine that if things like radar and ultrasonic stabilities understand extremely accurately what is happening in front of the car we can use that to learn future behaviors that are very precise, we can also build very accurate predictions of how things will continue to happen. in front of us, so a really exciting example is that we can watch cyclists and people and not only ask where you are now, but where you are going, and this is actually the heart of the ordinary art of our automatic braking system of next generation emergency, which will not only stop people who cross your path, but also suffer people to encounter. in your way and that's running in shadow mode right now, we're going out to fleet this quarter.

I'll talk about shadow mode in a second, so when you want to start a feature like this to navigate on autopilot on the highway system, you can start by learning from the data and just look at how humans do things today, what is their assertiveness profile, how they change lanes, what causes them to abort or change lanes, like their maneuvers, and you can see things that aren't immediately obvious, like oh. Yes, I will do it. Constant merging is rare but very complicated and very important, and you can start generating opinions on different scenarios, like a vehicle overtaking quickly, so this is what we do when we initially have some algorithm that we want to try.

We can put them in the fleet and we can see what they would have done in a real world scenario, like this car that is passing us very quickly. This is taken from our real simulation environment and shows different paths we have considered taking and how they overlap. the real world behavior of a user when you tune those algorithms and feel good about them specifically and this really removes that for the neural network, puts it into that vector space and ultimately builds and tunes these parameters on top of it . I think we can get through more and more machine learning to get into a controlled rollout, which for us is our early access program and this is getting this out to a couple thousand people who are very excited to give them very thoughtful feedback. but useful about house it doesn't behave like an open circuit but like a closed circuit in the real world and you see their interventions and we talk about when someone takes control we can get that clip, try to understand what happens and one thing we can actually do en In fact, we can play this again in an open loop way and ask, as we build our software, if we are getting closer or further away from how humans behave in the real world and one thing is great about fully autonomous computers that actually we are. build our own racks and infrastructure so that we can basically deliver completely autonomous, fully integrated computers, build them on our own cluster, and actually run this very sophisticated data infrastructure to really understand over time as we tune, these algorithms are becoming approaching. and they behave closer to humans and ultimately we know if we can exceed their capabilities and once we had this we were very good about it.

We wanted to make our launch wide, but to start, we asked everyone to confirm the behavior of the cars via stock confirmation, etc. We started making a ton of predictions about how we should navigate the highway, we asked people to tell us if this is right or wrong and this is again an opportunity to fire up that data engine and we spotted some really difficult and interesting long things. In this case, I think it really finds Apple like these are very interesting cases of simultaneous merging where you start going and then someone moves behind her before you don't realize it and what is the appropriate behavior here and what are the neural network adjustments we need to make to be super precise about the appropriate behaviors here we worked, we tweaked them in the background, we improved them and over time we got 9 million successfully accepted lane changes and used them again with our continuous integration infrastructure to really understand what we think we are ready for and this is something where we are completely autonomous is also very exciting to me as we own the entire software stack directly from kernel patch all the way through.

I suspect that by tuning the image signal processor we can start collecting even more data that is even more accurate and this allows us to better and better tune these faster iteration cycles, which we were fighting for earlier this month. We're ready to employ an even smoother version of Autopilot navigation on the highway system and that perfect version doesn't require a stock confirmation so you can sit there, relax, put your hand on the wheel and simply monitor what it's doing the car and in this. In this case, we're actually seeing over a hundred thousand automated lane changes every day on the highway system and someone is great for us to implement this at scale and what excites me most about all of this is the actual lifecycle of this and how we journey to make that data engine work faster and faster over time and I think one thing that's really becoming very clear is the combination of the infrastructure that we've built and the tools that we've built on top of it. that and the combined power of the fully autonomous computer, I think we can do this even faster as we move forward now to being an autopilot from the highway system to the city streets, so yeah, with that I'll deliver the only which is okay As far as I know all of those lane changes have occurred with zero accidents, that's correct, yes I watch every accident so it's conservative obviously, but having hundreds of billions of lane changes and zero accidents is a great achievement. you have a team yeah thanks so let's look at some other things that you're familiar with mention that to have a self-driving car or a robotaxi you really need redundancy throughout the vehicle at the hardware level so you need to start at Every It was October 2016, all cars made by Tesla have redundant power steering, so we ended up with the motors in the power steering, so if one motor fails, the car can still direct all the power and power lines.

Data has redundancy so you can cut any given power line or any data line and the call will still drive the auxiliary power system even if the main pack loses all power in the main pack, the car is able to turn and brake using the auxiliary power system, so you can completely lose the main line. package and therefore the car is safe, the whole system, from a hardware point of view, has been signed to be a Robo taxi basically since October 2016, so when we were all that hardware autopilot version 2, we do not expect to update the manufactured cars.

Before that we thought it would actually cost more to make a new car than to upgrade it, just to give you an idea of ​​how hard it is to do this, unless it was designed yesterday it's not worth it, so we've gone through the future. of klitz autonomous driving products is the hardware, its vision and then there is a lot of software and the software problem here should not be to minimize some to a massive software problem that yes, manage large amounts of training data against the data, how do you control the car based on vision, it's a very difficult software problem, so going after a guy like Tesla's master plan obviously made a lot of forward-looking statements as they call it, but let's go over some of our forward-looking statements that We haven't backed down when we built the company we sit in, both Tesla Roadster said it was impossible and then, and that even if we built it, no one would buy it.

It's as if the universal opinion was that building an electric car was extremely stupid. and would fail. I agree with them that the probability of failure was high, but this was important, so we built the Tesla Roadster that will go into production in 2008 and the shipment of that car, it is not a collector's item, it never released a car more affordable with the Model S we made. that again they told us that it is impossible, they called me a fraud and a liar, it is not going to happen, all this is false, okay, famous, the last words now are that we went into production with the Model S in 2012, it exceeded all expectations, still in 2019 there is no car that can compete with the 2012 Model S it's been seven years and I'm still waiting so it will be an affordable car maybe very affordable it is affordable more affordable with the model 3 we bought the model 3 we are in production I said we would get over five thousand cars we have model 3 right now five thousand cars a week is a walk in the park for us it's not even difficult so we do large scale solar energy which we did through the acquisition of Seoul city and we are developing to play.solar roof which is going very well now we are in version 3 of the solar tile roof and we expect this to be a significant production of the solar tile roof this year.

I have it in my house and it's great and who would like it. Like I started doing the power wall and the power pack and we made the Power Woman power pack, in fact the power pack is now deployed in massive grid-scale utility systems around the world, including projects of the world's largest operational batteries exceeding 100 megawatts. and in the next year, or probably in the next year, tears at most we hope to have a group on the scale of the drum project that I completed, so all these things I said we would do, we did, we say we do, we did.

We're going to do the Robo Taxi thing just to criticize and it's fair and sometimes I don't show up on time, but I do it and the Tesla team does it, so what we're going to do this year is we're going to reach a combined production of 10,000 per week between the air six and three, we feel very confident about that and we feel very confident that the future will be complete with autonomous driving next year, we will expand the product line with the Why and semi model and we hope to have the first operational Robo taxis on next year with no one in them next year it's always hard to like it when things are going at an exponential rate of improvement it's very hard to correct your mind about it because we're used to extrapolating linearly But when you have massive amounts of hardware in the Along the way, the accumulated data increases exponentially and the software improves at an exponential rate.

I feel very confident predicting Rover's self-driving taxis for next year. It's not a state order for all jurisdictions because we won't have regulatory approval everywhere, but I'm sure we will have the least amount of regulatory approvals somewhere literally next year, so any customer will be able to add or remove their car to the Tesla network, so wait. this to operate is kind of a combination of maybe the uber and airbnb model so if you own the car you can add or subtract it to the

tesla

network and Telsa would take 25 or 30% of the revenue and then on some places. where there aren't enough people sharing their cars, we would just have dedicated Tesla vehicles, so when you use the car we will show you our ridesharing app, it's your tag to hail the car from the parking lot, get in and go for a ride.

Driving is really simple, just grab the same Tesla app you currently have, we'll just update the app and add a Tesla summary or commit your car to the fleet, so make sure some are in your car or as many Teslas. or add or subtract your car to the fleet, you'll be able to do it from your phone, so we see potential to smooth out the demand distribution curve and make a car run at much higher utility than a normal car. For example, car use is typically 10 to 12 hours a week, so most people will drive one and a half to two hours a day, usually 10 to 12 hours a week. of full driving, but if you have a car that can run autonomously, then you can most likely get that car to run for a third of the week or more, so there are 168 hours a week, so You probably have something on the order of 55 to 60 hours a week of operation, maybe a little more.

So the fundamental utility of the vehicles increases by a factor of five. So you look at this from a macroeconomic point of view and say that if this were like we were operating a large simulation, if you can improve the simulation to increase the utility of cars by a factor of five. factor of five, that would be a massive increase in efficiencyeconomics of the simulation, just gigantic, so we will use the models 3, 3 and of the lease contract. We want them back. If you buy the car you can keep it but if you rent it you have to go back on the grid and like I said we are in any location. where there is not enough supply to share, Tesla will simply make its own cars and add them to the network at that location, so that the current cost of the Robo model three taxi is less than $38,000;

We expect that figure to improve over time and by redesigning the The cars being built today are all designed for one million miles of operation, they are the power units, design, design and testing validated for one million million miles of operation. operation, the current battery pack is approximately 300 to 500 thousand miles of the new battery pack. which will probably go into production next year is explicitly designed for one million miles of operation, the entire vehicle battery pack included, well, it's designed to operate for one million miles with minimal maintenance, so we'll actually adjust the design of the tires and we will really optimize the car. for a hyper efficient Robo taxi and at some point it won't need steering wheels or pedals and we'll just leave them so as these things become less and less important we'll just remove the parts that won't be.

If you say that probably within two years we will make a car that has no steering wheels or pedals and if we need to speed up that time, we can always easily remove parts. Yes, it probably says three year long term rubber cabs with parts removed. it might end up costing $25,000 or less and you want a super efficient car, so the illustrated electricity consumption is very low, so we're currently at four and a half miles per kilowatt hour, but we can improve that to five and beyond and there really is no company that has full integration, we have the vehicle design and manufacturing, but the computer hardware is internal, we have the internal development of AI and artificial intelligence and we regret by far the biggest. suite, it's extremely difficult, not impossible, maybe, but extremely difficult to catch up when Tesla has a hundred times more miles a day than everyone else, because this is today, this is the cost of running a gasoline car, the cost average of operating a car in the US this is taken from triple-a so it currently costs about sixty two cents per mile on thirds of fifteen hundred miles of fifty million vehicles amounting to two trillion a year, These are literally taken from the rideshare cost on the triple-a website. to your left, it's two to three dollars a mile, the cost of operating a mobile taxi we think is less than eighteen cents a mile, and falling like that, it's an automobile, this would be current, this current cost, the future cost will be less if you say what the likely growth of a single Robo taxi would be, we think probably something on the order of $30,000 a year and we expect Wordsworth to literally design it, we are designing the cars the same way commercial semi-trucks are and the semi-trailers.

The commercial semi trailers designed are designed for a million mile life and we are also designing the cars for a million mile life so no nominal dollars can be spent i.e. a little over three hundred thousand dollars over the course of 11 years. higher, I think these consumptions are actually relatively conservative and this assumes that 50 percent of the miles driven are art, there is nothing or it is not useful, so this is only 50 percent useful for mid-year. Next year we will have more than a million Tesla cars on the market. road with full self-driving hardware features at a reliable level that we would consider no one to need to pay attention to, which means that from our point of view you could go to sleep if you fast for a year, just look, maybe a year, maybe a year in three months, but next year we will surely have over a million Rover taxis on the roads, the fleet wakes up with an over-the-air update, that's all it takes, say what is the current value net of a Rover taxi, probably on order. of a couple hundred thousand dollars, so buying a Model 3 is a good deal, good questions, well, I mean, in our own fleets, I don't know, I guess in the long term we will probably have on the order of 10 million vehicles, I mean our production rates in general.

If we look at a compound annual production rate from 2012, which is our first full year of Model S production, we went from 23,000 vehicles produced in 2013 to about 250,000 vehicles produced last year, so over the course of five years we increase. production by a factor of ten. I would expect something similar to happen in the next five or six years as far as sharing sharing versus I don't know, but the good thing is that essentially customers advance us the money for the car, it's great, so um. in terms of one thing is the snake charger, I'm curious about that and also how did you determine the price?

It sounds like you're undervaluing the average Lyft or Uber ride by about 50 percent, so I'm curious if you could talk a little bit about that, the pricing strategy, we sure hoped it was solving the solution for the snake charger. It's pretty simple, it's a view of the frontal vision problem, it's like a known situation, any kind of known situation with vision is like a charging port. that's trivial, so yes the car would just park automatically, but and plug in automatically, there would be no one, no human supervision would be required, yeah, sorry, what was the price? uh that we only have three numbers there, I mean, I think it's like definitely plug. at whatever price you think makes sense, we just randomly said, well, maybe a dollar and things like that, there are about two billion cars and trucks in the world, so robotaxis are going to be in extremely high demand for a long time and from My observation so far is that the ordering industry takes a long time to adapt.

I mean, I said there still isn't a car, Lord, that you can buy today that's as good as the Model S was in 2012, which suggests a pretty slow pace of adaptation. the automobile industry and so probably a dollar is conservative for the next 10 years because I still, if you also think that there's not really enough recognition for the difficulty of manufacturing, manufacturing is incredibly difficult, but a lot of people with the What I'm talking about thinks that if you have the right design, you can instantly make as much of it as the world wants. This is not true, it is extremely difficult to design a new manufacturing system for new technology.

I mean the ones with major problems make Neutron and they are extremely good. in manufacturing and if they have problems, what about the others? So, know that there are on the order of 2 billion cars and trucks in the world, on the order of approximately 100 million units per year of vehicle production capacity, but only of the old ones. design, it will take a long time to convert all of that into fully autonomous cars and they really need to be electric because the cost of operating a gasoline and diesel car is much higher than an electric car and any robotex that is an electric will not be not at all competitive Elin, it's a Colin Oppenheimer race here, you know, obviously we appreciate that customers are spending some of the cash to get this fleet being built, but it sounds like a massive bottom line commitment by the organization.

As time goes on, can you talk a little bit about what your expectations look like in terms of funding over the next three years, three, four years to build this fleet and start monetizing it with your you know? customer base, we aim to be approximately cash flow neutral during the fleet build phase and then I hope to be extremely cash flow positive once Robo taxis are enabled, but I don't want to talk about financing, for what everything is difficult. We talk about financing rounds in this place, but I think we will make the right moves. Oh wait, I think I'll make the move, so you think, think about what you did.

I have a question, if I'm Ober, why would I do it? Didn't I just buy all your cars? Do you know why I would let you put me out of business? There is a clause that we put on our cars. I think it was about three or four years ago, they can only use the Tesla network, so even a private person likes that if I go out and buy ten model threes, I can't. I can run on the network, that's a business now. You had only already used it, it does not work on the network, but if I use the test, the network. in theory I could run a car sharing Robo taxi business with my ten model threes, yes, but it's like the AppStore where you can just add them, add them or delete them through the Tesla network and then it tells you that you get a share in income, but similar.

However, for Airbnb, I have this house, my car, and now I can rent them out so I can make extra income by having multiple cars and just renting them out like I have a model three. I aspire to have this Roadster here next time. You build it and I'll rent my model for three hours. Why would I give it back to you? You know, I suppose you could operate a fleet of rental cars, but I think this is very difficult to manage, yes, no, it seems easy, okay, try it. Now, to operate a robo taxi, Orkut, it seems that you have to solve certain problems such as, for example, the autopilot today.

If you turn it too much, it allows you to take control, but if it does, you know if it's a ride-sharing product that someone else is doing it. sitting in the passenger seat, like moving the steering, you can't allow that person to take charge of the car, for example, because they may not even be in the driver's seat, so the hardware already exists for it to be a robbed taxi and you could get into situations. like having a police officer pull you over where some human might need to intervene, like using a central fleet of operators that interact remotely with humans or I mean, it's all that kind of infrastructure already built into each of the cars, does that make sense?

I think there will be some kind of phone home where if the car gets stuck you will just call home, Tesla, and ask for solutions, like having them pull it over for you. A police officer knows it's easy for us to program it and it's not a problem, it will be possible for someone to take control using a steering wheel or at least for a period of time and then probably in the future just cover the steering wheel so there is no steering control, well just remove the steering wheel, put it a cover and if you end up here, give the car's hardware a couple of years of modification to allow it, or yes, we.

We literally just unbolt the steering wheel and put a cap where the steering wheel handle is, Karlie, but that's a similar car to the future that you would take out, but what about today's cars where the steering wheel is a mechanism to take over control of the autopilot? So if it's in robo-taxi mode, someone could take care of it by simply moving the steering wheel? Yes, I think it will be a transition period where people will take control and they should be able to take control of the rover. taxi and then once the regulators are comfortable with us not having a steering wheel, we will just remove it and for the cars that are in the fleet, obviously with the owner's permission, if it is owned by someone else, we would just take the Remove the steering wheel and put a cap where the steering wheel currently connects, so that there can be two phases in the Robo taxi, one where the service is provided and you enter as the driver, but you could potentially take control and then in the future , there might not be an option for the driver that is how you see it or like it in the future, that is how the probability of having the steering wheel taken away in the future will be, since one hundred percent of people will. will demand, but initially I would follow it.

This is not clear. This is not me prescribing a point of view on the world. This is not me predicting what consumers will demand. Yes, consumers will demand in the future that people are not allowed to drive these two-tonne death machines. I totally agree. with that, yes, but for a model tree today to be part of the problem with the taxi network, when you call it, then you would sit in the driver's seat essentially just because, just to be safe, okay, with common sense, Thank you, as you already know. There were amphibians, you know, but then things become like terrestrial creatures, a bit, a bit of inspection phobia phase, uh, hello, sorry, what's okay?

Yes, the strategy we have heard from other players in the Robo taxi space isselect a certain municipal area to create geofenced autonomous driving that way, you are using an HD map to have a more confined area with a little more security. Hey, today we don't hear much about the importance of HD Maps, how much is an HD Map necessary for you in a second. We also don't hear much about implementing this in specific municipalities where you are working with the municipality to get their acceptance and also getting a more defined area, so what is the importance of HD maps and to what extent are specific municipalities being looked at for their implementation.

I think HTML is a bug. We actually had a team after a while. Actually, we can't do it because either you need HD maps, in which case if something changes in the environment. the car will break down or you don't need h TMS, in which case why are you wasting your time during HD maps? So HD maps are like the two main crutches that are turned off and not to be used and in hindsight, hindsight, the obviously fake and dumb maps or lidar and HD mark my words hello, if you need a geofence area, you don't have driving Real autonomous driving, it just seems like maybe the battery supply might be the only bottleneck left for this vision and it might also just clarify how to make battery packs last a million miles.

Hi, I think the cells will be a limitation, that's all, that's a topic for a completely separate topic, there's a completely separate topic and I think we're actually going to want to get that out. standard battery type more battery than our long range battery because the energy content in the long range pack is 50% more kilowatt hours, so it can essentially let you know 1/3 more cars if only if, all type standard range plus instead of the long range pack, those of us have 50 kilowatt hours and the others around 75 kilowatt hours, so we are probably biased in our sales intentionally towards the smaller battery pack to have greater volume. than you basically want, but the obvious most you can do is maximize the number of autonomous units or the amount of maximizing production that will subsequently result in the biggest autonomous leap in the future, so we're giving for a number of things. in that sense, but it's not for today's meeting, the million, my life is basically about getting the package cycle length; you know basically you need it you know in the order like I say you have basic math if you have a 250 mile range package you know you need four thousand cycles so it's very tractable we already do that with our stationary storage , so they remain stationary storage solutions like the power pack or we are ready to use the power pack with 1000. life cycle capacity yes, can I ask sorry, yes, it's like we're 12 years old, not my day, yes, it obviously has very constructive significant margin implications to the extent to which you can drive the Tatra, it is much higher than the full self-driving option.

I'm just curious. If we can level out where you are in terms of those connection rates and how you hope to educate consumers about the Robotech scenario so that connection rates improve materially over time, it's a little hard to hear your question, yeah, I'm just curious to know where and where. We are today in terms of fully autonomous driving at rates in terms of financial implications. I think it's enormously beneficial if those attached fees increase materially because of the higher gross margin dollars that flow in as people sign up for fully autonomous vehicles. FST is just curious how you see that increase or what the fixation rates are today versus when you expect, how you hope to educate consumers and make them aware that they should attach FSD to their vehicle purchases.

We increased that tremendously later. Yes today, I mean, if the fundamental, really fundamental message that consumers should convey today is that it is financial folly to buy anything other than a Tesla, they will be like owning a horse in three years. I mean, okay, if you're known as a horse, but you should go in with that expectation if you buy a car that doesn't have the hardware needed for fully autonomous driving, it was like buying a horse and the only car that has the hardware needed to several autonomous vehicles for Tesla. like people should really think about buying any other vehicle, it's basically crazy to buy any other car other than

tesla

, yes we should make that clearly convey our point and we will have to be thankful for bringing the future to the present.

Today there is a very important informative moment. I was wondering if you didn't talk a lot about the Tesla pickup and let me give you some context. I could be wrong, but the way I see testing on the network will be as an early adopter and somewhat of a test. servant I think that the Tesla truck maybe is the first phase of putting the vehicles on the network because the usefulness of the Tesla truck would be for people who are carrying a lot of things or who are in the construction profession or with few strange items here and out there like Picking up stuff from Home Depot, I'm sure I'd say you know it might be necessary to have two-stage process vans exclusively for testing on the web as a starting point, then people like me can buy them later, but what do you think at the end? regard? today was really just about autonomy, there are a lot of things we could talk about, like the cellular production pickup truck and future vehicles, but today we only focused on autonomy, but I agree that it is an important thing.

I'm very excited about what a truck says. revealing it later this year will be great Colin Lang and UB just so we understand the definitions we need to refer to a full self-driving feature. It sounds like you're talking about level 5 without geofencing, that's what's expected by the end of the year. well and then the regulatory process, I mean, have you talked to the regulators about this? It seems like a pretty aggressive timeline from what other people have posted. I mean, do you know what the hurdles are that are needed and what the timeline is to overcome? approval and you need things like in California no, they're tracking miles, you know what that operator is behind there, you need those things, but what is that process going to be like, yeah, I mean, we talk to regulators around the world all the time. as we present, you know additional features like a navigator and an autopilot, you know that this requires regulatory approval depending on the jurisdiction, so what, but I think fundamentally, in my experience, the regulators are convinced by the data, so that if you have a lot of data that shows that autonomy is safe, they listen to it, it may take them a while to digest the information, their process may take a little time, but they have always come to the correct conclusion from what I have seen, oh, I have a question.

Here, as he says, I have lights in my eyes and a pillar. Well, I just wanted to let you know about some of the work we've done to try to better understand the Ride Heil market. It appears to be highly concentrated in major dense urban areas. centers, so the way to think about this is that Robo taxis would probably be deployed more in that area and the additional failure of full autonomous driving for personally owned vehicles would be in suburban areas. I think probably yes, like Tesla owned Rover taxis would be in dense urban areas along with customer vehicles and then as you get to medium and low density areas it would tend to be more people owning from the car and occasionally land it, yes, there are many extreme cases in Manhattan and let's say in downtown San Francisco, but although those are, you know, and there are several cities around the world that have challenging urban environments, but we don't expect that this is a major problem.

When I say future-complete, I mean it will work in downtown San Francisco and downtown Manhattan this year. Hello, I have a question about the neural network architecture uses different models, for example, for planning and path perception or different types of AI and more or less how that problem is divided between the different parts of autonomy, essentially, right now, iron neural networks we actually use for object recognition and basically we still use it as fixed frames to identify objects and soul frames and put them together in a perception path planning layer from then on, but what What's happening all the time is that the neural network is eating up the software base more and more and over time we expect the neural network to do more and more now from a computational cost standpoint there are some things that are very simple to do. a heuristic and very difficult for a neural network, so it probably makes sense to keep some level of heuristics in the system because they are computationally a thousand times easier than a neural network as a new kingdom that is like a cruise missile and if you're trying to swat a fly just use a fly swatter or not a cruise missile so with a little time I would expect it to actually move to train it against video and then a video in the car is the steering and the pedals out or basically video in that lateral longitudinal acceleration almost completely, that's what we're I'm going to use the dojo system because there's no system that can do that currently, maybe here, just going back to the sensor array discussion, the area of ​​which I'd like to talk about is the lack of side radars in a situation where you have an intersection with a stop sign where there's maybe 35 to 40 mile per hour cross traffic, are you comfortable with the sensor suite? than a human would do, I think you can be human is basically like a camera on a slow gimbal and yes it's quite remarkable that people can drive the car the way they do because if you know what you can't look at all directions at once, the car can literally look in all directions at once with multiple cameras, so humans can drive by just looking this way, looking that way, they are stuck in the driver's seat and can't get out. the driver's seat, so it is like a kind of camera on a gimbal and is capable of making a conscientious driver drive with very high safety.

The cameras in the cars have a better point of view than the person, so they are as above. on the B-pillar or in front of the rearview mirror, their guts really have a great vantage point, so if you're turning onto a highway that has a lot of high-speed traffic, you can just do what the person is like crouch down, turn a bit, don't go all the way onto the road so the camera sees what's going on and if things look good and then the rear cameras don't show up on oncoming traffic or if you go and if it looks sketchy you can back up. a bit like a person, if the behaviors are remarkably similar, it starts to become remarkably realistic, it's quite disturbing, it's actually hard to start behaving like a person, here, here, see, ventriloquist, here, okay, given everything. the value they are creating in their automotive business by wrapping all this technology around it.

I guess I'm curious why they would still be taking some of their cell capacity and putting it into the power wall and power pack, right? It makes sense to put every single unit you know you can make into this part of your business. They already stole almost all of our cell lines that were meant to go to the power wall and package them and use them for bottle three, I mean last year, in order to produce our model three and not sell stock, we had to convert all 2170 lines in the gigafactory to sell two cars and it's our actual production and in total gigawatt hours of stationary storage compared to vehicles is an order of magnitude different and for stationary storage we can basically use a bunch of various cells that They exist so that we can bring together cells from multiple suppliers around the world and their You don't have a homologation or safety problem like you have with cars, so basically our stationary battery business has been running on scraps for quite some time, yes, but we really think of manufacturing as being a massive constrictive production system, there are many, many limitations, it's like IKEA, but the degree to which manufacturing is underestimated in a supply chain is surprising, it's a whole series of limitations. and what the constraint is in one week may not be the constraint in another week it's tremendously difficult to make a car especially one that's evolving rapidly so yeah but I'll just take a few more questions and then I think the desires are broken so that you can test the cars, well Elon Adam Jonas and questions.

About security, what data can you share with ustoday? How secure this technology is, which is obviously important in the regulatory or insurance space. We publish accidents per mile every quarter and what we see now is that Autopilot is about twice as safe as a normal driver, you know, on average, and we expect that to increase quite a bit over time. As said, in the future it will be the consumers who will want to ban. I don't think they will achieve it nor am I saying that I agree with this position. but in the future consumers will want to prohibit people from driving their own cars because it's not safe if you think about elevators, elevators used to be operated with a big lever like to go up and down the floor and there's like a big relay and yet , elevator operators, but then periodically they would get tired or drunk or something and then they would turn the lever at the wrong time and cut someone in half, so now you don't have elevator operators and it would be quite alarming if you walked into a elevator that had a large lever. that could just move between floors or betray early, so there are just buttons and in the long run, again, it's not a value judgment.

I'm saying I want the worldview this way. I'm saying that consumers will most likely demand that people not be tracked by many cars Drive, can you share with us how much Tesla spends on Autopilot or autonomous technology by order of magnitude annually? Thanks, that's basically our entire expense structure question here about the economics of the Tesla grid, just so I understand that it seemed like if you got a model three on a lease, $25,000 would go to the balance sheet and be a active and then you would generate a cash flow of approximately $30,000 a year. That's the way to think about yeah, kind of, yeah, and then just in terms. of the financing of it and there is a question earlier that you mentioned that you would ask: neutral cash flow for the Robo taxi program or neutral cash flow for Tesla as a whole.

Sorry, cash flow in terms of you asked a question about the financing of the Robo. taxes, however, it seems to me that they are self-funded, but yes, you mentioned that they would be basically cash flow neutral. That's what he was referring to now. I'm just saying that between now and when Robo taxis are fully deployed around the world, the sensible thing for us to do is to maximize the rate and drive the company to cash flow on its troll once the Robo taxi fleet is active, I would expect extremely positive cash flow and this, so you were talking about production, yeah, do it with producible, okay, thanks maximize. the number of autonomous units manufactured, thank you, okay, maybe one last question, yes, if I add my Tesla to the Robo taxi network, who is responsible for an accident?

Is it Tesla? If the vehicle is in an accident and probably damages Tesla, it is probably Tesla. Yes, the right thing to do is to make sure there are very few accidents. Well, thank you all, please enjoy the trips, thank you.