23. April 2019
by Paul Fosse
In this article I will investigate the claims made by Tesla on Tesla Autonomy Day April 22. Although I am a small investor in Tesla and arguably a fan of the company and their cars, I try to be as objective as I can be and show where Tesla's claims are proven or undisputed and where they are unproven and require a shot of believe. To get a robot ax, Tesla must have the following pieces of the puzzle:
1. Cars. Tesla claims to have built around 500,000 vehicles that have the required sensor array (some Tesla built since October 2016 have 8 cameras, 12 ultrasonic sensors and 1 radar), and these cars can get the upgraded Full Self Driving (FSD) computer. Tesla is planning to produce another 500,000 cars next year, so it will have about a million cars available to compete with Uber and Lift on a switch.
Of course, Uber and Lyft have more than a million cars registered on their platforms (thought to be over 2 million only on Uber). Not all the millions of Tesla cars are willing to put a stranger in them for money. On the other hand, Uber and Lyft drivers only run a few hours a day, while Teslas can run up to 24 hours a day, since they do not need a driver.
It seems that Tesla will not have the scale to do significant damage to Uber and / or Lift when it first rolls out services. On the other hand, investors are trying to anticipate the future with their investments, and if they perceive Tesla's history to be credible, it will do a lot of damage to Uber and Lyft – if Tesla can scale over the next 5 years without paying a percentage For the drivers, it is obvious that the company will have much lower costs than Uber and Lyft, unless they can access millions of self-driving cars.
2. Redundancy. Tesla needs cars that can accelerate, brake and steer using electric motors. Elon and team mentioned that they have full redundancy in braking and steering (they did not mention acceleration) so they can get faults in a steering motor and a motor brake motor and still safely steer and stop. I think you want to stop and get the problem fixed, not continue with the single steering and brake motor. Although this can be done with gas cars, most people claim that controlling a car is a little easier with electric cars. This point is not really contested by the company's critics.
3. Electric cars against gas or diesel cars. You can probably build a self-propelled gas or diesel car (if you can of course find self-driving), but it is undisputed that electric cars have much lower fuel costs (about a quarter of the cost in most areas). If you only drive a few miles a day, it tends to counterbalance the higher original purchase price of the electric car. If you drive the car a lot, like for 24 hours a day to maximize income, the lower cost of an electric car becomes very important. Tesla is the only electric car manufacturer in the United States that has significant scale. It appears that car manufacturers across the industry electrify their lineups, but there is much controversy as to how fast this will happen and even if it will happen.
Elon claimed that their model 3 engine and body can go a million miles and their battery packs can go 300,000 to 500,000, but it's untried. At the presentation, Elon claimed that a new battery pack would come out next year, designed for multiple charge cycles, so it would last a million miles. This is untested, but Elon's record with such demands is excellent. He has always delivered the promised battery performance, but not always in the promised time frame. It is believed that maintenance costs on Tesla cars are much lower than petrol cars, and although this benefit is sometimes disputed, the evidence is quite strong that it exists.
4. Sensor array. Does Tesla have the right sensors?
No one disputes that cameras, ultrasonic sensors and radar are very useful, but almost everyone thinks it needs suffering. I've written about it here. CleanTechnica has also addressed it here and here. The problem is that although the lidar makes it easier to find the safe areas to run, since it gives you a 3D map of the room without using any artificial intelligence (it just shines a laser and measures the time to jump back), it works Not in bad weather, and it does not help with many other problems you need to solve to make self-driving. Lasers do not help with stop signs, traffic lights or knowledge of bicycles or pedestrians or cars or predict future behavior of any of the three. Lidar does not help reading directions or signs or any of the aids used worldwide to help billions of human drivers.
Lidar is great if you just want to put a car in a science project and make it walk around on the road and not run into some stationary objects in perfect weather. Then you don't need any fancy software, you can just tell the car where the stationary objects are and find a path around them.
As you can see clearly (pun intended), lidar does not help with any of the problems of complex urban environments – moving people, bicycles, animals, cars and trucks controlled by humans or animals that make unpredictable things in any weather . For that you will need some intelligence, either human or artificial.
What a coincidence this was made living today …
So we have pronto, zoox, Commaai and Tesla …  #TeslaAutonomyDay pic.twitter.com/9B8GLkYONL  – Anner J. Bonilla🇵🇷🛩️🔋🔧 (@annerajb) April 23, 2019
It's Anthony Levandowski talking over.
5. Intelligence to understand the environment around the car. To understand the way forward, Tesla claims that you need a little modest CPU and graphics processing power and a massive amount of multiplication and additional power for linear algebra. As I wrote, almost a year ago, Tesla saw what the industry had available to meet its computing requirements and found no one working on a chip that met its performance requirements (especially the ability to process a single image at a time – instead of batches of 256 images – with very low power). If you use too much power, you significantly affect the car's driving range.
Elon recruited a top team with experience from Digital Equipment, Intel, Apple and AMD to create a custom chip. Since they had modest requirements for CPU and graphics needs, they licensed the existing designs and just put them on the chip. But since they had unique needs for high performance multiplication and very low power additional operations, and they couldn't find any acceptable solutions available for license, they designed a very simple processor with very high performance. It is a known truism in the semiconductor industry that you can make a piece faster for an operation if you do not need to handle a complex instruction set.
You also see this in cryptocurrency mining. If you are willing to design a piece to do mining, you can perform the operations much faster and use less power than using CPUs or GPUs to do the math operation. The reason why each currency does not have a chip designed to remind is that it costs a great deal to design each chip, and it is difficult to predict which cryptographic curves will be used to repay the first chip design cost. Of course, this is not a concern here – if this piece solves the full self-purchase problem, there is no dispute as to whether there is tremendous demand.
I've heard critics claiming that Tesla is unlikely to be able to design a chip that is better than "experts" on Intel and AMD, but I find their plan viable for several reasons:
- They hired top quality talent from the industry .
- The only specially designed parts of the piece that they had unique requirements for. The licensed proven (but not leading) designs for CPU and graphics engines. This project would be much more risky if they had specially designed the whole piece.
- They produce the chip on a Samsung fab. Elon may love vertical integration, but he is wise enough to realize that the construction of a 14 nanometer lithography process that moves to a 10 nanometer process is a headache they did not need to touch.
6. You need a lot of training data.
There is no doubt that Tesla has many more cars running around with cameras than all other players in the world combined. It is disputed that they can afford the mobile data charges to send the data back to the mother ship. If they can't (and they are likely to be able to send a small portion of the data back), do they choose the right trial to get the edge cases they need to make the cars safe? They use driver interruptions to help them decide what routine video they do not need, and what is special they need to look at and train image recognition software to handle.
7. Image recognition and depth perception.
Elon made it clear in the presentation that Andrej Karpathy was not only a PhD and a professor of artificial intelligence at Stanford, but that he developed the very popular class taught it in image recognition and is undoubtedly the one top expert in the world training neural networks to recognize images. I think few would dispute that Andrej is a top expert, but many (including me) are not convinced that image recognition could go as fast as Elon claims. I have read many articles on this and it probably sounds, but there is a big leap in ability that I cannot help but doubt whether they can make so much progress in such a short time.
I would say that an example where my skepticism was misguided was Alexa's natural language ability. I had seen 30 years of PC products claiming to make speech recognition, and they took a lot of training for disappointing results. So suddenly, Alexa (and I've heard Google have a good one too) solved the problem, and it seems to understand what I'm saying pretty well. It still seems pretty silly to do complex tasks, but it does a good job in simple.
Tesla has a good team, but this problem is just incredibly difficult. This is really the area Tesla just has to prove that it seems because the world is not going to trust them, whatever they say.
8. Drive the car when you recognize which objects are out and where they go. This is not too difficult with the exception of the chicken game that drivers play when they try to change paths. Tesla must prove that they can find a way to be confident enough to merge into a crowded lane without causing a minor accident. This is difficult for people and it will be difficult for computers as well.
I got away from Tesla Autonomy Day impressed by Tesla's strategy and enthusiasm, but convinced that they will be able to pull it off next year. In my 35 years of software development career, I have seen many examples of a project that I expect to take 4 years, completed in a year of excellent management and programming talent. I have also seen several projects that could have been completed in a year, be interrupted after several years of unfinished, usually due to management who had great vision but insufficient talent to pull it off. Too complex development processes have also killed some projects, but I do not expect it to be a problem at Tesla. Elon has developed commercial software since he was 12 – he won't let a bad process kill this project.
My opinion is that they can pull it off, but I don't really know if they can do it next year or not. As Yogi Berra said, "Predictions are difficult, especially about the future."
If you want to take advantage of my Tesla referral link to get 1,000 miles of free Supercharging on a Tesla Model S, Model X, or Model 3, here's the link: https://ts.la/paul92237 (if someone else helped you, please use the code instead of mine). I urge you to buy before the Full Self Driving (FSD) prize goes up on May 1st if you believe in Tesla's ability to make it work soon.