Click on the blue word to follow us
Before combing through Tesla's AI trajectory, let's first understand Elon Musk's "AI phobia".
The only thing on the planet that can give Musk a nightmare is not Bezos's bald head or Schrödinger's brake pads, but artificial intelligence. In 2014, he tweeted: "We need to be extra careful with artificial intelligence, it may be more dangerous than nuclear weapons." In a later interview, he said alarmistly: "When AI becomes an immortal dictator, the world will never be able to break free [from its control]." ”
Perhaps as if the analogy wasn't shocking enough, Musk further compared the threat of AI to North Korea in 2017 [2], saying on Twitter that AI was "Vastly more risk than North Korea." It then strongly declared that "humans should regulate AI in the same way that food, drugs, airplanes, and cars are regulated." ”
Why so scared? In a 2018 conversation with Westworld screenwriter Jonathan Nolan at the "South by Southwest" conference, Musk explained: "I don't usually advocate regulation and tend to reduce this shackle, but "AI scares me out, it's more capable than almost anyone knows, and the rate of evolution is exponential." ”
In Westworld, Musk's ex-wife Riley plays an advanced AI
However, on the one hand, Musk maintains the persona of "the most carbon-based organism that fears artificial intelligence", and on the other hand, he is investing in AI quickly.
In 2013, Musk personally invested in DeepMind; In 2015, he participated in OpenAI's crowdfunding launch and Vicarious' Series B financing. In 2016, Musk founded NeuraLink, a brain-computer interface company; Tesla has also acquired artificial intelligence companies such as DeepScale, GrokStyle, and Perceptive Automata.
Tesla began to lay out artificial intelligence very early.
In 2013, Tesla exceeded $10 billion with the hot sales of Model S, and immediately began planning to enter autonomous driving. In a conversation with Google's founder in May, Musk put it this way: "Airplane Autopilot is a great thing, and cars should have it." ”
At the time, "autonomous driving" was more of a science fiction concept for traditional car manufacturers. In the 1970s, the global automotive giants defined DAS (driver assistance systems), and then cautiously advanced along this route, "autonomous driving" on the one hand the big manufacturers do not want to do (will bring endless legal nightmares), on the other hand really can't do it.
In 2014, the Society of Automotive Engineers (SAE) divided "autonomous driving" in a broad sense into six categories. It can be seen that traditional car companies have basically been standing in place between L0~L1 in the past few decades, if you want to reach L2 or even higher, the car must use artificial intelligence, and to do this, it is necessary to make the car more like a computer, rather than a simple combination of mechanical and electronic components.
6 levels of autonomous driving, a future think tank
Tesla has achieved electronic and electrical architecture innovation on the Model S, making the car more like a "computer on four wheels". This concept was later summarized by former Huawei Su Jing in the vernacular: traditional car manufacturers believe that the base of the car is the car, and then embed the computer into it; We think that the base of the car is a computer, and then we hang the car up.
The original intention of Modle S to change the electronic and electrical architecture is to reduce costs, such as reducing expensive and heavy car wiring harnesses, but the new architecture can at least make all parts of the car obey the unified command of the "brain", which is equivalent to building a set of embryo rooms for the landing of artificial intelligence.
The embryo room is ready, but what else is needed to make AI really "move in" - to achieve "automatic driving" above level 2?
What we usually know as "autonomous driving" is that the car uses various sensors to collect data about the surrounding environment, and then the brains of the car (the core is chips) analyze these data according to algorithms to control vehicle behavior. For example, the camera sees a dog suddenly swooping out of the front, and the brain analyzes it and sends out an emergency brake command to stop the car.
In this process, identifying whether it is a German shepherd or a black garbage bag that comes out of it requires a set of "algorithms". These algorithms need to be loaded into the "brain" of the car in advance, input the data collected by various sensors of the car, and then make real-time judgments to control the behavior of the car.
The car has to collect data, load algorithms, and make rapid judgments during driving, and its own computing performance can not drop the chain, especially when traveling at high speed, decision-making 1ms late may cause a disaster, if the "stuck machine" is even more a disaster. Therefore, the performance of the chip installed on the car cannot be fooled, and there must be enough computing power.
When Musk starts to engage in AI, he must have a feeling: compared to manufacturing, there are too many AI and chip talents in the United States.
When Tesla prepared to develop its own driverless car in 2015, it was already a popular fried chicken in the technology circle, and Musk had the capital to pry away all kinds of cows and gods from Silicon Valley factories. From 2015 to the present, the structure of Tesla's driverless team has been adjusted many times, and the personnel are bustling, but whether it is hardware or software, the team leaders selected by Musk are basically the world's top scientists or engineers.
We can get a glimpse of the extremely high talent density of Tesla's Autopilot team from the resumes of several big bulls: Jim Keller, the pioneer of AMD K7/K8/Zen architecture, Pete Bannon, a core member of Apple's chip team, Chris Lattner, the inventor of the Swift programming language, Andrej Karpathy, chief scientist of OpenAI...
Tesla team (from left): Ganesh Venkataramanan, director of hardware and head of Dojo; Director of Engineering Milan Kovac; Andrej Karpathy, Director of Artificial Intelligence; Ashok Elluswamy, Director of Software; Elon Musk, 2021 Tesla AI Day
At its peak, Tesla's Autopilot team had more than 300 top engineers (not including more than 1,000 data annotators), 200 of whom specialized in software and 100 specialized in hardware and chips, Musk said in an interview.[8] These elites "can find jobs wherever they go, and no one is their real boss." ”
2018 is the sprint node of Tesla Autopilot's self-development: Andrej Karpathy, director of artificial intelligence, leads the team to train algorithms through large neural networks; Hardware god Jim Keller and successor Pete Bannon presided over the research and development of terminal FSD chips; Veteran executive David Lau led a team of nearly 100 people to improve data collection and vehicle-machine interaction...
Can Tesla hand over a satisfactory answer? Not only Nvidia wants to know, but people all over the world who want to copy homework are also waiting.