Author: Mr.Yu
There are only half a month left until Tesla AI Day, which will be held on August 19, 2021, local time, when Elon Musk announced it.
According to Musk, Tesla will showcase its latest developments in artificial intelligence software and hardware during the event, and the real purpose of the event is to recruit talent.
Musk said that anyone interested in the field of AI should join Tesla and claimed that this would be the “fastest way to practice ideas”.
As the global pandemic normalizes, humans are gradually finding a way to coexist with the new coronavirus. Even without a grandiose launch event, Tesla showcases its “muscles” in different technological areas every year with a different theme. Last year’s theme was “Tesla Battery Day,” and the year before that was “Tesla Autonomy Day.”
Considering the exploding market for automotive technology talent in China, Musk clearly has a more sober understanding of this issue. He explicitly stated that the only goal of the AI Day event is to attract the best AI talent to join Tesla.
Implicitly, to attract AI talent and resources, there might be a certain threshold to participate in this event.
Hopefully, Musk will remember to set his alarm clock in advance. After all, he previously arrogantly said that if there was nothing important, he would not bother to attend financial reports and telephone conferences.
The protagonist of the event will be Tesla’s supercomputer “Dojo”.
On June 20, 2021, at the International Conference on Computer Vision and Pattern Recognition (CVPR) 2021, Tesla’s Senior Director of AI, Andrej Karpathy (referred to as A.K. below), publicly revealed the details of Tesla’s self-developed supercomputer Dojo for the first time.
Under the introduction of Li Feifei, a famous female scientist and a former student of A.K, the research director of Tesla’s AI team, the supercomputer named “Dojo” in Japanese, was almost ranked among the top five in the world in terms of computing power almost immediately, reaching a peak speed of 1.8 EFLOPS.
In Chinese, “Dojo” corresponds to the verb “kicking the dojo”. That is to say, Tesla, which is pronounced as a car company but is actually a technology company, wants to use its self-developed supercomputer to compete with research giants.## A.K Reveals a Key Information: Introduction of Transformer Model for Large-scale Unsupervised Learning, which is Now Dominating NLP Field and Slowly Entering the Image Vision Field.
Everyone knows that Musk has always shown great support for the technology route of achieving autonomous driving through pure vision, as well as clear hostility towards the laser radar route adopted by most of his competitors. Musk claimed that through a large amount of video data, Dojo can help achieve fully vision-based autonomous driving.
We can see the source of Musk’s confidence from the information released by A.K.
Artificial intelligence in the broadest sense is to simulate human intelligence with machine intelligence. However, if we want computers to react to constantly changing environments in the way humans do, we need powerful supercomputers and massive datasets. Based on this dataset, autonomous driving technology is trained by injecting a large amount of time into neural networks.
This is the original intention of Tesla’s Dojo project: to make Tesla cars realize autonomous driving using optical cameras.
A.K explained Tesla’s approach at the CVPR:
-
Tesla’s supercomputer collects information from the eight cameras installed around the car at a speed of 36 frames per second, extracts image features using the deep residual network (ResNet), and then uses the Transformer algorithm to fuse them together.
-
Based on the Transformer algorithm’s understanding of pixel position relationships, it is applied to image stitching to form complete scene perception.
-
Tesla used a combination of Transformer, CNN (convolutional neural network), and 3D convolutional neural network to perform cross-temporal fusion and generate 3D information output with a spiritual sense based on 2D images.
Barring unforeseen circumstances, Dojo, as Tesla’s third supercomputer, will be used to train Tesla’s Autopilot and FSD AI.
According to the available information, Dojo has already tagged 6 billion objects from over 1 million 10-second-long videos, occupying a storage capacity of 1.5 PB (1 PB=1024 TB, 1 TB=1024 GB).
For a company that heavily relies on computer vision to achieve autonomous driving, there is no precise standard to prove how much data is needed for a “reliable” autonomous driving system. However, larger data sets, longer training time, and more powerful supercomputers are undoubtedly indispensable.
But this is not the ultimate goal.
Tesla’s future in turning cars into profit interfaces of a business model will be more stable.
On the one hand, Tesla has continually lowered the price of its vehicles. Correspondingly, following the release of Tesla’s Fully Self-Driving (FSD) Beta V9 subscription service in July of this year, the price was set at $199 per month, and the previous option of a one-time payment of $10,000 for the FSD package will no longer exist.
According to Gene Munster, an analyst at venture capitalist research firm Loup Ventures, in ten years, FSD’s annual operating profit will exceed $100 billion, and the market value of this service alone will reach $850 billion, even surpassing Tesla’s current market value of $700 billion (as of August 3, 2021, Eastern Time USA).
In other words, in the near future, Tesla will break away from the traditional model of manufacturing and selling cars, and instead transform Tesla’s cars into a continuous revenue-generating interface. By continuously enhancing its internet service operation model, Tesla will bring considerable sustained cash flow income.
Gene Munster also predicted that by 2032, over 80% of Tesla owners will equip their Teslas with the FSD function, and the profit margin of this part will increase to 64%, and profit will skyrocket from the current $600 million to $102 billion, which is five times the Toyota’s annual net profit in 2020.
With a massive user base, millions of Tesla vehicles are racing on the world’s highways. Each Tesla car is not only a product but also an interface for FSD subscription services and a real road testing fleet that continuously provides valuable road testing data for Tesla to continuously improve its FSD autonomous driving system on giant supercomputers like Dojo.
As mentioned earlier by Mr.Yu, whether or not Elon Musk will tell jokes in Tesla’s upcoming AI Day in half a month seems less important.
This article is a translation by ChatGPT of a Chinese report from 42HOW. If you have any questions about it, please email bd@42how.com.