"This is the world's first end-to-end AI autonomous driving system."
A few days ago, Musk was doing a test drive live broadcast of FSD Beta V12, and he proudly made an introduction.
In the 45-minute test drive, Musk took over only once.
Musk's magic wand is it to hit the face of Chinese car companies "far ahead"?
Since October 2020, FSD Beta has been pushed to early bird users, and FSD has been completing iterations with small steps and fast research and development progress, of which Musk showed the 12th major version a few days ago.
Friends who are familiar with Musk's style should remember that almost every year he has to say "FSD will land this year".
In the face of V12 version, Musk also said at the end of June this year, "FSD V12 will cancel the beta version and push it to all users." ”
What can support this statement is that this FSD adopts an end-to-end technical roadmap.
In May this year, Musk revealed that the V12 version will adopt a new end-to-end solution.
Then in early July, Musk delivered a speech at the 2023 World Artificial Intelligence Conference, in which he claimed that L4/L5 autonomous driving will be realized later this year, which shows Musk's confidence in the performance of the V12 version.
At the end of July, he again revealed that he had tested the V12 alpha version of FSD Beta.
On August 2, Musk said that Tesla was developing the last piece of the FSD puzzle "vehicle control", and he revealed that after using the end-to-end solution, the original more than 300,000 lines of C++ control code was directly reduced by about 2 orders of magnitude.
At the time, he also said "Versailles", "Tesla is training these neural networks, and it is the scale of training calculations that limit our progress at the moment, not by engineers." ”
On August 26, the end-to-end based FSD Beta V12 was finally released.
In the debut of Beta V12, Musk chose the old Model S with Hardware 3.0 hardware.
In the context of its computing power of only 144TOPS, Tesla tries to teach the "computing power theory" of Chinese car companies a good lesson.
In addition, Musk pointed out that the buildings and road signs encountered in this test drive have never been seen before.
Do you believe that you chose a line that has never been run to complete the global debut of the Beta V12 version?
During Musk's 45-minute test drive, the Beta V12 version performed relatively smoothly, such as the Model S smoothly completing corners when passing through relatively complex roundabouts.
It's just that after the car in front left, the Model S did not accelerate during the corner, but kept a low speed to enter the corner, which was a bit low in terms of efficiency.
Musk said, "The team never programmed the concept of roundabouts, and we just showed it a bunch of videos about roundabouts. ”
In addition, in a construction area enclosed by roadblocks on both sides, Model S can pass smoothly in such a passage that can only accommodate a single vehicle, which shows the accuracy of Beta V12 in identifying obstacles and handling path planning when dealing with irregular roads.
From this point of view, Beta V12 has little problem reaching the "usable" level.
As for the only place where Musk took over, it was when waiting for the traffic light at the intersection to continue going straight, but Model S suddenly started after recognizing that the opposite left turn signal turned green, and Musk quickly stopped the vehicle.
In this regard, it is the Model S's misjudgment of the intention of this traffic light.
Musk immediately said, "Let's feed more traffic light videos, and the problem will be solved." ”
He added that FSD V12 is still being tested, which is why V12 has not been released.
Summarizing this Beta V12 debut, Tesla easily demonstrated its high-level strength in the context of the simple road conditions in North America.
Despite the simple road conditions, on the basis of low computing power and the absence of high-definition maps and lidar, Beta V12 can have such high driving continuity and stability, and it is not difficult to beat most new power brands.
More importantly, don't forget that this is the "world's first end-to-end" smart driving system debut.
Musk emphasized that "the V12 system is implemented through AI from start to finish, and we are not programmed. ”
"Instead of having a programmer write a line of code to identify roads, pedestrians, etc., and having Tesla stop at a stop sign or wait for another car, we left it all over to neural networks."
FSD V12 saves hundreds of thousands of lines of code through extensive high-quality driving video training.
Musk also said that all inference decisions are done on HW3 hardware, the inference computing power is about 100 watts, and all inference is done completely locally, without the need for a network connection.
As for the possible existence of problems with the system, Musk mentioned that because it hardly rains in California, where Tesla is headquartered, it needs to be trained with the help of driving videos in different weather conditions around the world.
The strength of FSD, no one denies.
However, this case-in-law, which is called "the world's first end-to-end" case, still arouses many people's doubts.
Is Tesla really running this line for the first time?
Is it really completely end-to-end?
This cannot be verified, and the only way is to check the program code.
There's a reason people don't believe Tesla's end-to-end.
Simply put, end-to-end is to output the final result by simply entering the raw data, which means unifying the three major parts of autonomous driving: perception, prediction and planning into a whole.
Further, the general automatic driving system has a more subdivided system architecture, including detection, tracking, static environment mapping, high-precision map positioning, target trajectory prediction, vehicle trajectory planning and motion control and other different parts.
Because of the large number of intermediate sub-modules, the industry has always been keen to open up the end-to-end intermediate links.
Previously, a paper published by the team of Li Hongyang, a young scientist at the Shanghai Artificial Intelligence Laboratory, proposed for the first time a general model of autonomous driving of "UniAD perception and decision-making integrated scheme".
The paper won the "Best Paper of CVPR 2023", which is the first time in 40 years that CVPR (International Conference on Computer Vision and Pattern Recognition) has awarded the best paper in the field of autonomous driving.
In the past, autonomous driving systems used artificial rules and explicit connections to open up submodules.
Today, UniAD uses a huge Transformer network to unify tracking, mapping, trajectory prediction, and occupation raster prediction, and uses Planning that does not rely on HD maps as an end target output.
If FSD V12 really adopts an end-to-end solution, then this will also be the first time UniAD has landed.
But why do people think this is not very reliable?
Because although end-to-end eliminates intermediate links and becomes more efficient, this move also means that AI models will become a black box.
Once an error occurs, it will be difficult for the manual to find the corresponding module to complete the adjustment and optimization.
Conversely, in the past, the form of series of submodules, although relatively inefficient, ensured that intermediate processes could be traced, thus providing the possibility of individual optimization of individual modules.
In the final analysis, eliminating intermediate processes to improve efficiency is not the ultimate goal, the most important thing should be to achieve 100% safety rate of intelligent driving systems.
Wu Xinzhou has previously said that he believes that "motion control" is a mathematical optimization problem, which will be solved based on artificial rules for a long time.
"It's silly for you to throw away all the original stuff."
At that time, Wu Xinzhu said categorically, "At the control level, we will definitely not go up to the end-to-end model. ”
As for the planning part, Wu Xinzhu believes that the road of using large models to improve the ceiling of intelligent driving ability can be passed, and Xiaopeng is already doing it.
Compared with Wu Xinzhu's caution, Musk is undoubtedly more bold.
"Just because there's no line of code doesn't mean it's uncontrollable, now just by having the data, you can still be in control," he says. ”
Whether FSD V12 is really that reliable is just a matter of seeing.
Through this live broadcast, Tesla once again interpreted the FSD's strength far ahead.
Being able to achieve such a high degree of intelligent driving performance on a chip with a computing power of only 144TOPS has fully demonstrated the outstanding ability of Tesla's algorithm.
At the same time, Musk once again expressed his views on the visual route, he believes that lidar is not the best choice, and the combination of digital neural network + camera is better.
With these two points, it has somewhat slapped the face of the Chinese car companies that now claim to be "far ahead".
At the level of engineering capabilities, Tesla's strength is undoubted, but to say that FSD has a real opportunity to enter China in the future, whether its performance can be so outstanding, there may still be a certain gap.
After all, the difference between China's road conditions and North America's is too great, and the urban generalization challenge encountered by FSD in China is not small.
Wu Xinzhou also said that if FSD wants to land in China, it is necessary to set up a research and development team of hundreds of people, and then do a large number of localization tests like Chinese car companies, and improve the experience one by one.
He believes that "the road we take, Tesla is indispensable." ”
It is worth mentioning that in mid-August, Tesla China said that it had established a data center in China, and all the data generated by the sale of vehicles in the Chinese mainland market will be stored in China.
In addition, it is also reported in the industry that Tesla plans to set up a data labeling team in China, with a size of about 100 people, to prepare for training FSD algorithms.
The pace of FSD landing in China should be getting closer and closer.
While accelerating the landing in China, Tesla FSD is also moving towards deeper capabilities.
Just last month, the Dojo supercomputer has been officially put into production, according to the official plan, Tesla's computing power will enter the global TOP 5 in February next year, and the total computing power will reach 100 Exa-Flops in October next year.
In terms of the scale of computing power, Dojo can help Tesla process massive video data and accelerate the iteration speed of the FSD system.
It should be noted that the computing power battle is still a "burning money" battle.
Tesla said it will invest more than $2 billion this year and next to enhance its computing power to meet the needs of developing self-driving technology.
How many Chinese car companies can follow up with such a costly war?
How many new forces can continue to endure even greater losses?
Because of this, it is probably not difficult to understand why Xiaopeng is willing to "sell itself" to Volkswagen and Didi, and Huawei also chose to sign a deepening strategic cooperation agreement with Moutai.
At this moment, the intelligent driving battle has advanced to the deep water area.
On the one hand, the NOA part of the urban area dominated by Chinese car companies, "how much to open the city" has become the main focus in the remaining months of 2023.
On the other hand, it is the deep optimization of algorithms led by Tesla, as well as high-dimensional gameplay that expands the scale of supercomputer computing power.
Although the priorities of the two sides are not exactly aligned, the difficulty and the degree of money burned should not be underestimated.
As for whose choice is more productive, this may not be able to compete until FSD really enters China.