Tesla FSDV12 can also be used by ordinary car owners. Tesla FSDV12 is launched in the United States, not just Tesla internal employees,Ordinary car ownersThere is also a chance to experience end-to-end autonomous driving.Say goodbye to traditiondriven by rules,Driving decisions are generated by AI algorithms. Although the FSDV12 update can be called"leap", however, in the actual measured video, Version 12, which still has the suffix Beta, cannot be called"The ultimate answer".

Tesla FSDV12 launched in the United States

FSDV12Beta is launched in the United States and does not rely on lidar. It uses data-driven AI algorithms and media demonstration videos to cope with most scenarios.

Turn left after going uphill without protection on a rainy night:


Wait for the children to cross the street before walking:


Parking on the side of the road:


Avoid sudden braking of the vehicle ahead:


Overall, the capabilities of the purely visual FSDV12Beta have been greatly improved. It can accurately identify irregular obstacles and restore some details very well.

Program compatibility has also been improved. This time, not only the original HW3.0 hardware solution has been pushed, but HW4.0 has also been pushed.

However, there are still shortcomings, such as not relying on lidar, and the current purely visual solution does not perform well enough in bad weather.

Bloggers of the first batch of tests said that FSDV12Beta took over many times when tested in heavy rain.


This is also consistent with Musk’s reply to netizens about a month ago. At that time, netizens urged Musk on social platforms. Musk responded that the FSDV12 program performed well in California, butYou also need to train more in areas with a lot of rain..


Therefore, the currently pushed v12.1.2 version obviously has insufficient capabilities. The blogger who posted the test video later said:

FSDBetav12.1.2 (the latest version number) cannot realize automatic driving.


This may be why the pushed FSDV12 still has the suffix Beta (test version). Musk has previously stated publicly that FSDV12 will not be a Beta version.


And although this time it was indeed pushed to ordinary car owners, not just Tesla internal employees, the number of car owners pushed is still relatively small. Tesla is still more cautious about this. Maybe it will be pushed to all employees later, and the "Beta" may be removed after the official public beta.

If it were to land directly in China with its current performance, I am afraid that many people would bid farewell to the "FSD worship".


However, despite its shortcomings, the update to FSDV12 can still be calledleap, opening up a new path for the development of autonomous driving.

What is the difference between V12 and previous FSD versions?

Tesla stated in the latest OTA annotation that V12 is trained with massive video data.end-to-end neural network, replacing more than300,000 lines of C++ code.

no longerrules drivenYes, butdata drivenYes, this is a major evolution of FSD.


so-calledrules driven, that is, in the past, traditional self-driving programs used programs to set rules and if-else logical judgments when making decisions based on perception.

CornerCase (edge ​​scenarios) are endless. It is difficult to write a program to cover as many scenarios as possible. If you encounter special circumstances, just add a section. If things go on like this, the redundancy of the code and the complexity of the program will continue to increase. In the end, the system will inevitably be called a "shit mountain" by programmers.

turn nowdata driven, use a large amount of video data to train the neural network, and continuously improve the model to become more like a human veteran driver.

According to a Tesla employee who participated in the internal testing of FSDV12 earlier, problems that were difficult to optimize in the past can be easily solved - (the model) can just imitate humans.


FSDV12 is not like the past autonomous driving, which is divided into three modules: perception, planning and control, and is implemented in a unified body.end-to-end, this end inputs the data sensed by the surroundings, and the other end outputs the driving decision.

What are the benefits of end-to-end neural networks?

From a research and development point of view, it used to be divided into perception, planning and control, which ultimately led to the realization of driving the car well. However, the optimization goals of the three modules are different. Each module will produce bugs or errors. The accumulation of these errors will lead to problems with the final output results.

It’s like the Chinese New Year is about to come and we make and smash dumplings together, mix the stuffing, mix the noodles, and make the dumplings. The ultimate goal is to make this plate of dumplings. However, the goals pursued by the fillings, noodles, and dumplings are different, and problems may occur at every step, resulting in the dumplings being overcooked.


From the perspective of commercialization, each module has its own system and mechanism, which means that teams and funds must be invested in each module, but it may not necessarily be able to obtain upgrade and iteration efficiency.

After converting to a single end-to-end model, the engineering complexity is not as high as before. It only needs to be optimized around one final goal and unified and controllable. The backbone of the model can also be shared to improve computational efficiency.

End-to-end also makes it possible to improve programs purely based on data.


Of course, new breakthroughs bring new problems.

In the past, it was divided into multiple modules, which made it more flexible to upgrade and optimize, and the C++ program was more interpretable. From assembly to compilation, the program ran very clearly.

End-to-end autonomous driving is a "black box" with poor explanationability, which increases people's cost of trust in autonomous driving.

But as AI guru He Kaiming previously said in a lecture at the Chinese University of Hong Kong:

The driver you met while taking a taxi is a stranger. Why would you trust him?

You trust him because his brain isexplainable, or do you think one iswell trained,ExperiencedThe driver should be able to drive very steadily?