A thousand calls began to come out,$Tesla (TSLA.US)$FSD v12 (FSD v12.1.2) has been officially launched to users.

“FSD Beta v12 upgraded the urban street driving stack to an end-to-end neural network, replacing more than 300,000 lines of C++ code after millions of video training.” In the update notes, Tesla said so.

In other words, after upgrading to FSD v12, artificial intelligence will replace engineer coding and control vehicle behavior in autonomous driving mode.

The evaluation video by Tesla review blogger Whole Mars shows that in the process of taking over the FSD v12, it will wait for pedestrians to cross the road in front of the zebra crossing, smoothly avoid cars that stop in the driveway, and quickly change lanes after turning left/right; if the car owner's destination is on the left side of a one-way street, the vehicle will also stop directly on the left; when other vehicles parked on the side of the road suddenly leave or stop suddenly while driving, the FSD v12 will also avoid it in a timely manner...

According to the evaluation account, it has been testing the Tesla FSD beta since 2020. From FSD v11.4.9 to today's FSD v12.1.2, “this is the biggest improvement I've seen in FSD Beta,” and FSD v12.1.2 has “achieved significant improvements in all key aspects.”

But at the same time, he also admitted that compared to FSD v11.4.9, FSD v12.1.2 has more time to take over — “FSD v12.1.2 is not ready for full driverless driving.”

Musk has said that Tesla will provide users with a non-beta version of FSD v12, but in the update instructions, FSD v12 has not removed the word “beta”.

A month ago, Musk revealed that since FSD v12 is a brand new architecture, Tesla is undergoing additional testing. It's doing well in California, but it also needs more training for areas with heavy rainfall.

End-to-end is the biggest highlight

Tesla FSD v12 is its most important upgrade, and the end-to-end solution is one of its biggest selling points.

Why choose an end-to-end solution? Musk previously gave more details when connected with Whole Mars, the review blogger mentioned above:

“That's what humans do,” he said, “photon input, hand and foot movement (control) output.” ——Humans rely on eyes and biological neural networks to drive. For autonomous driving, cameras and neural network AI are the correct general decision-making solutions. Although it is difficult for AI neural networks to explain specific details, accordingly, when taking a taxi, human passengers cannot accurately know what the driver is thinking; they can only see the driver's comments.

Almost all of the end-to-end algorithms are built using neural networks. Open Source Securities points out that the advantages are that they are very flexible, have good adaptability, high upper limits, and can drive in a highly anthropomorphic state, greatly improving the driving experience. In Musk's live broadcast of the FSD v12 test drive in August 2023, the overall process was very smooth. FSD v12 is close to the driving habits of human drivers, showing good driving experience and results.

Analysts have soared further. Sophisticated algorithms, massive data, and large-scale computing power are necessary conditions for FSDV12. Among them, data is the top priority for training good algorithms. How to collect massive amounts of effective data, and let the model show corresponding driving behavior in corresponding scenarios through training tests the developer's technical ability. Computing power is the foundation for training autonomous driving models, and Tesla's computing power will reach 100 eFlops by the end of 2024. A well-designed algorithm is also a core part of building an end-to-end algorithm.

Editor/Corrine