Source: Finance Association
① On non-predefined roads, vehicles can be polite to pedestrians, avoid roadblocks, turn at intersections, and choose between two direct roads with fewer cars.
② Through video training data, AI can learn to drive on its own; high-quality data from excellent drivers is the key to training Tesla's autonomous driving.
③ Tesla is about to launch a new computing power cluster, which includes 10,000 Nvidia H100s.
As promised before, Musk used a Model S equipped with HW3 to show the outside world a beta version of the Tesla FSD 12 live broadcast last weekend.
In this 45-minute live broadcast, Musk, sitting behind the wheel and holding his phone, only interfered with the vehicle's behavior once. On a road other than the default road, the vehicle was able to excuse pedestrians, avoid roadblocks, turn around at intersections, and choose one of the two straight roads with fewer cars.
Musk said that FSD 12 can be used offline in unfamiliar environments; if intervention occurs, the system will record and send back to Tesla for analysis.
And about 20 minutes after the live broadcast began,Musk intervened and took over the whole process for the only time.At the time, the Model S needed to go straight, so I stopped and waited for a red light. However, when the left-turn signal light turned green, the vehicle actually started immediately. Fortunately, Musk and the engineers on his side stopped it in time.
After that, Musk said he wanted to “feed” the FSD with more videos of left-turning traffic lights.
Can you “feed” an “AI chauffeur” by posting a video?
In fact, in this live broadcast, when the vehicle decelerated on its own in the speed bump and avoided the scooter riders, Musk emphasized many times,There isn't a single line of code in FSD 12 that artificially sets the vehicle to perform these actions——They haven't been trained to read road signs, and they don't know what a scooter is,FSD 12 accomplishes these actions entirely as a result of extensive video training. Through video training data, AI can learn to drive on its own,“Do things like humans”.
If FSD doesn't make the right decision in a specific scenario, Tesla will invest more data (mostly video) into its neural network training.
Of course,Mediocre and random data is not enough; the data supplied to the neural network needs to be carefully selected. Musk also placed special emphasis on high-quality data from excellent drivers, which is the key to training Tesla for autonomous driving.
“A large amount of mediocre data doesn't improve driving, and data management is quite difficult. We have a lot of software that can control what data the system selects and what data to train.”
For Tesla, a major source of data is its fleet of cars from all over the world. Musk also revealed that Tesla has many FSD test drivers around the world, including New Zealand, Thailand, Norway, and Japan.
Since 2020, Tesla has begun to shift Autopilot decisions from programming logic to neural networks and AI. After 3 years of development, as can be seen from Musk's FSD 12 live broadcast, almost all decision-making and scenario processing has been transferred to Tesla's neural networks and AI.
FSD 11 has more than 300,000 lines of C++ code in its exclusive control stack, while there is very little code in 12.Musk also previously pointed out that vehicle control (vehicle control) is the final piece of the “Tesla FSD AI Puzzle,” which will reduce more than 300,000 lines of C++ code by about 2 orders of magnitude.
Full AI end-to-end driving control
Tesla FSD 12 is itsThe most important upgrade was to achieve full AI end-to-end driving control.
As for why choose an end-to-end solution? Musk gave more details when connecting with WholeMars before Live.
“This is how humans do it,” 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, corresponidentally, when taking a taxi, human passengers cannot accurately know what the driver is thinking; they can only see the driver's comments.
Brokers pointed out that one of the key differences between the end-to-end solution and the previous one is that the traditional modular architecture divides intelligent driving into separate tasks and leaves them to special AI models or modules for processing, such as perception, prediction, planning, etc.; while end-to-end AI is “integration of perception and decision making,” that is, integrating “perception” and “decision making” into one model.
Currently, the vast majority of Tesla training still relies on Nvidia GPUs, and Tesla's own Dojo supercomputing is used as an aid.Since this year, Tesla has spent $2 billion on training.
Tesla is still working overtime,A new computing power cluster is being prepared, which includes 10,000 Nvidia H100s, and is expected to go live this Monday (August 28).It is worth mentioning that this cluster uses Infiniband for connection transmission.Musk went on to say that today InfiniBand is less than GPUs.
edit/lambor