Cao Yuan is from the Vice-Pilot Temple
智能车参考|公众号AI4Auto
What Musk said also scored context.
Luminar, a star company of lidar, has just revealed that Tesla is its largest customer.
Musk ridiculed the doomed failure of relying on lidar and the sensor as a crutch for the autonomous driving system, but he quietly became the big brother of Luminar.
And Musk has just made another surprising statement, revealing that FSD mileage is almost useless for model training.
What's going on?
Luminar: Tesla is the number one customer
Luminar, which has just announced a 20% layoff, has emerged as more reasons for the layoffs have emerged with the release of its first-quarter earnings report.
Judging from the several key indicators given, Luminar's performance in this quarter can be said to be unsatisfactory.
Revenue increased by 45% year-on-year to US$21 million (about 150 million yuan), and it was still in the red, with a net loss of US$126 million (about 910 million yuan), a year-on-year decrease of 14%.
However, compared with the previous quarter, revenue fell by 5%, which was explained in the earnings report because of the decrease in sales of lidar non-automakers.
But Luminar later added that this reduction in sales was partially offset by Tesla's sales.
And more importantly, Tesla is also Luminar's largest customer in the quarter, accounting for more than 10% of sales.
It seems that Musk is well versed in the essence of "integrity of the mouth", and complaining about lidar is complaining, and it should be used.
This also proves the rumors in 2021, when it was reported that Tesla and Luminar had reached a partnership to develop autonomous driving with Luminar's lidar on the Model Y.
At that time, there was speculation that it was to test FSD, and some puppets encountered Model Y with LiDAR "big flower pot" road test.
It seems that the cooperation between the two sides is indeed a thing, and it is the kind of long-term cooperation.
It just so happens that FSD has recently achieved intelligent driving with three domains, and is ready to continue to iterate to a higher level of capabilities.
But Musk recently suddenly said that FSD mileage is of little help for training.
As we all know, the iteration of the capabilities of the autonomous driving system is inseparable from the accumulation of mileage data from actual road tests, but Musk began to say that it is useless.
And what is Musk's reason?
Elon Musk: FSD mileage plays a small role
After the full push of FSD V12, the popularity in the field of autonomous driving remains high. A Tesla investor analyzed Tesla's leadership in technology and business model in detail with others in his video.
Highlights of their discussion include, but are not limited to:
1. The transformative nature of Tesla's FSD V12: from the traditional code-based autonomous driving model to the end-to-end neural network model;
2. Tesla's advantages in data collection and business model: data is the core of training and improving the training of autonomous driving models, Tesla has a large fleet, and unlike Waymo and Cruise, Tesla collects data by selling cars, rather than using expensive test vehicles and engineers to collect data and train models;
3. After switching to neural network, FSD V12 has significantly improved the ability improvement speed compared with the previous version;
Subsequently, Musk affirmed the analysis in the video, saying that the analysis was very accurate, but corrected his opinion on the data.
Only about 1 in 10,000 miles is useful for FSD training.
The limiting factor hindering the progress of FSD was the computing power of AI, but now it is the verification capability link, because user intervention is very rare nowadays.
Although at first glance this is to say that the FSD mileage is useless, after all, after running 10,000 kilometers, there may only be one kilometer of data useful.
But it's actually saying that only data that contains special cases, or even bad decisions, is the key to training AI.
For example, the system judges that the front can continue to move forward, but the user intervenes and applies the brakes, which means that the system is still insufficient and needs to be trained in this scenario.
But according to Musk, FSD now allows users to actively intervene less and less, so there is no data that can be used for training.
Although Musk seems to be advocating the uselessness of mileage on the surface, in fact, he should still want to say that the high-quality data required to train autonomous driving models is very scarce.
Looking back now, whether or not to use lidar may not be so important.
Because according to the FSD system, what is more critical is the ability of the neural network to process data and make decisions, and as Musk said, the link of verifying the system's capabilities in the actual situation.
Perhaps this is why Musk is in a hurry to launch RoboTaxi. Because after the lifting of the ban on RoboTaxi in San Francisco, there have been a lot more accidents in the Cruise next door, which means that there are more and more complex road conditions than before.
And these are all high-quality data that can be used for training.