Musk quietly added hardware sensors to the new car, and Tesla officially removed the term "assisted driving" from its FSD introduction: This is autonomous driving.
The "counterattack" of the L2 camp has arrived quietly.
The first shot overseas: Without surprise, it's Tesla.
"Quietly" because just yesterday, Intelligence Vehicle Reference introduced the new version of Model S/X, which added an extremely crucial but unnoticed configuration:
Tesla didn't make a big fuss either, but quietly changed the description of the mass - production plan.
The official definition of the nature of the entire FSD has also changed.
Comprehensive Upgrade of Tesla's New Cars
Can you tell what has been added to the new car?
An exterior front - view camera.
Actually, the previous Cybertruck and the refreshed Model Y also had the same configuration.
At that time, Tesla's official description was:
The front - view camera provides a wider field of view for automatic assisted driving and the Smart Summon function.
With the release of the new version of Model X/S, the official wording has changed to:
The front - view camera provides a wider field of view for Autopilot, self - parking, and the Smart Summon function.
The word "assisted" has been completely removed, and FSD has been re - defined as "automatic".
Autopilot can be understood as autonomous driving or "automatic cruise". In FSD, the "cruise NOA function" where the system follows the navigation autonomously has already covered urban, suburban, and parking scenarios, and the experience is fully automated.
This should also be the original intention of Tesla and Elon Musk. It's just that this time it's more subtle. Instead of using "fully autonomous driving" to describe it, they've subtly made FSD (Full Self - Driving) the name of the product plan to capture users' minds.
From fully autonomous driving, to automatic assisted driving, and now to autonomous driving, isn't this Musk's "salami - slicing" tactic in marketing?
Against the backdrop of the domestic official's clear regulations on L2 promotion, all domestic players emphasize the word "assisted" in their promotions. Now, Tesla is doing the opposite, discarding "assisted" and not caring about "intelligent", directly announcing "automatic". It's quite courageous.
However, it's still unknown whether it will promote in the same way in China.
The question is, behind the change of words, we need to figure out whether Tesla really has the confidence or if Musk is getting over - confident again.
Can Tesla's current plan and technical system support "autonomous driving"?
After several iterations of software and hardware updates, there are actually many versions of Tesla's FSD suite on the road.
We won't discuss the early plans because they were limited by the software and hardware technology at that time (such as the early versions using NVIDIA and Mobileye chips) and couldn't experience Tesla's latest technological achievements.
Starting from the HW 3.0 hardware put into production in 2019, Tesla switched to self - developed computing power chips, self - developed highly integrated SoC + MCU, with a full set of chip design, image recognition algorithms, multi - sensor fusion, and application - layer software development. The single - chip computing power reaches 72 TOPS, and the board - level computing power reaches 144 TOPS.
The first - generation Model 3/Y, including Model X after 2019 and Model S after 2021, are all equipped with HW 3.0. The corresponding sensor plan is: forward - facing millimeter - wave radar + surround - view cameras + ultrasonic radar.
The "refreshed" version based on HW4.0, including the new S3XY, has a computing power of over 700 TOPS. The biggest feature is that Musk removed all millimeter - wave radars and ultrasonic radars at once. The whole car has 8 cameras to handle all driving and parking tasks.
All vehicles in the North American market are of this 8 - camera pure - vision type. In the Chinese market, the millimeter - wave radar has been added back in subsequent versions.
And this update has increased the number of exterior cameras of the pure - vision version HW4.0 from 8 to 9.
So, there are at least 4 versions of the FSD hardware plan for current Tesla models:
HW3.0, HW4.0 with 8 cameras, HW4.0 with 8 cameras + 1 millimeter - wave radar, and the latest HW4.0 with 9 cameras.
However, multiple versions don't prevent the FSD software from being updated and pushed.
The latest version V13.2.9 (pushed in May), the core technological upgrades include
36Hz full - resolution AI video input, native AI4 (i.e., HW4.0) input, 4 - times data scaling ability of the neural network architecture, 5 - times expansion of training computing, 2 - times reduction of latency, optimization of speed strategies on urban streets and highways
In terms of functions, FSD has enhanced the understanding of non - standard scenarios during U - turns and parking, improved reward prediction to reduce collision risks, increased the efficiency of camera cleaning, upgraded the control for a smoother FSD experience, and added the function of autonomous avoidance of temporary road closures.
The official has revealed that it will also add the ability of fixed - point autonomous parking in different garages, lanes, and parking spaces in the future.
These updates fully cover all configurations and models of HW3.0 and HW4.0.
Tesla's entire autonomous driving R & D team (excluding hardware) has only about a few hundred people, far fewer than the domestic teams, which usually have at least a thousand people, or even four or five thousand.
Now Tesla is very stingy with words or even doesn't mention its technical route at all. But from the current OTA scope, R & D team, and the performance of FSD itself, we can actually find some clues.
At the back - end, Tesla must have a unified autonomous driving base model, which may be much larger in scale than any domestic player.
Previously, XPeng revealed that its base model has 72 billion parameters and a computing power cluster of 10 EFLPOS. While Tesla's computing power reserve, Cortex, has a built - up scale of about 30 EFLPOS...
The ultra - large - scale base model may very likely have the VLA architecture. The way to solve corner cases has also shifted from relying on a large amount of data to the more crucial cognitive and understanding ability of the large model: Where am I? What's happening around me? What measures should I take?
After all, when we take the driver's license test, we don't pass by imitating other drivers' behaviors. Instead, we learn the simplest driving operation knowledge based on understanding the road and vehicle conditions.
Musk's "first - principles" are reflected not only in the sensor level but also in the algorithm architecture level.
The base large model needs to have basic cognitive ability and driving skills through pre - training, and then ensure safety and reliability through targeted reinforcement learning. Finally, the trained large model needs to be "distilled" into a "small model" that can be installed on the vehicle.
Of course, this "small" model on the vehicle is also very large and requires at least several hundred to a thousand TOPS of computing power support.
As for the sensor and computing power configurations of different models, only minor adjustments and adaptations are needed. For example, HW 3.0 is further distilled to get a model that retains most of the basic abilities but can run with 144 TOPS.
Actually, it's easy to find that the base large model with the ability to cognitively understand the real world has far more applications than just intelligent assisted driving.
It's reported that the HW 5.0 hardware to be unveiled next year has a computing power of thousands of TOPS. This may be the fundamental reason why Tesla dares to remove "assisted" and directly say "autonomous driving".
Domestic players are also following this route. For example, XPeng's self - developed Turing AI chip, NIO's Tianji chip, and Huawei's ADS 4.0 plan with four lidars but without specific computing power disclosure...
In essence, it's the same as Tesla's "autonomous driving", but it's called "L3" in China.
Two more things
Today are two "small things".
Tesla's Robotaxi hit the road a few days ago and is regarded as the "biggest catfish" and a possible "disruptive factor" in the L4 track. Because Tesla's Robotaxi is completely different from traditional L4 players like Waymo, Apollo, and Pony.ai, which use a large number of redundant sensors + partial rule - based algorithms to ensure safety and reliability. Instead, it uses ultra - large computing power + models, betting on the intelligent emergence of AI drivers.
The model is the Model Y. When netizens asked if it could be used on mass - produced cars, Musk only replied that "the software is a new version". It's very likely that the hardware version of Tesla's Robotaxi is exactly the same as the mass - produced car, and the software may be the long - awaited FSD V14.
Policy - wise, North America is also about to give the green light to Tesla's Robotaxi: It's reported that NHTSA is considering allowing "autonomous driving companies" like Tesla to produce vehicles that don't meet the current motor vehicle regulations, such as vehicles without steering wheels.
The mass - production and implementation process of traditional L4 is just beginning to show signs of hope, but the counterattack of the L2 camp is more intense than ever. If a large number of "upgraded" Robotaxis are put into use and work well, how will the L4 camp respond?
Finally, FSD is finally going to be widely promoted in China.
Previously, only the test - drive vehicles of the refreshed Model Y in some stores had it. Since June, FSD has begun to cover more models.
It may also introduce a more flexible monthly subscription model to lower the experience threshold.
Once the FSD experience can be converted into word - of - mouth and then become a selling point, Tesla's current sales slump should be reversed.
The question is, how much is appropriate for the monthly subscription? The one - time purchase is 64,000 yuan. If it's a friendly price for the monthly subscription, 1,999 yuan?