Elon Musk Drops the Pretense: The New Version of FSD Clearly Aims for "Human Removal," and Full Self-Driving OTA Is Imminent

This might be the last generation of FSD that can't achieve "fully autonomous driving":

Just now, Tesla pushed the FSD version V14.1.7:

KOL users uploaded a large number of real - world tests immediately. Some people were amazed that FSD really has "consciousness" as Musk said and can understand the scene environment:

Some people were shocked by FSD's operational ability in extreme scenarios:

Of course, there were also continuous complaints. Some people wondered why FSD seemed to be unable to handle seemingly simple scenarios properly...

In response to all these, Musk replied uniformly: 95% of the problems have been solved. In the next V14.2 version, fully autonomous driving will be comprehensively realized.

It has been 10 years since Musk first promised "fully autonomous driving".

Improvements and remarkable features

One month ago, FSD V14 was updated for the first time. The biggest feature was "integration", integrating various parking and shuttle functions with the intelligent assisted driving experience on open roads.

That is the door - to - door driving ability that domestic users are more familiar with. Now, V14 can handle a complete door - to - door journey.

With the minor update of 14.1.7, the intuitive experience is that the handling in various scenarios is smoother.

For example, at a regular intersection without extreme corner cases, FSD now doesn't hesitate at all. It also won't make unnecessary early deceleration and avoidance when it has confirmed that it's safe to pass:

This shows that the proportion of "rules" in FSD is getting lower and lower.

When encountering pedestrians crossing the road, FSD is obviously more polite than other human drivers on the road. It will stop in advance at a relatively far distance and wait for pedestrians to pass:

Experienced drivers may think this strategy is too cautious and reduces traffic efficiency. But don't worry. The FSD V14 version has a style selection function. There is a "Mad Max" mode in which the system will change lanes frequently and actively to pursue efficiency.

The performance in game - theoretic scenarios may better reflect its level:

When entering the highway and the large vehicle beside doesn't give way, FSD still chooses to slow down actively and then merge into the traffic after passing the vehicle.

It's not difficult to handle this scenario safely. Most domestic players can also do it. But users who often use intelligent assisted driving systems can feel the difference in FSD's "thinking" -

After merging into the highway from the ramp, instead of staying in the right - most lane for observation, FSD tries to change lanes directly to the fast lane.

In contrast, many domestic players' solutions still force the vehicle to drive into the right - most lane first after merging into the highway or making a turn and keep it there for a few seconds before changing to the fast lane, even if the road is empty.

From the perspective of "predictability", FSD is more likely to be used frequently by users.

The 14.1.7 version also performs more smoothly when passing through narrow spaces:

Even in extreme scenarios like passing through width - limited bollards with only a few centimeters of space on each side, FSD can now handle it independently, which the previous V13 version couldn't do at all:

There's something even more amazing:

In an extremely narrow road, FSD independently planned a complete set of 14 - point U - turn process!

For a 3 - point U - turn on a narrow road, most current industry solutions need to be optimized separately for the scenario, and the "3 - point" process is fixed. It can't be activated when the road is too narrow.

From this perspective, FSD has shown certain AI Agent capabilities.

Does this mean that the native non - plug - in VLA system is essentially the same as an AI Agent?

What problems remain unsolved

Since Musk himself said that 95% of the problems in V14 have been solved, what about the remaining 5%?

The most frequently reported problem is the optimal route selection problem.

Some users found during real - world tests that in some scenarios, especially in non - standard environments such as parking lots and charging stations, FSD's route selection was often very strange. For example, in this scenario:

When leaving the Supercharger, the vehicle wrongly turned into a small road, pulled over and then the system exited.

This problem is not unique to the V14 version. It has been a "chronic problem" of FSD.

The most complained - about problem is FSD's hesitation when starting at intersections. It seems to "creep" forward hesitantly for a few steps before really starting, with obvious jerks in between.

There are also other problems. For example, when turning, the system turns on the turn signal far away from the intersection, making other vehicles think it's going to change lanes...

Instead of replying to each feedback, Musk simply said that the next version, V14.2, would definitely solve these problems and achieve fully autonomous driving before being pushed to users.

Maybe Musk will finally fulfill his promise of "fully autonomous driving", even if it's 10 years late.

After FSD 14.1.7, will fully autonomous driving be achieved?

From a functional perspective, the update of FSD V14, including the 14.1.7 version, was rather uneventful:

A new arrival option was added, allowing users to choose where FSD should park: in a parking lot, on the street, in a lane, in a parking garage or by the roadside.

The operation of pulling over or giving way to emergency vehicles (such as police cars, fire trucks, ambulances) was added.

Navigation and path - planning functions were added to the vision - based neural network to handle road blockages and detour routes in real - time.

Additional speed modes were added to further customize driving style preferences.

The handling of static and dynamic garage doors was improved.

The recognition of road debris (such as tires, branches, boxes) was improved.

The handling of the following situations was improved: unprotected turns, lane changes, vehicle cut - ins and school buses.

The ability of the FSD management system to handle failures was improved, and it can recover smoothly from degraded operation, thus enhancing reliability.

It seems to be a continuous optimization of existing functions, but the most crucial statement is:

Navigation and path - planning functions were added to the vision - based neural network.

That is to say, the biggest difference between the major version V14 of FSD and the previous versions is the completion of the one - stage end - to - end transition.

Musk revealed that the number of parameters is at least 4.5 - 10 times that of the V13 version. By comparing Musk's statements for each version, it's reasonable to assume that the number of parameters of the underlying model of FSD in the V14 version is at least tens of billions or even hundreds of billions.

This means that currently only Tesla users with HW4.0 hardware can experience V14.

The "hardware - level" update of 14.1.7 is also an adaptation for all HW4.0 hardware models, including the Cybertruck.

It shows that Tesla has solved the problem of model optimization with limited computing power. After all, with the same hardware foundation of hundreds of TOPS, most industry players can run models with billions to tens of billions of parameters at a frequency of about 10Hz.

The advantages of V14 are beginning to be fully realized, proving that the role of self - developed full - stack software and hardware is indeed irreplaceable. More players may bet on it in the future.

There is another statement in the update notes that is worth noting:

The ability of the FSD system to handle failures was improved, and it can recover smoothly from degraded operation, thus enhancing reliability.

This is the first time in Tesla's history and also the first time in the entire automotive industry to have such a statement.

"Breakpoint recovery" - this is completely in line with the ability requirements and language system of L4. After all, for L2, the "breakpoint" doesn't need to be handled by the system. Humans can take over.

Starting from V14.1.7, FSD has clearly shown its intention to iterate towards a "safety - driver - free" direction.

In other words, Musk has no intention to argue with those who are complacent about L2, those who are on the fence about L3 or those who are unmoved by L4. He simply doesn't want to be in the same camp with any of them.

The "upgraded" route of autonomous driving that Musk believes in and Tesla represents, or more accurately, general autonomous driving with large models, is about to undergo the most crucial large - scale verification.

For the L4 track, the previous launch of Tesla's Robotaxi may have just been an "appetizer". The real impact is just beginning: Tesla is the first to blur and remove the barriers between L2 and L4.

The two different commercial implementation routes of Robotaxi may start a "race against time".

Actually, only one month has passed since FSD V14 was first pushed to users until the update of V14.1.7.

V14.1.7 is currently being pushed in a small - scale test, mainly to media KOL users and long - term volunteer "testers" among Tesla owners, with a total of more than 200 vehicles. An interesting phenomenon is that they found the version number jumped directly from V14.1.4 to V14.1.7.

This shows that versions 1.5 and 1.6 were only used for internal testing and not made public, which also proves from the side that FSD is now being iterated almost once a week.

This also means that if there are no accidents, the push of FSD V14.2 with "fully autonomous driving" promised by Musk may happen within one or two months.

One more thing

What kind of impact will the "fully autonomous" V14.2 have on the tech and automotive circles in the near future?

Andrej Karpathy, the former head of AI at Tesla and an AI expert whom Musk praised as a "tested brother and comrade - in - arms", bought a latest Model X with HW4.0 hardware (V13 version) more than two years after leaving Tesla. He was shocked by FSD and directly wrote a long article to praise it: