Photo: Not a Tesla App, @BLKMDL3 via X,
After installing the FSD V12.1 to employee vehicles before Christmas, Tesla is rolling out the software to over 15,000 vehicles that are part of the wave1 testing group. Unlike previous rumors, the twelfth iteration of Tesla's self-driving software is still in Beta and will remain so for the foreseeable future.
This summer, Elon Musk emphatically announced that version 12 of Tesla's Full Self-Driving software "won't be beta." This version was supposed to change everything thanks to an entirely new architecture, being the first "end-to-end AI." Even the most optimistic among us thought that Musk was jumping the gun again, announcing a fait accompli when, in fact, FSD was still very much a work in progress.
Dropping the Beta from its name seemed unrealistic, as every piece of software needs thorough testing before being deployed to customers. Indeed, FSD V12 might drop the beta moniker at some point, but not necessarily in its early builds and not for actual beta testing. This proved true once Tesla released the first FSD Beta V12 build to employee vehicles on November 24. This is similar to how FSD Beta V11 was released on November 11, 2022.
The software has since made progress, with V12.1 already passing the first testing phase. The point one update is now rolling out to a broader group of Tesla employees, known as the "wave1" testing group of personally owned Tesla employee vehicles. This means that this version is mature enough to attempt more comprehensive testing on public roads. However, since the wave1 testers are bound by strict non-disclosure agreements, it will take some time before we find out how FSD V12.1 is doing or even what's in the release notes.
Based on the limited information at hand, the FSD Beta V12.1 upgrades the city streets driving stack to a single end-to-end neural network. This is what Elon Musk meant by "from images in to steering, brakes & acceleration out." This change makes over 300,000 lines of C++ code obsolete, streamlining the software and making it behave more human-like.
Right after the wider rollout was confirmed, Elon Musk offered an interesting insight. He thinks the FSD Beta V12.1 works very well in California, but it needs more training in heavy precipitation areas. This is intriguing, considering that FSD Beta has performed consistently well in the rain. Whatever the reason, Tesla thinks the software needs more testing before being released to customer vehicles. Considering this is an entirely new architecture, it's a wise choice.
Switching from rules-based software code to network-path-based software is a huge step, which will allow the software to "learn" and base driving decisions on this acquired knowledge, the same as humans do. Until now, Tesla has been computing-constrained as far as neural network training is concerned, as Musk revealed in August. However, with the Dojo supercomputer coming online and a significant expansion of the Nvidia-based supercomputer already completed, FSD Beta is advancing much faster.
This is already obvious with the V12.1 deployment. Last year, Tesla rolled out V11 to specialized groups in November, to team leaders in February, and to the wave1 group in March. With V12, the same path took less than a month, so I expect a V12 build to get to customer vehicles much sooner, possibly in January. Unless, of course, the NHTSA has some "one more thing" moment as it did with V11 in January.
Dropping the Beta from its name seemed unrealistic, as every piece of software needs thorough testing before being deployed to customers. Indeed, FSD V12 might drop the beta moniker at some point, but not necessarily in its early builds and not for actual beta testing. This proved true once Tesla released the first FSD Beta V12 build to employee vehicles on November 24. This is similar to how FSD Beta V11 was released on November 11, 2022.
The software has since made progress, with V12.1 already passing the first testing phase. The point one update is now rolling out to a broader group of Tesla employees, known as the "wave1" testing group of personally owned Tesla employee vehicles. This means that this version is mature enough to attempt more comprehensive testing on public roads. However, since the wave1 testers are bound by strict non-disclosure agreements, it will take some time before we find out how FSD V12.1 is doing or even what's in the release notes.
Based on the limited information at hand, the FSD Beta V12.1 upgrades the city streets driving stack to a single end-to-end neural network. This is what Elon Musk meant by "from images in to steering, brakes & acceleration out." This change makes over 300,000 lines of C++ code obsolete, streamlining the software and making it behave more human-like.
Right after the wider rollout was confirmed, Elon Musk offered an interesting insight. He thinks the FSD Beta V12.1 works very well in California, but it needs more training in heavy precipitation areas. This is intriguing, considering that FSD Beta has performed consistently well in the rain. Whatever the reason, Tesla thinks the software needs more testing before being released to customer vehicles. Considering this is an entirely new architecture, it's a wise choice.
Switching from rules-based software code to network-path-based software is a huge step, which will allow the software to "learn" and base driving decisions on this acquired knowledge, the same as humans do. Until now, Tesla has been computing-constrained as far as neural network training is concerned, as Musk revealed in August. However, with the Dojo supercomputer coming online and a significant expansion of the Nvidia-based supercomputer already completed, FSD Beta is advancing much faster.
This is already obvious with the V12.1 deployment. Last year, Tesla rolled out V11 to specialized groups in November, to team leaders in February, and to the wave1 group in March. With V12, the same path took less than a month, so I expect a V12 build to get to customer vehicles much sooner, possibly in January. Unless, of course, the NHTSA has some "one more thing" moment as it did with V11 in January.
Google News