The highly anticipated Version 12 of Tesla FSD software is being rolled out to employees, an important first step before the release reaches the public.
Overnight, ‘not a tesla app‘ posted an exclusive story that FSD v12, build number 2023.38.10. They sighted a ‘trusted source’ which we wouldn’t typically put too much confidence in, however, Musk has since replied on X, confirming this is indeed true.
Musk also replied to a post by Whole Mars Catalog, suggesting that V12 could be the moment that changes everything. Musk replied saying simply, ‘it feels human’. While there may be a day when autonomous cars are better than humans in every way, right now our hope is that the car is indistinguishable from other human drivers (the good ones).
On November 13th, Elon Musk had replied to a question about the timeframe for V12, to which Musk replied About 2 weeks’. Two weeks has been a long-running joke in the Tesla community, as a placeholder for the ever-changing timeline for FSD, apparently, this time Two weeks, meant two weeks (although to employees first).
Version 12 has been in the works for some time at Tesla and the fact they are kicking off a rollout to employee vehicles, suggests its performance is exceeding that of V11, at least in a majority of circumstances.
The last and only time we’ve actually seen FSD V12 was during Elon Musk’s livestream back in August. While most of the drive looked fairly similar to V11, as well as the visuals, there was one big new feature in V12, the ability to pull over to the gutter to park at the of a route. Given the new functionality and new architecture, users will need to reset all expectations of how the car operates.
With V12, FSD will move to use end-to-end neural networks which effectively learns from massive amounts of video training data, made up of millions of clips of data from ‘good drivers’ and the model effectively learns how to drive, then can take live inputs from the car its running on and navigate the dynamic world around it, without explicitly being told what each object is.
End-to-end neural networks are a type of machine learning model that takes raw sensor data as input and produces desired output actions directly. This is in contrast to traditional autonomous driving systems, which use a pipeline of separate algorithms to process sensor data, make decisions, and control the vehicle.
Tesla is transitioning to an end-to-end approach for FSD with the release of V12. This means that a single neural network will be responsible for all aspects of autonomous driving, from perception to planning to control. This approach is expected to have several advantages, including:
- Improved performance: End-to-end neural networks can learn more complex relationships between sensor data and desired actions, which can lead to better performance in a variety of driving situations.
- Increased adaptability: End-to-end neural networks are better able to learn and adapt to new situations, which can make them more robust and reliable.
- Simplified development: End-to-end neural networks are simpler to develop and maintain than modular systems, which can reduce costs and accelerate development.
V12 being released to employees is the first step in an important release schedule.
- Early public beta testers
- Point releases to resolve bugs/improve
- International release
- Drop the Beta label.
Countries like Australia that have been waiting on FSD now have their best chance at seeing Tesla’s Full Self Driving software rolled out to countries outside of the United States and Canada. Given the change in architecture, this path should now be accelerated.