Right now, version 12.3 of Tesla’s Full Self Driving software is rolling out across millions of vehicles from North America to Canada, showing that Tesla has confidence the technology is ready for prime time.
In the last week, the company announced that all compatible cars in NA would receive a 1 month free trial of the software. This has never been done before.
Over the years of FSD development, it has often been thought that Tesla was limiting the number of people in the FSD Beta program, but with this now rolling out to people with the widest range of skills and abilities, it seems Tesla is comfortable it’s ready.
The latest version of FSD is version 12.3.3 which includes the following in the release notes.
Under your supervision, Full Self-Driving (Supervised) can drive your Tesla almost anywhere. It will make lane changes, select forks to follow your navigation route, navigate around other vehicles and objects, and make left and right turns. You and anyone you authorize must use additional caution and remain attentive. It does not make your vehicle autonomous. Do not become complacent.
Tesla is clearly changing the language around FSD Beta as it moves from a group of enthusiasts, to the broader public. Globally Tesla just shipped it’s 6 millionth car, many of those in North America and Canada, with the vast majority featuring the HW3+ necessary to access FSD Beta (also now known as FSD Supervised).
Side note Tesla Australia also passed a milestone this week, speeding past 100,000 sales locally.
Tesla’s been clear from the start, the goal is to deliver autonomy, Full Self Driving, however right now, the car does still leverage the driver as a backup.
Watching the vast array of YouTube videos from users experiencing V12.3, it’s clear the technology is getting closer than ever. It’s certainly not perfect with some users highlighting issues where they had to disengage, but compared to previous builds the feedback has been overwhelmingly positive.
What started as one or two drivers reporting zero intervention drives, we’re now seeing many, many people share that they effectively watched the car drive from point A to point B safely.
While V12 may seem like just another build like V10 and V11 that came before it, in reality, it’s a complete architecture change, moving from human labelling and human written code, now using millions of video clips of good driving behaviours to create an AI model that effectively learnt how to drive like a human.
There are still some edge cases that need to be overcome, but the trajectory now seems clear, find examples of edge cases experienced by the fleet, then feed the model lots of examples of what the vehicle should do when faced with that style of environment. Where Tesla doesn’t have a large set of video data around an edge case, they now have the capacity to generate video variants of that edge case to train the model on.
This is now a pipeline of development that runs as fast as the compute capabilities at Tesla allow. Recently Elon Musk shared that the constraints they were facing (in processing so much video) have now been largely met by capacity. Estimations were that they would invest to the tune of $2 billion in the last year to resolve this, much of which went to Nvidia as their hardware is particularly good at processing this training data.
So with V12 demonstrating significantly improved driving behaviours, often in really complex environments, the question is, how long will the driver need to have their hands on the wheel? Are we approaching the point where the car understands the environment well enough and monitors the driver well enough, to remove the steering wheel nag?
The next question is, when does Tesla feel confident that this is ready to roll out to other countries (like Australia) that have also been sold the FSD software package, but as yet do not have the additional functionality provided by Beta, of navigating city streets.
It is difficult to tell with any level of accuracy, but if this next month goes by without any serious crashes as a result of FSD, then I think we should feel confident about seeing it in Q3 or Q4 this year.
The road to full autonomy, where the driver is allowed to text and drive, or be drunk, or in the back seat is likely still some time off, but 12.3 is Tesla’s best demonstration to date that the car can drive using computer vision alone and this is ultimately now a compute problem, assuming you can capture or generate the right training data.
The reason Tesla’s approach is so significant in the industry is that this software is going out to people across the whole of united states and can be used anywhere. Unlike Waymo that has serious geographic restrictions, GM’s Super Cruise or Ford’s BlueCruise, this goes way beyond the highway and tackles all aspects of driving. It’s also worth noting that FSD is available in the Model 3, Model Y, Model X and Model S (not yet available for Cybertruck or Semi), while other automakers reserve their best autonomous technology for luxury vehicles only, meaning it’s only available to a small number of wealthy individuals.
The threshold of being ‘safer than a human’ is a very difficult one to quantify, given the wide variance in human driving ability, but being able to constantly monitor the environment around you in all directions gives the machine a significant advantage. It also doesn’t get distracted, doesn’t get tied and when we combine that with it’s ability to navigate the roads and highways successfully (with supervision), it certainly feels like we’re reaching that point.
Below are some hands-on videos of 12.3.x in action. As always, I recommend you watch videos from a variety of content creators.
