Tesla has sold their Full Self Driving software package internationally and while many of the features listed on the website are available, but Autosteer on city streets is only available with FSD Beta.
The world outside of the United States and Canada has watched and waited patiently to get hands-on with Beta, but as yet, it hasn’t arrived.
The official timeline remains a bit of a mystery, but there are signed that Tesla is gearing up for a broader rollout, hiring ADAS testers in other markets (including Australia), we’ve also seen them deploy software builds that include the Beta code. We understand that Tesla staff are actively testing Beta in Australia, but as yet, we haven’t seen it reach consumer cars.
The FSD Capability is currently priced at A$10,100 with no subscription option, those who have invested the money are keen to know when.
The exciting development today was the list stream of FSD V12, our first look at the new build, even skipping the normal route of release notes being leaked.
This version is not simply an iteration of what’s come before, but a major architectural change.
This change moves the software from being a number of neural nets to manage different parts of the driving operations, to a single end-to-end solution that simplifies everything.
This new approach takes video training data from Tesla’s fleet of vehicles and understands how humans drive, how they move through intersections, or they perform lane merges and basically every driving task.
This learning is condensed into an AI model that is deployed in a software release. A car with V12 then takes the inputs from the 8 cameras around the car and compares what it sees to the AI model, multiple times per second.
The live stream today really showed how well this works, which is important as conceptually it sounds to simple to actually work. What it does require is a massive amount of video training data and a massive amount of compute to create the model.
When it comes to the timeline for an international rollout of V12, we expect V12 to arrive before the end of the year on customer vehicles in the US, it may even be as soon as a month or two from now.
The rollout of FSD beta to international markets, with Autosteer on city streets, will be much faster, because there is almost no explicit code.
Previous builds of FSD had lots of heuristic code, written by humans to accommodate the differences in different locations. With V12, that all changes and something like the variance in road rules between counties, can be accommodated simply by throwing enough video data at the problem and allowing the data engine to learn from it.
While we still don’t have a firm date for FSD Beta in Australia, we now have a much better runway to achieve that, than we had before.
The vision-only approach that Tesla uses for autonomy, has always been said to have far better-scaling prospects than other HD-maps and lidar-based systems. While other ride-sharing solutions like Waymo and Cruise already offer services, Tesla’s model has the potential to not only catch up, but overtake and spread to new markets in weeks, rather than months or years.
A large dependency for the rate of rollout, is the ability to train on data. This is being addressed with continued investment in compute to process the vast amount of video data into the AI model that now drives the vehicle.
From Musk’s earlier today on an X Space, we know Tesla is working on scaling out a new 10,000-unit Nvidia H100 cluster adding to what they already have and their own Dojo hardware. Each Nvidia H100 costs around US$30,000 making this investment somewhere close to US$300 Million.
