Discuss: Tesla to Launch a Larger FSD Model After the Holidays; Will Remove Robotaxi Safety Monitors in Three Weeks

[Not a Tesla App]

Elon Musk confirms Tesla will launch unsupervised Robotaxis in Austin in three weeks, removing human monitors for the first time.… Read More

Tesla to Launch a Larger FSD Model After the Holidays; Will Remove Robotaxi Safety Monitors in Three Weeks

Elon Musk confirms Tesla will launch unsupervised Robotaxis in Austin in three weeks, removing human monitors for the first time.

www.notateslaapp.com

 

[Procal]

So I'm confused about this “larger model”. The way I understand it is that it isn't larger in the sense that it is one model that handles all task, but rather smaller models all designed for specific tasks clustered into one “larger model”. So will this mean, for example, that when the car is in a parking lot it will be using a “parking lot behavior model”, and similarly when on the highway it will use a “highway” model etc.?

As a side note, does this also mean we aren't going to see anymore FSD updates until January? That would be a shame, as I feel it can use at least 1 more minor update to improve it further. I'm noticing that some issues of previous versions are still present, even if they are rarer.

 

[ALIEN SUPERSTAR]

i could be very wrong, so please bear with me, but i have a slight feeling that this new, bigger model will come with a bit of an FSD rebrand. in general FSD seems like a difficulT subject to teach people about without a PR department but also now we have versions and build numbers, which to the masses might be confusing. Tesla has been overall simplifying everything else. for example, starting FSD is different, seeing how FSD is active is different, and picking different driving modes is different. all of these changes have felt like ways to overall make the experience simpler and easier to understand.


i also think that once we get to unsupervised, we won’t really have to keep track of what version we are on. i don’t think we are close yet but only time will tell. once it’s a robust, reliable product, the need for us to monitor version numbers will likely disappear completely.

 

[phxazcraig]

I just reread the article on FSD reasoning and the hiring ad for a new engineer to implement FSD reasoning. My first impression was that it was pretty late in the game to be hiring a new engineer to develop what sounds like 14.3. My second impression, after reading the article for the second time, was that it could be a good long year of 'FSD reasoning updates' before we get to a model that not only reasons well, but gets FSD to obey and follow the reasoning.

We saw a lot of development of v13 and v14 versions just to start approaching their originally-promised goals. I don't know why they never implemented things called out in the release notes, but in general, everything is harder than you first think it will be. Getting cars to think ahead sounds like one of those 'harder than it sounds' projects, and it sounds pretty hard.

 

[Procal]

After doing some digging and having a bit more understanding (but still mostly clueless), I wonder if this is how they plan on implementing a v14 “lite” for HW3 cars. If they can manage to split FSD into smaller more specialized models and run them independently rather than simultaneously, this could in theory mean that less processing power and less memory would be needed. It would also likely mean that they could have a slimed down specific model to replace the old Autopilot stack.

If this is how they are planning to make these changes it could mean massive performance improvements for FSD, however, I worry about how this will impact responsiveness. If FSD will now have to determine which model to use for different situations, this could potentially add extra processing time, which would have an impact or delay on how fast FSD can respond to a very dynamic scenario.

I would love a deeper dive into explaining this in a more nuanced way that someone who isn't familiar with how AI works in an article.

 

[phxazcraig]

I wish I had a clue how this stuff works too. My take at this point is that driving training is split into many different areas, possibly called contexts. One for stopping at a stop sign, one for stopping at red lights, one for driving between lines, one for merging onto freeways, etc, etc.

Each of these training tasks may be broken out into individual segments within the FSD 'brain', meaning that each segment takes a certain amount of memory and expects a certain amount of tasks/millisecond execution time. All of segments add together to create an intelligent mechanism.

So if HW3 has less processing power AND less memory, then two things have to happen to convert HW4 FSD to HW3 FSD:
1: segments only get about half the memory. I don't know how one makes a 10mb segment become a 5mb segment, but certainly there must be some loss of capability along the way. Could be that some segments are simply omitted to save memory.
2: processing power being reduced might mean something like skipping every other clock cycle compared to HW4. Thus, somewhat slower overall.

It may be that advanced compression techniques gain back a lot of the memory loss.