Next Level AI Transformers for Tesla FSD Version 14

by

Tesla’s Full Self-Driving (FSD) V14 will use auto-regressive transformers. This technology will help FSD predict the plans and paths of the surrounding environment, allowing it to more accurately anticipate the actions of other vehicles, pedestrians, etc., and make human-like decisions. FSD V14 will also increase the model size and context size, and will enable more sophisticated driving by utilizing audio inputs to recognize things like emergency vehicle sirens. These features are expected to be introduced in June 2025, along with the launch of Tesla’s robotaxi network.

Auto-Regressive Transformers Will Help FSD Anticipate

An auto-regressive transformer processes sequential data in time, using that information to predict future elements based on previous ones. Imagine completing a sentence: You use the words already written to guess what comes next. This process isn’t just about filling in the blank; it’s about understanding the flow of the sentence and anticipating the speaker’s intent.

FSD could analyze a sequence of camera images to identify pedestrians and predict their likely path based on their current movement and surrounding context. The system’s auto-regressive nature allows it to learn from past sequences and improve its predictions over time, adapting to different driving scenarios.

Today, FSD mostly reacts to what it sees, but FSD 14 will anticipate even more.

4 thoughts on “Next Level AI Transformers for Tesla FSD Version 14”

  1. Unsupervised FSD?

    Maybe it’s really here. I sure hope so.

    I’m buying a new Tesla once I see true UFSD in the wild.

  2. I don’t see why version 14 would be tied to robotaxis. In fact, it would be much smarter to roll it out to supervised cars so that they can test it out before deploying it to cars where there is no supervision in real time. Much safer.

  3. The main reason for removing the steering wheel in the self driving taxi is basically the expectation of a typical drunk getting in and trying to drive…. the steering wheel abuses would end up on countless youtube videos.
    The implementation of the GPT in self driving will hopefully be quite radical and what you may end up with from the taxi perspective is custom city based models that are bound by the drivable area of the taxi. Why would you need to teach a model on a nation wide basis when certain city training specifics could deliver a lot better option and performance. The only limitation on this approach would be training compute, which Musk seems to be ramping up.
    The taxis will undoubtly end up in Vegas quite early on replacing the fleet in the boring tunnels…. on the bucket list for 2026.
    I’m looking forward to the release and seeing what the performance is like in my car assuming this approach will work in older 2022 hardware. Drivng style preferences still seem to have a conflicting input into the training though, where the typical driver on some road points, typically does not actually follow the intended rules of the road and so the self drive does the same.

  4. I love that they are making a steering wheel-less car, that force them to be extremely aggressive on FSD, which is something many people have been wanting for a long time.
    Unsupervised FSD, is the only kind people want.

Comments are closed.