Elon Says FSD V12.4 Is Way Better than Very Good FSD V12.3

by

Elon Musk says Tesla FSD V12.4 is another big jump in capabilities and the training compute limitations are almost gone.

AI expert James Douma had said that with additional training and data, he expected FSD V12.X to reach 0.1-1% of the interventions for Tesla FSD v11.X.

If the level of improvement expected by James Douma were achieved then 1% intervention levels would be reached with three steps of 20% less each time.

100 interventions, 20 interventions, 4 interventions, 0.8 interventions.

If the rate of FSD training enabled 1 full release or improvement of an existing release every month, then 100X better than FSD v11.X would happen in May-June 2024.

IF the Nvidia H100 chip sales for 2023 is correct, then Tesla had 15,000 Nvidia H100s in 2023. These should be installed and operating by now. This means that Tesla has at least 65 Exaflops of compute for training.

Tesla had targeted 100 exaflops of training compute by October, 2024. If Tesla bought 10,000 Nvidia H100s in 2024 then Tesla will have over 100 Exaflops of training compute installed by May 2024.

12 thoughts on “Elon Says FSD V12.4 Is Way Better than Very Good FSD V12.3”

  2. FSD is a surprisingly amazing driving experience! It makes regular driving even better as it is actually interactive and engaging. It’s smarter than you realize and actually can see what’s happening around you before you! Also knows where you are headed and makes sure you get there (no missed exits trying to follow the gps monitor or voices guidance. Steers really well too! It needs your assistance at tough intersections but it’s an interactive — you use the AI to drive better!
    Don’t be afraid of the future — it’s is actually a better more hi-tech and sophisticated DRIVING experience than can be imagined! (autopilot pales in comparison however—just fancy cruise control).

  3. I think FSD V12.3 is remarkable. Looking forward to FSD V12.4.
    12.3 is almost intervention free. My wife, who was very critical of V11, did not notice engagement with V12.3. That in itself is a huge step forward.

  4. When they solve the phantom braking, maybe I’ll start to trust anything said on the matter. Until then all autononsense is staying disabled. It’s a good car otherwise but these half-baked features are just pushed recklessly.

  6. Until it’s good enough that I can safely take a nap while on a long drive, I don’t freaking care. If I have to be alert and paying attention anyway, I might as well be the one driving. It would actually be less exhausting than continually double guessing what the car was about to do.

    But maybe that’s just me, and other people do find value in a self driving program you have to continually monitor anyway. I mean, other people use cruise control, and find it valuable.

    OTOH, give the car a “take me to the nearest emergency room, stat!” button, and that I could see some real value in, even if it wasn’t as safe as a human driving. Because I’d only be using it when I wasn’t a safe driver anyway.

    • Kinda depends on what “safely” take a nap means. You could do that now, just not legally or within FSD’s guidelines so the Tesla wouldn’t let you. Interventions don’t imply an accident would have occurred otherwise. Regulators aren’t going to allow it until it’s demonstrated to be unambiguously safer for you to take a nap than have the option of controlling the vehicle. Probably requires data about interventions to be individually reviewed to judge whether they were justified or not and what the consequences of not intervening would have been.

    • I used to feel the same way. Recently though I had to take several trips between NY and Florida and autopilot on highways made a pretty big difference in tiredness- 9 hour a day drives would feel like half that with autopilot by end of day. But as they say, you’re milage may vary.

Comments are closed.