There are quite a few improvements expected with V12.4, which Musk has said will be focused on user comfort by reducing hard accelerations and sudden braking. According to Musk, FSD v12.4 is also expected to have a 5-10x improvement in time between user interventions, which means that Tesla is rapidly approaching its goal for the Robotaxi announcement on August 8, 2024
Are they going to announce Robotaxi?
Andy
First, they have to survive their “internal testing”.
Is that in question? Isn’t that the point of the event?
I mean…that’s what Musk tweeted. “Tesla Robotaxi unveil on 8/8.”
So you are saying that he is going to announce a Robotaxi? Now that is interesting.
Andy
I mean…of course? He scheduled an event to unveil the Robotaxi, so it’s a pretty safe bet that at this event they will unveil the Robotaxi. Otherwise, why have an event to unveil the Robotaxi?
I still feel I’m not understanding the uncertainty here….
And a year or two of Federal review before it’s approved for use on Federal Highways
intercst
Because I am not sure it’s ready. Everything I have read, even from people on this forum who own a Tesla, state that it isn’t. So to get to a Tesla robotaxi, it would have to be full FSD. Otherwise it isn’t the Robotaxi that Musk promised. So you think that the Tesla robotaxi will be here on 8/8/2024?
Andy
Then you agree Intercst that the Robotaxi isn’t ready? @albaby1 it sounds like I am not the only one that doesn’t think it’s ready.
Andy
Ok, now I understand.
Just because Tesla’s unveiling the Robotaxi on 8/8 doesn’t mean that it’s going to be ready on 8/8. We don’t know yet whether the “Robotaxi” he’s unveiling is a vehicle (a prototype on stage) or a service (walking through the app and how its structured) or both - but no matter what, it could still be years before the vehicle/service is actually implemented. If it’s a vehicle, it wouldn’t be put into production for around a year in the best case - and it could end up being like the Cybertruck (or New Roadster, where years later it still hasn’t been produced).
Whatever they reveal, it will be along the lines of “Once we solve FSD, this is what the Robotaxi will look like…” Whether they explicitly state it or not - if they show a vehicle with no driver controls, or talk about what the app for the Tesla Network will look like, it won’t actually be rolled out unless they solve FSD. Which could take months or years.
There is no “ready.” There is only, “is it good enough?” A record player, a pianola, an industrial robot, a cuckoo clock are ready once they do the task exactly as programmed. Not so with neural network AI where there is no heuristic programming. Right now Tesla is releasing FSD v.12.4 and they are working on v.12.5 and v.12.6. Elon Musk noted that as FSD improves it becomes harder and harder to get useful data for further training. Most of the data is about what FSD already does well. What is needed are new edge cases so FSD can learn something new. Think of airline pilots, the training never stops.
When is FSD 'ready?" When it’s better than most humans, when accidents are fewer than with humans.
The Captain
FSD v 12.3 is super good. Very few interventions. v12.4 is a step change.
40K Americans die each year due to auto accidents. Tesla will fix it.
Be on right side of innovation. Ignore the naysayers and negative nancies.
Do you know how far v12.3 can go before an intervention is necessary, generally speaking? How much better is v12.4? I’ve tried to find that type of information online, but no luck - Tesla doesn’t appear to release it.
It is a function of complexity of scenarios (not time or distance). It still has minor issues with Unrestricted Left Turns especially congested scenarios.
Elon has tweeted that 12.4 is a step change and improvement is 5x - 10x. It is one integrated stack now. He is already on v12.6.
In my view it is far more important that the cadence of releases has increased significantly (including support for new form factors). This means that that Tesla’s touchless automation is now working End to End (Data, Dojo, Compute, Annotation).
Hence the confidence in Robotaxi annoucement on 8/8.
I recall reading somewhere, 12.3 is getting an intervention about 3500 miles averaged across the fleet. 12.4 is projected to have an intervention every 17,500 to 35000 miles.
12.4 may be good enough if it reaches the 35,000 mile bogey, but if it doesn’t and we are only getting a 5x improvement per iteration, we can expect 12.5 to be better than 80,000 miles per intervention.
12.5 should be out this year. It would likely be good enough for limited Robotaxi release. 12.6 is likely in 2025 and should be exceedingly good. Good enough for worldwide release.
Cheers
Qazulight
This isn’t Sprint’s cell phone service where that mindset was prevalent. With FSD it is going to have to be better than good enough. I know where you are coming from Captain and I am rooting for Tesla but this is much different than good enough.
Andy
Is that from personal experience Dividend? Could you expound on that?
Andy
Depends on how you calibrate “good enough!”
The Captain
That seems way too high to work, even for a limited Robotaxi release.
As we talked about in the other thread, the average manual driver goes about 170,000 miles between crashes; the average manually driven Tesla goes about 700,000 miles between crashes. While we don’t know for sure, it sure seems that when an AV encounters a scenario where it would normally require a human intervention and there isn’t a human to intervene, you’d have a pretty high incidence of crashes. Not every intervention is going to be in a safety-critical situation, of course. But a lot of them will be. And when there is no human when the safety-critical failure happens, the outcomes are likely to be pretty bad.
If v12.5 “only” gets you to an intervention rate that’s 10x higher than the Tesla crash rate, that seems like it’s not good enough for rollout.
A lot depends here on what the fall back is when the FSD runs into a situation it knows it can’t handle. When there is a driver, it can simply turn over control, but with robotaxi, what is the fall back … remote operation?
There are at least two types of issues here – not knowing what to do and making a mistake about what to do.
Is there a typo in there?