Maybe they'll relax the Autopilot speed limits on S/X with the added radar, the S Plaid has a 200 mph top speed so Autopilot should scale up proportionally.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Elon: FSD Beta tweets
- Thread starter Terminator857
- Start date
clydeiii
Member
The roundabout I drive on every day during my commute has gotten significantly better in the last few versions...10.69 was ok, 11.3 was pretty good, 11.4 is driving it like a human would. Can't speak to all roundabouts in the world of course.
I must have Completely different expectations on progress than a few. I have used FSDb for over a year and the progress and its capability has progressed drastically and far from a party trick. Perfect no but with only a few rare confirmations via taping the accelerator etc most drives are minimal if any interventions. Could I find a construction area type road block to stump it, sure. But in general it seems most have some agenda of Perfect or nothing. I’m just along for the ride in most cases and rather impressed. If it was Easy every auto company would have done it. It’s clearly a Major accomplishment which I respect taking time to complete.
I can’t tell if u are being sarcastic here. Cruise control max speed should not be related to the car’s max speed lol. Should a Bugatti have cruise control up to 200 mph lol?
Why not, it will run out of gas in 12 minutes so what could happen.
DanCar
Active Member
> Exponential improvement coming to FSD Beta once Dojo is up and running.
I doubt it will be a very big improvement. You need an inference engine capable of handling a large training model. FSD hardware V3 isn't it. Maybe V4 is. In other words: just because you train on a lot of data, doesn't mean you get exponential results at the end of the pipe.
I doubt it will be a very big improvement. You need an inference engine capable of handling a large training model. FSD hardware V3 isn't it. Maybe V4 is. In other words: just because you train on a lot of data, doesn't mean you get exponential results at the end of the pipe.
diplomat33
Average guy who loves autonomous vehicles
I think we will see more gradual improvement, not smooth exponential progress. It will be "two steps forward, one step back". So there will be improvements but also some regressions, more improvements, some regressions etc... I say this because regressions are a natural part of ML. And with such large data training, the chances of a regression in one part, will be greater. I don't believe the notion that Tesla will dump massive data into Dojo and L5 will pop out.
I also agree with you that hardware will be a limiting factor. So I would not be surprised if Tesla announces HW5, HW6 etc...
But an easy way to see if the improvement is exponential is for Tesla to publish safety intervention rates. If we see the intervention rate improve exponentially that would imply exponential improvement. But we also know that the march of 9's gets harder and each next 9 is harder that the last. So we could also see the intervention follow a S curve where it starts to grow exponentially at first but then hits a plateau.
Doesn't AEB involve very loud beeping? I haven't heard people saying that their phantom braking is accompanied by loud beeping.
diplomat33
Average guy who loves autonomous vehicles
I would add that the main advantage of Dojo, as I see it, is that it will greatly speed up training of large datasets. So Dojo will allow Tesla to train on much larger datasets in a reasonable amount of time that would otherwise take too long on less powerful compute. Speeding up training is certainly very useful but it will not guarantee that the training will be correct. It will still take time to validate the training to make sure it is correct. Tesla will likely need to do retraining to fix regressions etc... And we also see from the graph that it will take about 6 months for Dojo to be in the top 5 supercomputers and another 10 months to reach 100 exo-flops. So it will likely be a year before Dojo is big enough to even make a difference. I think we will see progress but I don't think we will see instant progress as soon as Dojo comes online.
Yes, AEB will sound an alarm. My example was thrown in there to illustrate the idea of patching a problem. It was just the first thing that came to mind.
jabloomf1230
Minister of Silly Walks
I agree with you that hardware (sensors) will be the limiting factor. I believe that issues with emergency vehicles, school zones, school buses, no right turn on reds, lane changes and multilane roundabouts will eventually be fixed with existing hardware. However, without side-facing cameras (or equivalent sensors) on the car's front bumper, blind ULTs and URTs will never be solved. The NN can't navigate what it can't sense. Even a wide angle cam placed at each repeater position isn't going solve blind UTs.
There or some stories, just a month or two ago amongst all the recent LLM buzz, about promising conclusions regarding training a large model on a huge training set, then reducing it to a much smaller execution engine. Besides that, huge LLMs like chatGPT is just one of the latest fads, obviously because it demonstrated breakthrough results somewhat ahea d of schedule.
It's not yet clear what architecture(s) will prove most suitable for the FSD problem (a safe statement even if I don't claim any special expertise
) , and it's changing month by month. I don't think anyone would disagree that more compute in the car would be advantageous, but it's too early to conclude that the 2019 computer is inadequate for future ML models.
jeewee3000
Active Member
Elon has kind of announced HW5 already. I got my HW3 Model Y last December and I'm intent on not getting a new car before HW5 is available
.From the Q4 2022 earnings call (emphasis mine):
Martin Viecha
Thank you. Next investor question is, Elon said previously that FSD Hardware 4 will most likely come first in Cybertruck. Is that still the current plan? Do you expect there to be an upgrade path for Hardware 3 cars to Hardware 4?
Elon Musk -- Chief Executive Officer and Product Architect
Yes, Cybertruck will have Hardware 4. And to be clear, for 2023, Cybertruck will not be a significant contributor to the bottom line, but it will be next year. So it's an incredible product. I can't wait to drive it personally.
It will be the car that I drive every day. I actually just -- I'm wearing the T-shirt with this matched glass. And it's just one of those products that only comes along once in a while, and it's really special. So yes, so with respect to upgrading cars on -- that have Hardware 3, I don't think that will be needed.
Hardware 3 will not be as good as Hardware 4, but I'm confident that Hardware 3 will so far exceed the average -- the safety of the average human. So [Inaudible] how do we get ultimately to -- let's say, for argument's sake, if Hardware 3 can be, say, 200% or 300% safer than human, Hardware 4 might be 500% or 600%. It will be Hardware 5 beyond that. But what really matters is are we improving the average safety on the road.
But it is the cost and difficulty of retrofitting Hardware 3 with Hardware 4 is quite significant. So it would not be, I think, economically feasible to do so.
(...)
Pierre Ferragu -- New Street Research -- Analyst
Excellent. Thank you so much. So as a quick follow-up, Elon, I was thinking about like FSD, and when you look at like the situation today compared to a year ago, it's -- like the progress has been like amazing in the quality of the product but also its rollout. And so, I was wondering, how much is this like impacting the take rate of FSD today? So do you already see that people are getting more excited by FSD because they see it around them on 400,000 cars and they see the value of the service already? Or is that too early to really see like -- to expect like an uptick in the take rate?
Elon Musk -- Chief Executive Officer and Product Architect
The trend is very strong toward use of FSD. And as you alluded to, the -- with each incremental improvement, the enthusiasm obviously increases. And so, I think something that still a lot of people out there don't quite appreciate is that Tesla -- of course, Tesla is as much as a software company as a hardware company, but Tesla is really one of the world's leading AI companies. This is kind of a big deal with AI on the software side and on the hardware side.
With the Hardware 3 inference computer, still the most efficient inference computer in the world despite being, at this point, five years old from the design point. And with Hardware 4 coming and then Hardware 5 beyond that where there are significant leaps. And the Dojo computer, we expect to be using that operationally at Tesla later this year. And we're seeing just a lot of world-class AI talent join the company.
There's also the long-term potential of Optimus where we're able to use our expertise in electric motors and power electronics, batteries and advanced manufacturing to be able to make a humanoid robot that is actually useful and can be made at high volume with exceptional capabilities because of the autopilot AI that -- where we take the -- because the car is like a robot on four wheels and Optimus is a robot on legs. But the -- as we get closer and closer to solving real-world AI, and we don't see anyone even close to us in achieving this, the value -- I think you appreciate this and a few others do, but most don't know what I'm talking about. And so -- but it's -- this is the thing that has order of magnitude, potential market cap improvement for Tesla.
EDIT:
Elon confirmed HW5 again today on Twitter!
FSD discussion
Looks like the main vehicle software releases are more regularly including FSD Beta. I was noticing 2023.12.10 with FSD Beta 11.3.6 updating to 2023.12.11, and yesterday a lot more vehicles from without-FSD-Beta versions like 2023.12.9 were newly getting this but generally outside of US/Canada...
teslamotorsclub.com
I don't understand why this is considered news. There's always going to be a HW n+1, same way as there will always be an iPhone n+1 if you wait long enough.
diplomat33
Average guy who loves autonomous vehicles
Sure. I think the reason it is considered news is because Elon famously claimed that HW2 would be good enough for FSD. In fact the Tesla website used to say in big letters about HW2, "your car has all the necessary hardware for full self-driving".
jeewee3000
Active Member
True, but there's a difference between the implication that hardware gets updated once every few years and concrete talk of the capabilities of each hardware-suite.
In other words: the "news" is what HW5 will be capable of.
Anything Musk says about the future is totally worthless though. I think he's very interesting to listen to about his ideas for possible futures, or on how his companies are doing the things they are currently doing because he's very knowledgeable about that stuff, but his statements about what will 'definitely happen' and when have been shown to be wrong so many times you have to learn the lesson eventually.
Either the bit of his brain that does estimates is just missing / replaced by more 'visionary engineering' stuff than the average human, or he knows his predictions are junk and just doesn't care. Either way, they're a waste of electrons.
jeewee3000
Active Member
I've always found this critique of Elon to be too harsh. The guy is trying to estimate how long it will take to do things that have never been done before.
I don't listen to his timelines personally. But his train of thought rarely changes unless his current path leads him to a dead end. (For example the switch to steel for starship, that was a big turnaround from the first BFR vision, since large composite wrapped tanks proved too expensive/unreliable/hard to make).
With FSD in Elons mind the question isn't anymore: "can it be done?", but "how many iterations of the software/hardware will we need?" In other words: he sees a clear path and is just following it.
With rapidly reusable rockets he says that path is less clear. He has only recently (after switch to stainless steel) thought it can certainly be done. For a long time he doubted even the possibility of success.
Anywho, I won't go more off-topic on this, but that's why I do see some value in Elon's road map regarding FSD (HW4, 5 etc).
DanCar
Active Member
Some of us are of the opinion that current hardware will never cut the mustard. News of new hardware brings hope that we will someday be able to do other things in our cars besides mundane supervision or driving.
Similar threads
