FSD Beta Videos (and questions for FSD Beta drivers)

[dgatwood]

I'm sure it would have stopped.

Unless Elon is fed up with Whole Mars Catalog asking when the next beta is coming.

 

[rxlawdude]

I don’t know how many of you realize this but Cruise requires remote assistance (to plan new paths or allow to proceed) every 5 to 10 miles in SF. In the easier to drive parts of SF.

And when the car requires remote assistance, it’s still considered autonomous, so in the Cruise videos you watch, keep it in mind.

new term: RoboDroneTaxi.

 

[rxlawdude]

You obviously no nothing about the FDA. The FDA would be thrilled if you could show drugs that improved at 300% from one trial to another. THRILLED

Drugs dont do this. Software can. You are really going to be an unhappy camper in 6 months or so when they release this to the general public.

On a side note, when a drug shows simply fantastic performance and few side effects the FDA will quickly dispense with most of the placebo trails. This was recently the case with what looks to be the cure for Cystic Fibrosis (CF) which is maybe the most miraculous drug in years. On another note why in the hell are you bringing FDA into a thread on FSD. If anyone is doing big trials in real world it is Tesla.

Overall your comments are looking increasingly bizarre. I am setting aside everyones frustration (rightful frustration) regarding being sold a product that did not exist. OTOH- if you bought a tesla you could probably afford the investment in FSD. Thanks to all you funders. It looks like a quite useable product is on the cusp of rollout, you could literally be watching the world change...and what do you spend your time doing?

Ignore is your friend.

 

[gearchruncher]

Ignore is your friend.

Someone really needs to look into how you can see and downvote posts from people you have ignored. Seems like there is a unique bug that only impacts rxlawdude, and it keeps happening! He can even quote ignored people!

Straight out of the Elon playbook:

 

[EVNow]

Sure, you can SPECULATE they'll magically be able to add a bunch MORE capabilities and somehow also magically reduce compute usage a TON with 'optimization'- enough to squeeze ALL of that back into the single node it overflowed out of before the first FSDBeta came out.

Actually I'm not "guessing" that they can. But surely you know the R10 is not optimized. This is the first release of the given set of features.

But it seems less likely of an outcome than "needing at least HW4" does.

Based on what ?

Let me ask this again. Are you an industry insider or are you like the many of us here (decades of IT experience, but not in this particular niche area) ? Otherwise its all my speculation vs yours.

 

[EVNow]

Pro tip: You can search the transcript in YouTube and then link to time code.

Musk just says he is confident HW3 can achieve FSD - at "I don't know ... may be 200% to 300%" ...

Doesn't tell me much with any kind of confidence. Definitely doesn't tell me HW3 can't achieve FSD, like many people claim here.

ps : Whether HW3 can achieve FSD / L5 or not is hardly the issue. Can it get to L3 level is the question. That is all I'm hoping for in the next couple of years.

 

[gearchruncher]

But surely you know the R10 is not optimized. This is the first release of the given set of features.

I'm no expert here, but we do have this bit of history:

They said HW2 was enough for FSD. They released videos. Elon was confident. It was not.

Then they doubled processing in HW2.5. Elon was confident. It was not enough. HW2.5 can't even do all the stuff HW3 does in basic AP, even after 2+ years of it being introduced, and tons of time to optimize. It can't do as much inferencing, as shown by not doing garbage cans and pedestrians. It can't see stop lights or signs.

Tesla claims HW3 is 2.5X the compute power as HW2.5. But HW2.5 can't even see a stop sign. So I'm not so sure that 2.5X the power can take us from road lanes only to any kind of L4 FSD. I mean, they've been optimizing V10 for 3 years now- isn't that the point of all the re-writes? There's an awful lot left more to add before it's a useful L4 system.

ps : Whether HW3 can achieve FSD / L5 or not is hardly the issue. Can it get to L3 level is the question. That is all I'm hoping for in the next couple of years.

People here have convinced me there is no useful or real L3. It's L2 or L4. The intermediate is just too hard to manage. If you can go 10 seconds before the human comes online, you can do 30, and then you can do forever, and trying to only do it on narrow parts of the highway is just a messaging mess. The only real difference between L2 that Tesla is doing now and L4 is reliability. L3 doesn't allow lower reliability, so what's the point?

 

[helvio]

I’ll state one more time. I’m not worried whether HW3.0 can achieve FSD or not. I want whatever capabilities HW4.0 can do on my HW3.0 computer. Doesn’t have to be as fast (smart summon/autopark) or even as safe. But I don’t want it to do less than HW4. This has nothing to do with L3/4/5. This has to do with what Tesla as a company achieves in terms of autonomy, full or partial.

The moment I begin to fight is when Tesla is liable for damages caused while the car operates in FSD. Not mine, any.

 

[nativewolf]

View attachment 709357
Credit: XKCD

again.

 

[Knightshade]

Actually I'm not "guessing" that they can.

No, you're guessing HOW MUCH they can.

And your guess is pretty large too.

But surely you know the R10 is not optimized. This is the first release of the given set of features.

That's great.

Of course as already mentioned they were WAY OVER needed compute in a single node long before R10. Or R9. Or R8.

They were needing to spill over to node B by mid-2020 on the production code, months before the very first FSDBeta code was released. (2020.24 specifically)

The problem keeps getting worse, not better. It'll keep getting worse still, and a LOT worse if they run out of B node too before they solve FSD and before HW4 is ready to go into production cars.

Based on what ?

Known facts?

We know they exceeded single-node compute back in mid-2020, before FSDBeta was even out. They keep exceeding it more and more as the capabilities of the system grow- and they still have more they need to grow to achieve FSD.

They keep saying they want to EXPAND the NNs to do more of the work- which requires MORE compute than traditional code generally (but ideally gets you much better, more generally applicable, results).


It's certainly POSSIBLE, in fact virtually certain, they'll be SOME future optimization of code. But it's unlikely it'll happen until they're DONE. Which might not fit into HW3 at all. And seems unlikely to shrink back into a single node.

See Elons 5-steps-of-engineering discussion for why you don't optimize solutions that are unfinished BTW.

(Also NNs aren't traditional code either and anything you know about optimizing traditional code, well, deep into the weeds, but very different world... you can certainly shrink your NNs, but you lose accuracy/resolution--- this might tie into Elons 200-300% better with HW3 and 10x better with HW4 remarks... he's been wrong every single time so far estimating how much compute he needs for even "fully works let alone is X safer" there's been no evidence in favor of him being right NOW, and all the stuff already mentioned suggesting he's underestimated again).

Let me ask this again. Are you an industry insider or are you like the many of us here (decades of IT experience, but not in this particular niche area) ? Otherwise its all my speculation vs yours.

This seems like a discussion of "I'm not a virologist, but here's some peer reviewed papers by experts who are and here's their current conclusions, and here's the most current available data of what we do KNOW right now...." versus "Well I have no evidence of any kind to support my beliefs but we're both equally speculating!"

Which is...not a productive discussion to really continue.

 

[marketisopen]

Hyperchange is the worst because he's really peddling his own wares essentially.

He simply can't be non-biased.

Clearly there are lot of hand picked FSD testers that are not objective, and only handful of almost accidental aces like Chuck.

Chuck is a combination of a normal owner, balls of steel (who even signs up for taking 10+ uncontrolled lefts in a row), and someone with a lot of patience.

This is flat out not true. SolvingTheMoneyProblem, AIDriver, etc is proof of this.

You don't make money if all you did was positive or negative videos about tesla from start to finish.
You make money and get views only if you make completely outlandish superfan vids about Tesla that's well done.
I remember SolvingTheMoneyProblem first videos. I was there for that whole fiasco when he decided to make vids.
TMIOTesla is another and i could keep going. some of these youtube acounts are not even a year old.

But of-couse its normal to get 10-100 million views your first year on youtube and has absolutely nothing to do with the type of content you are doing.

lol, yes TMIO tesla and I ran the channel together. To be honest though, I am not 100% sure if he does it because it his opinion or if it will make money. He does focus solely on Tesla because it makes money but he genuinely believes his points and has always been very biased towards Musk.

He was always a Musk fan even before his channel. We argued about the hyperloop in 2013. And he was pretty optimistic it could be done fairly quickly while I said it wouldn't be.

 

[gearchruncher]

I want whatever capabilities HW4.0 can do on my HW3.0 computer. Doesn’t have to be as fast (smart summon/autopark) or even as safe. But I don’t want it to do less than HW4.

Yeah, and I want the same thing for HW2.5 and HW3. HW3 can show things my HW2.5 car can't like trash cans. HW2.5 can't read a speed limit sign, where HW3 can. These are not FSD features, they are things the SW can do on newer HW even without a FSD license, so it's only because of older HW that I can't do them. Despite Tesla saying my car had everything it needed to do FSD...

Yet guess what? Tesla won't replace my computer for free, and most people think that's fine and appropriate:

Will Tesla Charge for HW3 to use a FSD subscription?

A Redditor noticed that the code is leaking out for the FSD sub, and it will point owners of non-HW3 cars to an upgrade link: It doesn't appear they are going to try and have different subscription prices or terms for pre HW3 cars. By definition, this only applies to HW2 and HW2.5 cars, which...

teslamotorsclub.com

HW3 to HW4 will be no different. If you think it will be, we're heading into that "but Tesla wouldn't do that to me because something about my situation is special, and Tesla told me when I bought...." territory. When Tesla does eventually try this on you, and you're frustrated, there will be a new cohort of people that bought a year ago going "I don't see the issue. Tesla never promised you anything. Anyway, I have HW4, and Tesla promised me when I bought that...."

 

[helvio]

No, you're guessing HOW MUCH they can.

And your guess is pretty large too.



That's great.

Of course as already mentioned they were WAY OVER needed compute in a single node long before R10. Or R9. Or R8.

They were needing to spill over to node B by mid-2020 on the production code, months before the very first FSDBeta code was released. (2020.24 specifically)

The problem keeps getting worse, not better. It'll keep getting worse still, and a LOT worse if they run out of B node too before they solve FSD and before HW4 is ready to go into production cars.




Known facts?

We know they exceeded single-node compute back in mid-2020, before FSDBeta was even out. They keep exceeding it more and more as the capabilities of the system grow- and they still have more they need to grow to achieve FSD.

They keep saying they want to EXPAND the NNs to do more of the work- which requires MORE compute than traditional code generally (but ideally gets you much better, more generally applicable, results).


It's certainly POSSIBLE, in fact virtually certain, they'll be SOME future optimization of code. But it's unlikely it'll happen until they're DONE. Which might not fit into HW3 at all. And seems unlikely to shrink back into a single node.

See Elons 5-steps-of-engineering discussion for why you don't optimize solutions that are unfinished BTW.

(Also NNs aren't traditional code either and anything you know about optimizing traditional code, well, deep into the weeds, but very different world... you can certainly shrink your NNs, but you lose accuracy/resolution--- this might tie into Elons 200-300% better with HW3 and 10x better with HW4 remarks... he's been wrong every single time so far estimating how much compute he needs for even "fully works let alone is X safer" there's been no evidence in favor of him being right NOW, and all the stuff already mentioned suggesting he's underestimated again).






This seems like a discussion of "I'm not a virologist, but here's some peer reviewed papers by experts who are and here's their current conclusions, and here's the most current available data of what we do KNOW right now...." versus "Well I have no evidence of any kind to support my beliefs but we're both equally speculating!"

Which is...not a productive discussion to really continue.

Migrating C++ code to NN can free up a lot of resources, namely memory leaks that are minimal to non existent in the newer stack. I don’t know how much code they have left on C++ stack but neither does you. What we do know is that the compute power to train is incredibly more than to infer. So I guess optimization would take HW3 farther than what you think. Far enough for redundancy? I don’t know. Dozens of pedestrians with their specific forecasted paths were not enough to saturate HW3 as shown on AI Day. Then we add gestures to all of them, increasing computing power needed, but also infer more on NN than on C++, reducing the computer power needed. All I know is that regardless on the outcome, it shall be an interesting development.

Remember, we went to the moon on something a Raspberry Pi 1 would overpower, squared.

 

[helvio]

Yeah, and I want the same thing for HW2.5 and HW3. HW3 can show things my HW2.5 car can't like trash cans. HW2.5 can't read a speed limit sign, where HW3 can. These are not FSD features, they are things the SW can do on newer HW even without a FSD license, so it's only because of older HW that I can't do them. Despite Tesla saying my car had everything it needed to do FSD...

Yet guess what? Tesla won't replace my computer for free, and most people think that's fine and appropriate:

Will Tesla Charge for HW3 to use a FSD subscription?

A Redditor noticed that the code is leaking out for the FSD sub, and it will point owners of non-HW3 cars to an upgrade link: It doesn't appear they are going to try and have different subscription prices or terms for pre HW3 cars. By definition, this only applies to HW2 and HW2.5 cars, which...

teslamotorsclub.com

HW3 to HW4 will be no different. If you think it will be, we're heading into that "but Tesla wouldn't do that to me because something about my situation is special, and Tesla told me when I bought...." territory.

You’re free to fight for that and I’ll support you! I’m not saying I’m special, but at the same time I’m not hugging the world. I just don’t have the energy to fight for everyone else. Or the desire TBH. But if you think you should, I stand by you!

 

[dckiwi]

Yeah, and I want the same thing for HW2.5 and HW3. HW3 can show things my HW2.5 car can't like trash cans. HW2.5 can't read a speed limit sign, where HW3 can. These are not FSD features, they are things the SW can do on newer HW even without a FSD license, so it's only because of older HW that I can't do them. Despite Tesla saying my car had everything it needed to do FSD...

I'm a bit confused by this. You didn't purchase FSD, and yet you think that you're entitled to a new computer because Tesla said your car had everything it needed for FSD. Even though you didn't buy FSD?

I do think this is fundamentally different than HW3 FSD owners feeling entitled to a HW4 upgrade if that upgrade is required for *FSD* functionality.

 

[gearchruncher]

I'm a bit confused by this. You didn't purchase FSD, and yet you think that you're entitled to a new computer because Tesla said your car had everything it needed for FSD. Even though you didn't buy FSD?

My car was advertised as having all HW needed for FSD CAPABILITY. No caveats. No requirement to buy the software for that to be true.
I went to subscribe to FSD and they told me I needed a HW upgrade for $1,000 to subscribe to the exact same software I would get as if I bought it.

My car does NOT have all HW needed for FSD CAPABILITY, and that is going to cost me $1,000 to fix. The most basic definition of damages. I don't want to acquire my FSD capability license via purchase, I want to do it via subscription, and Tesla never once indicated that my car only had the HW needed for FSD purchase, not FSD subscription. Tesla does not advertise FSD subscriptions and purchases independently- they are clearly the exact same thing, just paid for differently. The subscription is literally named "Full Self Driving Capability" the exact thing Tesla told me my car had the HW for.

 

[dckiwi]

My car was advertised as having all HW needed for FSD CAPABILITY. No caveats. No requirement to buy the software for that to be true.
I went to subscribe to FSD and they told me I needed a HW upgrade for $1,000 to subscribe to the exact same software I would get as if I bought it.

My car does NOT have all HW needed for FSD CAPABILITY, and that is going to cost me $1,000 to fix. The most basic definition of damages. I don't want to acquire my FSD capability license via purchase, I want to do it via subscription, and Tesla never once indicated that my car only had the HW needed for FSD purchase, not FSD subscription. Tesla does not advertise FSD subscriptions and purchases independently- they are clearly the exact same thing, just paid for differently. The subscription is literally named "Full Self Driving Capability" the exact thing Tesla told me my car had the HW for.

Ah ok, that makes sense and I agree with you. If you purchase the FSD subscription they should upgrade your hardware. I thought you were saying you expect to get an upgrade just for some of the newer visualizations, even if you don't purchase FSD.

 

[Iain]

So you're 100% sure that if Tesla said "no more HW3 development" right now, and no HW upgrades, what you have today is all you get, you'll be totally fine?

Wow!!! You sure can twist what he said . I paid for FSD I’m extremely happy with being part of the journey. It’s a personal choice and i see vast improvements with the betas. My car is way better at everything since I bought it in 2020, the only car I’ve owned in forty years I can say that about.

Did you purchase FSD and regret it ?

 

[gearchruncher]

I thought you were saying you expect to get an upgrade just for some of the newer visualizations, even if you don't purchase FSD.

Well, in my case I would truly subscribe (once CSA is out), but I do think you should be able to get the upgrade for free no matter what.

Customer: My neighbor's car shows animals and reads speed limit signs. He doesn't have FSD. My car doesn't. What's up?
Tesla: Oh, speed limit signs and cat detection are only on cars with hardware that is capable of FSD.
Customer: You told me my car had all HW needed for FSD capability.
Tesla: Well, it doesn't but if you buy FSD, we'll fix that.
Customer: I don't want FSD. I want the exact same thing you give my neighbor without FSD, since both of our cars are capable of FSD.
Tesla: Well, your car isn't capable of FSD without a hardware upgrade, and that is $1000.
Customer: But you advertised it as capable. And if it WAS capable, it would be able to read speed limit signs and see puppies, right?
Tesla: Yes.
Customer: But you want me to pay $1K for something that you said was already there?
Tesla: Yes. Welcome to Tesla ownership!

Tesla put themselves in this situation. They do more on cars that are FSD Capable than cars that are not even if those capable ones do not have FSD. They have created a product difference that they advertised would never exist. That's all a customer really needs to have to request they fix it. And remember, every one of these cars is still under warranty. The car literally doesn't do what it was advertised to do, while still under warranty, and Tesla wants $1000 to resolve it.

 

[Knightshade]

Yet guess what? Tesla won't replace my computer for free, and most people think that's fine and appropriate:

Sure they will. If you own FSD you get the new computer for free.

If you don't own it, you don't need it.

Name a single feature you were promised WITHOUT OWNING FSD that you don't have due to the computer thing.

Nobody ever promised "displays recycling bins" to anybody as an FSD feature.

HW3 to HW4 will be no different.

Which is why I, as an FSD owner, look forward to my next free computer upgrade

My car was advertised as having all HW needed for FSD CAPABILITY. No caveats. No requirement to buy the software for that to be true.
I went to subscribe to FSD and they told me I needed a HW upgrade for $1,000 to subscribe to the exact same software I would get as if I bought it.

As you've had explained to you many times already- the subscription did not exist when that promise was made to you.

The thing that DID exist, buying FSD, remains available to you, and incurs $0.00 in added HW costs to you.

So you have $0.00 in damages.

"They didn't specify I couldn't get a thing that wouldn't exist till years later (the subscription)" would be an hilarious argument to hear in court though. You were told you can file in small claims quite some time ago. But never did. Almost like you know you'd be laughed out of court.

Migrating C++ code to NN can free up a lot of resources, namely memory leaks that are minimal to non existent in the newer stack.

But it's not lack of memory that's the issue. At all.

It's lack of compute

Running more, and larger, NNs, cost more compute.

That's what spilled them over to node B, not adding traditional code.

That's why Elon expects HW4 to be many times safer than HW3 even if they COULD get it working on both-- larger NNs that HW3 can't run at all.

I don’t know how much code they have left on C++ stack but neither does you.

But roughly, I DO know this. So does anybody who follows it.

Green discusses this frequently.

Perception is largely NNs.

Nearly everything else (until a few months ago EVERYTHING else) was C++.

But even before that they kept adding newer, and larger perception NNs in that time, over a year ago having run out of compute on Node A alone to do it.

What we do know is that the compute power to train is incredibly more than to infer. So I guess optimization would take HW3 farther than what you think.

Wut?

There's no training done on the car at all so no relevance to the car HW there.

Training is done back at HQ on a GPU supercompute cluster right now... and Dojo in the future.

Remember, we went to the moon on something a Raspberry Pi 1 would overpower, squared.

While I appreciate the reference, they had plenty of time to do a LOT of math by hand in advance there. And even then- the actual landing was done manually with an astronaut on the stick.

And Apollo 11 didn't need make a single unprotected left either.