FSD v12.x (end to end AI)

[drtimhill]

That's the Tesla-stan FUD. Reality: There is not a single aspect of Waymo's system that isn't at least 10x more advanced than what Tesla has. Cameras, Radar, planner, safety, object detection et.c.

Care to explain how you have such detailed privileged knowledge to make that call? Not sure what kinda of "reality" you appear to be living in.

 

[Dewg]

It will blow or minds and be "on fire"?

Here's where we are today:

Logical Fallacy: Extrapolation

 

[powertoold]

Waymo sucks compared to what we've seen V12 do. Not saying V12 is better than Waymo in every way, but considering V12 is driving in "real-time" like humans do, whereas Waymo has a cm-level HD map of every single thing along with semantics of all lanes and curbs and islands, etc.

Waymo has no real understanding, it's constantly doing stupid stuff like these, and I didn't even look really hard:

 

[powertoold]

And here is Waymo deciding to reroute because it doesn't like a particular street? (maybe I'm wrong about why it's doing that?)

And here is Waymo's horrible driving, wobbling the wheel all over the place:

Again, I'm not even looking very hard for these examples. They're everywhere; Waymo's driving is not natural, has little to no subtlety and is basically a glorified obstacle avoidance train. You can tell from the wobbly wheel that it is considering the cm-distance of every little obstacle and turning the wheel one-way or another to avoid a set distance from it.

Anyone looking at Waymo vs V12 and thinks that Waymo is somehow superior at driving vs V12 needs to have their eyes checked, seriously. Waymo's driving is becoming sadder and more antiquated every week. Once V12 is wide released, like I said, it's over for Waymo

 

[Electroman]

Would someone hack the software for TMC website to remove the 'Funny' emoji for one member of this forum?

 

[stopcrazypp]

I've said this many times on this forum, and here we go again:
If autonomy was solved by "more data" and "more compute" it would have been solved five years ago.

Given the previous implementation is mostly heuristics based (with NN gradually driving more of the perception stack), I'm not sure how you can extrapolate that 5 years ago if would have been solved. That doesn't make any sense. We have yet to see what is possible using E2E. It could be another dead end, but it's way too early to judge at the moment.

The number 10% is optimistic for this decade (in less than six years) for camera only imho, but if you have any evidence (in research or otherwise) to the contrary, I'm eager to hear it.

Given the vast majority of dumb moves by FSD Beta don't appear to be perception related, I'm not sure if "camera only" matters at this point. Maybe Tesla will reach the limits at some point where that's the bottleneck, but it's not there yet.

 

[Daniel in SD]

Drop off a Waymo vehicle to my driveway tomorrow. It won't even be able to drive itself around my block.

How do you know this?

HD maps. Next question?

I have questions.
Has Waymo had a single collision caused by their use of HD maps?
If you can use AI to drive a car why can't you use it to make HD maps?
It just seems odd to me that people are convinced that you can train neural nets to drive a car using massive amounts of driving data but automating the generation of maps can't be automated well enough to scale.

 

[Daniel in SD]

Once V12 is wide released, like I said, it's over for Waymo

That's not much of a prediction. The same could be said about V11.
How many miles per disengagement do you expect by the end of year?
When do you expect wide release?

 

[Supcom]

That's not much of a prediction. The same could be said about V11.
How many miles per disengagement do you expect by the end of year?
When do you expect wide release?

Even if V12 ends up being everything anyone could hope for, I suspect the clever folks at Waymo could figure out how to create and train their own E2E system.

 

[Daniel in SD]

Even if V12 ends up being everything anyone could hope for, I suspect the clever folks at Waymo could figure out how to create and train their own E2E system.

Sure, but can they figure out how to manufacture cars as cheaply as Tesla?
I don't think E2E is going to work. It will probably work a lot like ChatGPT, human like most of the time.

 

[spacecoin]

Care to explain how you have such detailed privileged knowledge to make that call? Not sure what kinda of "reality" you appear to be living in.

Privileged knowledge? It's self-evident. Look at the cameras, the perception, object and gesture detection, look at the planner, look at the safety record, look at the actual performance. It's ridiculous even comparing the two. Waymo was driving like FSD 10-15 years ago.

Only fools brush off Lidar, highres-radar and HD-maps and other safety related additions to compensate for NN weakness as "rails" or "crutches". Waymo would most likely drive better and safer than FSD by at least a factor 10x even if you removed everything but the cameras.

 

[spacecoin]

Even if V12 ends up being everything anyone could hope for, I suspect the clever folks at Waymo could figure out how to create and train their own E2E system.

Waymo uses end to end techniques in some aspects of their systems since at least 2019.

 

[Bladerskb]

If Tesla spent a single week fine-tuning the lane semantics / parking lots specifically for the routes that Waymo takes in Chandler, I think V10 would be comparable to Waymo's disengagement rate (which is around 1 in 500 miles in JJRick's videos). This is how much of a joke Waymo is becoming. In the coming months, it'll become clearer and clearer as we look back at JJRick's Waymo videos.

Waymo clearly over since v10

 

[JulienW]

So this is the orange vs apple thread?

 

[lzolman]

Once V12 is wide released, like I said, it's over for Waymo

Well, once V12 is wide released and it doesn't make people fear for their lives (or at least the integrity of their car and others) then it'll be game over.

 

[Knightshade]

well, yeah, but that's true for a lot of words, especially for "terms of art" in various disciplines.

From a software developer perspective, you'll find very broad agreement that "beta" typically denotes software that is considered feature complete, but has not been thoroughly validated. "Validation" would consist of bug regression and functional testing.

It's important to note that software out of beta may still have known bugs.

What I imagine a lawyer would do in a courtroom setting is bring in an expert witness and ask them to define "beta" in this context.

I suspect not much? Term of art most often comes up when it's one company IN that industry suing another-- not same lay person customer who wouldn't be privy to the nitty gritty of terms of art in an industry anyway.

Also not much since as you just pointed out software OUT of beta may also have known bugs--- Plus Elon himself has said, going back years, Tesla does not use the term the same way as is "standard' among many others in the software industry.

So not only does it lack any legal meaning to begin with, you've got the CEO telling you, going back years, they don't even mean the same thing by it as the more common terms of art definition either.

Further- it's not like Tesla hasn't ALREADY had a number of lawsuits over beta-labeled AP/FSD... and AFAIK has not had the beta label be relevant to any (and generally won all of them liability-wise as well).

Yes but Tesla never defined what constitutes an "action"

No need- The fundamental POINT of a contract is It creates an obligation to take or not to take certain actions so that's pretty solidly understood in contract law in the first place... . just like they don't need to tell us what "may" vs "must" means in a contract, nor what other terms defined elsewhere in law like, say, "drivers seat" means (See FMVSS for that one for example).

Beta OTOH has no such definition anywhere in law that I'm aware of- but I remain open to a citation otherwise.

There is a definition, just not a codification. There’s a difference, and the lack of codification does not preclude using a definition for legal purposes so your statement makes no sense.

Can you cite where I can find the legal definition you claim exists? Or any legal case where that definition was used in resolving a lawsuit or liability claim? Because if there's not a LEGAL definition than your claim you CAN use one for legal purposes is the statement not making sense.

Elon said the beta tag is to encourage people to pay attention. Obviously that will not be necessary for unsupervised V12 so they will remove the beta tag in that extraordinarily unlikely scenario. Just my prediction!

So if they remove it while it's still supervised that would falsify your claim?

 

[Supcom]

Sure, but can they figure out how to manufacture cars as cheaply as Tesla?
I don't think E2E is going to work. It will probably work a lot like ChatGPT, human like most of the time.

If you are running a commercial robotaxi service, the cost of the car is less important than it is for an individual.

I don't know whether E2E is The Way. But if it is, others will follow. The advantage Tesla has is the sheer volume of training data that comes from a large fleet of cars just doing their normal daily driving.

 

[Daniel in SD]

So if they remove it while it's still supervised that would falsify your claim?

Absolutely. Of course I’ll still think it wasn’t their original plan.
I’ll always think that FSD beta is a beta version of FSD robotaxi and not a driver assist.

 

[FloridaJohn]

Anyone looking at Waymo vs V12 and thinks that Waymo is somehow superior at driving vs V12 needs to have their eyes checked, seriously. Waymo's driving is becoming sadder and more antiquated every week. Once V12 is wide released, like I said, it's over for Waymo

Aside from the whole nobody-in-the-driver's-seat thing.

 

[powertoold]

Y'alls pretend like Tesla has no data advantage at all

Waymo's approach is basically the hydrogen of cars: dead end