FSD rewrite will go out on Oct 20 to limited beta

[bdub85]

They’ve got a great base to iterate off of. AI improves at an exponential rate once they hit human level.

 

[Redbrick]

Is it ethical to engage AP if one of your passengers denies you explicit permission to use it?

Man, imagine all those people doubting Elon all this time that kept repeating Tesla needs Lidar and that Waymo will be first to FSD...

I would be feeling pretty rediculous looking back at all the time I insulted others who had faith in Tesla to get it done, to have FSD in every Tesla be end of 2020...

well done Tesla, this has changed the game.

oh just roll that Grenade and walk outta here huh?

 

[powertoold]

Yes, it will get a lot better over time.

A lot faster than we think... I still remember when I didn't think they'd ever release automatic stop and go for traffic controls.

Elon has been spot on lately with AP and FSD developments.

 

[M109Rider]

For comparison, Waymo has a disengagement rate in CA of 1 disengagement per 11,000 miles.

Yes, it will need to be much better.

Wow, did you sell Tesla and buy Waymo stocks or something. ?
Why so negative towards Tesla ?
Their new FSD release is great progress on a much more difficult platform to perfect than Waymo.
That means time, and incremental improvements.
BTW- where do we get released data on things like disengagement’s and other fails from the Waymo cars ?

 

[ChrML]

The update we see now is hopefully the beginning of a ketchup effect. It usually happens in software when you do an architecture revamp that a lot of things down the road depend on. Until you eventually hit the local maximum of the new architecture. So far, so good.

 

[willow_hiller]

Just for clarification, can Tesla achieve L3 with this set of features by having the system programmatically determine when it needs human intervention and alerting the driver in enough time for them to resume control? And then can they achieve L4 by having the car doing the same task as above, but also being able to find a safe place to pull over and putting on the hazards (same fallback performance of dynamic driving task as Waymo), correct?

 

[nepenthe]

The hacker in question has full access to the system and can see all the details of the collection campaigns, and all the actual data being sent to Tesla.

It's exactly as described.

There's no secret magic going on (and indeed later remarks from Elon and Karpathy largely confirmed his findings- all "shadow mode" is doing is targeted campaigns to send specific pictures of specific things back to HQ to be manually labeled by humans for later NN training- and only then does an updated, trained, NN, get sent back to ALL the cars as an OTA update- which is the only time car behavior changes.)


Folks who have these crazy fantasies about how their car "learns" from them or is secretly running magic code to see how its decisions compare to what the human does have no idea WTF they're talking about or how any of this stuff actually works.

He actually did not have access to "all" and has been called out on that fact before.

 

[mikes_fsd]

Wow, did you sell Tesla and buy Waymo stocks or something. ?
Why so negative towards Tesla ?
Their new FSD release is great progress on a much more difficult platform to perfect than Waymo.
That means time, and incremental improvements.
BTW- where do we get released data on things like disengagement’s and other fails from the Waymo cars ?

He's been this way for a while.
He's a Waymo (or more broadly - lidar) evangelist.
It bugs him (even if he says he would be happy to have L4/L5 on his M3) that Tesla can do it without Lidar and that Elon was right all along.

 

[ChrML]

For comparison, Waymo has a disengagement rate in CA of 1 disengagement per 11,000 miles.

Yes, it will need to be much better.

But it's a start, and a much welcome update. As long as there are good leaps at regular intervals, they're good for now.

 

[diplomat33]

Wow, did you sell Tesla and buy Waymo stocks or something. ?

No, I did not sell TSLA. I am long on TSLA. I do not own Waymo stock. I only own TSLA and AAPL stock.

Why so negative towards Tesla ?
Their new FSD release is great progress on a much more difficult platform to perfect than Waymo.
That means time, and incremental improvements.

I am not negative. I love my Model 3. Best car I've ever owned. I posted earlier that I am very excited for the FSD Beta. I think the FSD Beta is amazing progress for Tesla. But being pro-Tesla and being excited for the FSD progress does not mean you can't also recognize limitations.

BTW- where do we get released data on things like disengagement’s and other fails from the Waymo cars ?

The disengagement data comes from the CA DMV report.

 

[powertoold]

Just for clarification, can Tesla achieve L3 with this set of features by having the system programmatically determine when it needs human intervention and alerting the driver in enough time for them to resume control?

The whole L3 L4 thing is silly to talk about in detail because they're flawed definitions and not based in real-world FSD development IMO.

By definition, if a car knows it can't handle something, then it recognizes what it is. The problem with this is that what the car doesn't know is an edge case. It's not possible to know edge cases unless it's seen them before. That's why we've yet to see an actual L3 car. If you can solve for L3, you can solve for "L4" and "L5".

To me, a lot of the SAE definitions become like the simulation paradox. If you can simulate it (edge case identification), then you already know it.

 

[diplomat33]

He's been this way for a while.
He's a Waymo (or more broadly - lidar) evangelist.
It bugs him (even if he says he would be happy to have L4/L5 on his M3) that Tesla can do it without Lidar and that Elon was right all along.

I am a big fan of Waymo, that's all.

No, it does not bug me at all. If Tesla achieves L4/5 without lidar, I will be the first to cheer. I just want real FSD on my car.

 

[diplomat33]

But it's a start, and a much welcome update. As long as there are good leaps at regular intervals, they're good for now.

Absolutely!

 

[NHK X]

I guess I just don't understand why you think that or what evidence you have to support that opinion.

The data the car collects and uploads is not encrypted, there's no reason at all he can't see 100% of everything. (nor is he the only one who sees this stuff, he's just the one who comments publicly on it more often, anybody else with access who has ever commented AFAIK has only confirmed his findings- and again numerous comments by Musk and Karpathy themselves have confirmed it... as recently as *checks* yesterday... with Musk commenting during the earnings call on the data collection of the fleet where the system is looking for specific corner case example to train against- exactly the targeted campaign system Green has described for a while now)


Certainly the re-write might do new things other than just collect specific largely static data as Green has cited- but the version 99.9% of everyone has been running since 2016 does not.

If I remember right I thought he didn’t see much of the data transfer going over cellular.

And there is also stuff he self admittedly says he doesn’t see or know. Green is great, I think he’s an asset, but does he know all? No.

 

[M109Rider]

Clickbait from the Verge, the facts in the article mostly positive, but lots of gloom and doom from the writer (not all inappropriate but the general impression for the reader is "OMG are they crazy?!"

Tesla’s "Full Self-Driving" beta is here, and it looks scary as hell

Total click bait. Clearly written from someone who is not on Tesla’s side.
Dime a dozen.

 

[diplomat33]

He's been this way for a while.
He's a Waymo (or more broadly - lidar) evangelist.
It bugs him (even if he says he would be happy to have L4/L5 on his M3) that Tesla can do it without Lidar and that Elon was right all along.

FYI, you probably should let me speak for myself, especially since you are mischaracterizing me. Let me tell people where I am coming from.

 

[AlanSubie4Life]

You never offered to bet that they'd achieve Level 4. You just bet they would not upgrade the sensors. If they don't upgrade the sensors and never achieve Level 4 you win your bet. You're right: I'm glad I didn't take that bet.)

Perhaps you should have taken the bet?

I think the way the bet was worded you wouldn't have won because it would require retrofitting to older vehicles (??? I don't track all the details of exactly how it was worded), but you would have had a moral victory; new sensors are probably coming (unsurprisingly) - and these aren't just "iterations on the camera so they don't have to source old discontinued sensors" - they're likely real capability improvements, as would be expected from any rational developer of autonomous driving:

That's a much better (and heated apparently) radar, and I'd guess the new ultrasonic sensors aren't just going to be a cost-reduction measure in keeping with Tesla's vertical integration.

Complete thread:
https://twitter.com/greentheonly/status/1319363500695785473?s=20

We'll see if it pans out!

 

[Cheburashka]

Hopefully, they will release this to a wider range of testers and we will get some good videos of how it works and what the limitations are. To me, it's not really clear why these people were selected to test and post videos of the new functionality. Does anyone know?

So far it looks pretty good. Smart summon also looked decent from a lot of the early videos but turned out to be pretty useless after it was actually launched though.

 

[helvio]

Found this useful link, rejoice!

All FSD footage : teslamotors

 

[powertoold]

He's been this way for a while.
He's a Waymo (or more broadly - lidar) evangelist.

I think we should celebrate the fact that there are differing approaches to FSD. Even if Tesla ends up with widespread robotaxis, it's nice to know there are other options in case the Tesla network has an issue, for example.

A lot of us enjoy Elon's assertiveness (and hubris) when it comes to tech, so we tend to hold some of his strong beliefs, but at the end of the day, it's great that Waymo is targeting an "overly safe" approach vs Tesla.