Elon: "Feature complete for full self driving this year"

[linux-works]

if you are writing for the linux platform (and tesla does use lots of linux inside; probably not exclusively, but for a lot) then being trained in the windows world actually puts you pretty much behind the curve, in my experience. there is so much you have to train people who have not spent time (or even had any interest!) in unix. its not the user commands (that's part of it) but its how unix does things and how to find the right way to solve something with unix.

we burn out our developers. we use them, throw them out (or even lay them off; bay area has a culture of 'regular layoffs' and I'm not kidding, either), and rehire cheaper, younger ones. lather, rinse, repeat (at they say). and if you're over 35 and in software in silicon valley - well, lets just say - many companies will prefer someone younger. its toxic, this work environment. and its been this way for a long while, now (I've been here over 25 years, myself).

people need to realize that tesla is, in many ways, very much like the majority of bay area hw/sw companies. and yes, quality in hw/sw has gone way down over the years for lots of reasons; none of them good.

I'll finish on a high point: given ALL THAT, I'm actually impressed with how good the model 3 is (I'm 3 months on, in my ownership). I have a feeling of how much work is involved in getting a car like this out and I'm kind of impressed with it. they don't have as many people as they need, they don't really have as big a budget of money or time as they need and they do amazing work for the very short time they have been around.

 

[Az_Rael]

You are late. I just posted it 20 minutes ago.

Yes, it is a new rewrite.

Sorry about that - missed your earlier post.


So with a fundamental rewrite being worked on now, I am guessing we might be over a year away from seeing new FSD features released to the full fleet (not EAP).

 

[S4WRXTTCS]

I wonder how that would go down. I sold my HW2.0 MCU1 car and I got the class action punishment checks and not the current owner after the fact. If they refund some significant amount of $$$$ for the FSD purchase price for cars ordered with FSD I wonder if the current owner or the original purchaser would get the $$$$.

The class action lawsuits were due to the failure of bringing out the features in a reasonable time frame. So it went to the impacted customer at the time.

Any kind of refund for FSD on AP2 will be tied to the car, and the current owner when/if FSD is refunded.

Like if I bought a used AP2 car with FSD today I would fully expect that I would be the one getting the refund, and not the previous owner. That's because Tesla/Elon up to this point has claimed that it will get HW3/FSD.

If the individual sold the car to me without taking into account the value of FSD that's not my problem.

If someone owns a Model S/X with AP2 and FSD they should assume the FSD is worth $3K today just like it was worth $3K when they bought it.

 

[S4WRXTTCS]

Sorry about that - missed your earlier post.
So with a fundamental rewrite being worked on now, I am guessing we might be over a year away from seeing new FSD features released to the full fleet (not EAP).

I agree that we potentially might not see any new FSD features within the full fleet. Anything beyond what it currently introduces a lot of danger. So much so that I predicted 2020 would be the death of L2 based FSD.

It looks like I'm wrong, and instead FSD development is going to take a different path. Where we're likely going to see much better performance for the current features before they release the dangerous stuff.

SW does tend to follow a flow where every other release will focus on fixing bugs, and improving the overall operation. Essentially under the hood stuff.

That's extremely important with FSD because as it currently stands a lot of us including myself are really irked about how bad it works. Where we're not able to use NoA for very long without it screwing up in way that is completely unacceptable.

 

[eli_]

I wonder how big the AP teams are. Elon likes to run things really lean or the software will get bloated (which is true).

 

[mongo]

The class action lawsuits were due to the failure of bringing out the features in a reasonable time frame. So it went to the impacted customer at the time.

If we are talking about the lawsuit I think we are: The issue was people ordered cars when AP1 was in production and had a defined feature set. Then they were delivered AP2 cars which did not have those features for a period of time. The money was compensation for the time they had the cars without the expected, preexisting, features.

 

[diplomat33]

Sorry about that - missed your earlier post.

So with a fundamental rewrite being worked on now, I am guessing we might be over a year away from seeing new FSD features released to the full fleet (not EAP).

I think we will see incremental improvements this year. Since the rewrite involves the NN labeling the path of the car in 3D, I am hoping we see improvements like better lane keeping especially on tight curved roads, or smoother on and off ramps. But I agree with you that a major release like automatic city driving will probably be delayed to the end of this year because of how big a rewrite it is. There is a lot of work that will go into this rewrite and it will probably take longer than Elon estimates. Although, we might see traffic light and stop sign response released separately first and then automatic city driving" released later like it suggests on the FSD order page.

I kinda wonder if maybe Tesla tried to do City NOA and found it to be too unreliable and they concluded that they really needed the NN to stitch together all 8 cameras and do perception and planning at the same time. It is also obvious to me that this rewrite is how Tesla is planning to do intersections. Basically, the NN will both see and label an intersection and also label the path to navigate the intersection like turning left or right or moving over to the turning lane etc.. Plus, the video labeling of paths can also be used for things like going around a double parked for example. The NN can both label the double parked car and label the path for how to go around it. So this rewrite represents a major step towards "city self-driving". When the rewrite is done, I think we will see a much better AP.

I would argue that this rewrite probably represents the true divergence between EAP and FSD. I also think we are seeing the third generation of Autopilot. Gen 1 was AP1. Gen 2 was EAP on AP2/AP2.5. Gen 3 is FSD on AP3.

 

[Az_Rael]

I kinda wonder if maybe Tesla tried to do City NOA and found it to be too unreliable and they concluded that they really needed the NN to stitch together all 8 cameras and do perception and planning at the same time

I was surprised to hear those comments that they weren't already working on sensor fusion. It seems like it would be absolutely required to get any of this working in a complex environment.

I am becoming resigned to the fact that I paid $3k for a computer upgrade, but will probably not see any other benefits during my car ownership period. Ah well.

 

[diplomat33]

I was surprised to hear those comments that they weren't already working on sensor fusion. It seems like it would be absolutely required to get any of this working in a complex environment.

I don't know for sure. But the fact that they promised "feature complete" by end of last year and Elon says that they came close but now tells us that they are doing a major rewrite of the entire AP code, seems to support my theory. And it seems eerily similar to what happened at the beginning of AP2 when Elon was promising FSD features, missed the deadlines and then Karpathy comes in and we learn he rewrote major parts of the AP code (2018.10.4).

I agree that it would be really odd not to work on sensor fusion sooner. But frankly, Tesla's entire FSD approach from day 1 appears to be pretty clumsy. I mean, we know from a Karpathy presentation that Tesla had mostly software 1.0 when they started FSD in 2016. Tesla even put out the FSD demo and sold FSD on that basis. When they struggled to get to AP1 parity, Elon hired Karpathy and they shifted to machine learning. Elon said that AP2 would be good enough for FSD, then realized the AP2 computer would not be good enough so they created the AP3 computer. Now, 3 year later, they are doing sensor fusion and video labeling, something that is pretty basic and Elon even admits, very foundational to autonomy. That does not look like a coherent approach to FSD. But they do seem to slowly be coming around to the right approach.

I remain optimistic that Tesla will eventually get. Who knows? Maybe this new rewrite will get to good FSD with driver supervision but Tesla will finally realize that they need to add more radar and even lidar to get to the safety necessary for a true safe driverless robotaxi.

I am becoming resigned to the fact that I paid $3k for a computer upgrade, but will probably not see any other benefits during my car ownership period. Ah well.

For your sake, I hope you get some additional FSD benefits while you still have your car.

 

[DigAdrenaline]

if you are writing for the linux platform (and tesla does use lots of linux inside; probably not exclusively, but for a lot) then being trained in the windows world actually puts you pretty much behind the curve, in my experience. there is so much you have to train people who have not spent time (or even had any interest!) in unix. its not the user commands (that's part of it) but its how unix does things and how to find the right way to solve something with unix.

The best resource for introducing someone to *nix is Think Unix, by John Lassar.

Think UNIX https://www.amazon.com/dp/078972376X/ref=cm_sw_r_cp_api_i_uK2mEb1ZAJQ89

It helps them understand why nix works as it does, and it’s a hands on read. The only possible complaint is that it’s old, so SystemV and other things have changed significantly. I think it’s based on Red Hat 3 or 5 (pre RHEL). It’s still a worthwhile read though for a quick but comprehensive introduction to the Unix family.

 

[linux-works]

guys, a 're-write' is a GOOD THING.

sometimes, old technical debt is better left behind.

its not often that you get a chance to say 'well, this was a good learning experience. now, lets write the real code.'

this will delay things (I bet its been in the works a while, now) - but I'm quite encouraged by the fact that mgmt is ok'ing a "re-write" of some key systems.

 

[willow_hiller]

guys, a 're-write' is a GOOD THING.

sometimes, old technical debt is better left behind.

its not often that you get a chance to say 'well, this was a good learning experience. now, lets write the real code.'

this will delay things (I bet its been in the works a while, now) - but I'm quite encouraged by the fact that mgmt is ok'ing a "re-write" of some key systems.

Honestly, this is the advantage of Tesla being a flexible dev team. It's the development version of the "sunk cost" fallacy. Just because time has been expended on one methodology doesn't mean another can't yield better results.

 

[nepenthe]

I don't know for sure. But the fact that they promised "feature complete" by end of last year and Elon says that they came close but now tells us that they are doing a major rewrite of the entire AP code, seems to support my theory. And it seems eerily similar to what happened at the beginning of AP2 when Elon was promising FSD features, missed the deadlines and then Karpathy comes in and we learn he rewrote major parts of the AP code (2018.10.4).

I agree that it would be really odd not to work on sensor fusion sooner. But frankly, Tesla's entire FSD approach from day 1 appears to be pretty clumsy. I mean, we know from a Karpathy presentation that Tesla had mostly software 1.0 when they started FSD in 2016. Tesla even put out the FSD demo and sold FSD on that basis. When they struggled to get to AP1 parity, Elon hired Karpathy and they shifted to machine learning. Elon said that AP2 would be good enough for FSD, then realized the AP2 computer would not be good enough so they created the AP3 computer. Now, 3 year later, they are doing sensor fusion and video labeling, something that is pretty basic and Elon even admits, very foundational to autonomy. That does not look like a coherent approach to FSD. But they do seem to slowly be coming around to the right approach.

I remain optimistic that Tesla will eventually get. Who knows? Maybe this new rewrite will get to good FSD with driver supervision but Tesla will finally realize that they need to add more radar and even lidar to get to the safety necessary for a true safe driverless robotaxi.



For your sake, I hope you get some additional FSD benefits while you still have your car.

A major rewrite that is almost complete, would have presumably been going on for a while. And reasonably, if it was in-process while all the added development we have seen has been on-going, one could easily be talking about a year plus of work already.

 

[eli_]

Cool augmented video of the vision networks in action:
https://www.tesla.com/sites/default/files/images/careers/autopilot/network.mp4

And screenshot of Autopilot code:
https://www.tesla.com/sites/default/files/images/careers/autopilot/code.gif

 

[willow_hiller]

Cool augmented video of the vision networks in action:
https://www.tesla.com/sites/default/files/images/careers/autopilot/network.mp4

And screenshot of Autopilot code:
https://www.tesla.com/sites/default/files/images/careers/autopilot/code.gif

Thanks for posting these. Are they linked somewhere in a blog post? I'd love more context.

 

[eli_]

Thanks for posting these. Are they linked somewhere in a blog post? I'd love more context.

They're on this recruiting page: Autopilot AI

I direct linked them because they're scaled down very small on the actual page.

 

[BigD0g]

It’s certainly not the first or second or even third rewrite..... who knows anymore....

 

[Eno Deb]

Here is a clip from Third Row Tesla's podcast interview with Elon Musk. Elon explains a bit more about what he alluded to in the earnings call about video labeling and gives some new info:

Significant rewrite in AP system almost complete: instead of having planning, perception, image recognition being separate, they are being combined together. "NN is absorbing more and more of the problem."

This is nothing new. Karpathy already mentioned that goal at the "autonomy day" last year. I wonder how far along they are with this.

Video labeling will mean that the NN will label paths in 3D that the car should take.

This probably refers to training the neural net to process sequences of images rather than individual images, which can be used to estimate distances with a monocular camera when the car is in motion (by observing the the motion parallax between subsequent images).

 

[Eno Deb]

guys, a 're-write' is a GOOD THING.

Yeah, well, depends. This is at least the second "rewrite" since Karpathy joined, and before that was Lattner's "extensive rewrite" ...

 

[linux-works]

still, this stuff is a huge learning exercise for everyone. being able to know, somewhat early, that this dog^H^H^Harchitecture won't hunt - that's wisdom. and actual 'courage' (don't talk to me about apple, please.)