Autonomy Investor Day - April 22 at 2pm ET

[CarlK]

I know some of you are laughing at the last comment, but I honestly think that it's the answer Elon is proposing to the question. In order to be at the point where you can self-drive a car, you have to have some serious machine learning for vision as the primary system.

That's what he is working on, and suggesting that the results are exponential. In the end we will wait and find out, but I think the wait won't be as long as many think.

True. Whoever think that was funny was laughing at his own ignorance.

Vision + machine learning is able to do everything human drivers could with added advantages including ability to do millions times more of learning and trail and error than human could. Lidar + HD mapping can get you to a point, maybe faster, but it just can not do everything human can (that Lavenlowski's short interview has a good explanation of that). It will eventually run into road blocks (sorry about the pun). Our road system is designed for human drivers. The minimum requirement for a self driving car is be able to do everything human drivers could. The Tesla vision based system is the only one at this moment can meet this requirement.

And how do you know that? Did they ever say the AP2 demo was a lucky run with hard coded parts? No they never said that, but maybe they weren’t really confident, that FSD will happen soon.

But now Elon is really confident and says:


Oh wait, that was back in 2016. So now 3 years later FSD is still just 1 year out. And that’s what I don’t get. Elon has pulled that trick off before and people are again falling for it! And I bet in 2022 there will be a big announcement about how FSD will definitely happen in 2023 and people will applaud and cheer...

When he started the project people had very little knowledge and experience with this technology. Elon made the best guess he could make, and as usual on the most optimistic side. We all know a lot more now. More accurate prediction is much easier to make this time around. Like American Airline can make much better prediction of cross Atlantic travel time than Christopher Columbus could.

 

[diplomat33]

And how do you know that? Did they ever say the AP2 demo was a lucky run with hard coded parts? No they never said that, but maybe they weren’t really confident, that FSD will happen soon.

But now Elon is really confident and says:

Oh wait, that was back in 2016. So now 3 years later FSD is still just 1 year out. And that’s what I don’t get. Elon has pulled that trick off before and people are again falling for it! And I bet in 2022 there will be a big announcement about how FSD will definitely happen in 2023 and people will applaud and cheer...

There is a big difference between 2016 and 2019. The 2016 FSD demo was done by porting AP1 software + some extra code unto the HW2 hardware. Tesla thought that if they take AP1 which was really good and combine it with the extra cameras of HW2 and write some extra code to handle turning at intersections and such, then they could do FSD. That strategy failed in part because of the Mobileye break up and without Mobileye, Tesla struggled to duplicate AP1. Plus, hardcoding for all the edge cases doesn't work in the long term. Remember too that we never got anything like "Autonomy Investor Day" back in 2016. Tesla said nothing about how they were doing FSD or what their road map was or explain their camera vision process. Tesla never even explained how they did the demo video. Nothing. We just got the announcement about the new HW2 hardware and a cool demo video.

2019 is different because Tesla is using a completely different method now, one that is working. Tesla is using machine learning and neural nets which are the right approach for doing FSD. And Tesla has the AP3 computer which is much more powerful than AP2 and designed to handle the FSD software.

FSD will work this time because Tesla is using the right approach now. Plus, we have seen with our own eyes the progress that Tesla has made from the early days of AP2 to Nav on AP, to the real FSD demo on Autonomy Investor Day. We have a list of FSD features and proof that Tesla really is close to feature complete. So we got something more concrete this time unlike in 2016 where we just a demo and vage promises of what FSD would do. So there are good signs that AP3 will be able to do FSD. Plus, what we saw on Autonomy Investor Day, was just alpha. Just think how good FSD will be in a couple more years when Tesla has worked on it even more and taken it beyond "feature complete"! When we get the final product for FSD, I think it will be very good!

 

[mongo]

There is a big difference between 2016 and 2019. The 2016 FSD demo was done by porting AP1 software + some extra code unto the HW2 hardware. Tesla thought that if they take AP1 which was really good and combine it with the extra cameras of HW2 and write some extra code to handle turning at intersections and such, then they could do FSD. That strategy failed in part because of the Mobileye break up and without Mobileye, Tesla struggled to duplicate AP1. Plus, hardcoding for all the edge cases doesn't work in the long term. Remember too that we never got anything like "Autonomy Investor Day" back in 2016. Tesla said nothing about how they were doing FSD or what their road map was or explain their camera vision process. Nothing. We just got the announcement about the new HW2 hardware and a cool demo video.

2019 is different because Tesla is using a completely different method now, one that is working. Tesla is using machine learning and neural nets which are the right approach for doing FSD. And Tesla has the AP3 computer which is much more powerful than AP2 and designed to handle the FSD software.

FSD will work this time because Tesla is using the right approach now. Plus, we have seen with our own eyes the progress that Tesla has made from the early days of AP2 to Nav on AP, to the real FSD demo on Autonomy Investor Day. We have a list of FSD features and proof that Tesla really is close to feature complete. So we got something more concrete this time unlike in 2016 where we just a demo and vage promises of what FSD would do. So there are good signs that AP3 will be able to do FSD. Plus, what we saw on Autonomy Investor Day, was just alpha. Just think how good FSD will be in a couple more years when Tesla has worked on it even more and taken it beyond "feature complete"! When we get the final product for FSD, I think it will be very good!

The 2016 video was made after the mobileye breakup. IMO it was a proof of concept to reassure buyers that AP2 could do fsd, because it was being released with zero features (due to the breakup).

As to the current state of things:Watch Tesla Autopilot Successfully Make Left Hand Turn At Stoplight
Yes, they say 45 instead of 90 like 3 times...

 

[diplomat33]

The 2016 video was made after the mobileye breakup. IMO it was a proof of concept to reassure buyers that AP2 could do fsd, because it was being released with zero features (due to the breakup).

Correct. I couldn't remember if it was before or after the break up. But my general point is still correct. A proof of concept video for new hardware is very different from an "Autonomy Investor Day" where Tesla gave us a deep dive into their FSD process and gave test drives of their current alpha software which as far as I can tell, has received very positive reviews. I think we have more evidence now that Tesla really is making progress with FSD that we did not have in 2016.

As to the current state of things:Watch Tesla Autopilot Successfully Make Left Hand Turn At Stoplight
Yes, they say 45 instead of 90 like 3 times...

That's still AP2, isn't?

 

[mongo]

That's still AP2, isn't?

Probably 2.x, definitely the non HW3.0 specific NN.

 

[CarlK]

Other than HW3 there is also software 2.0 neural net. Some if that had already been implemented in at least AP2. I remember someone posted a video that the system seemed to know car on adjacent lane was going to change lanes and took evasive action. That was explained on Monday that the NN now can predict other car's intention by observe its small movements just like we human do.

 

[diplomat33]

Probably 2.x, definitely the non HW3.0 specific NN.

Watching the video, it seems like Autopilot is just following the curved dashed lane lines that happen to curve around in that particular intersection. So it's not actual intersection handling that we will get in AP3.

 

[mongo]

Watching the video, it seems like Autopilot is just following the curved dashed lane lines that happen to curve around in that particular intersection. So it's not actual intersection handling that we will get in AP3.

Yah, the driver mentions the well marked lane. Still, didn't AP have an issue with tighter turns previously?

 

[diplomat33]

Yah, the driver mentions the well marked lane. Still, didn't AP have an issue with tighter turns previously?

Maybe but it has obviously gotten a lot better recently. I know that AP in my car handles very tight turning in exits perfectly.

 

[croman]

Yah, the driver mentions the well marked lane. Still, didn't AP have an issue with tighter turns previously?

I've been able to coax AP into taking well marked left turns for over a year but only with a lead car.

 

[acoste]

I have to admit I used to still have some doubts, or questions, of how a self driving car or Tesla FSD could reach absolute L5 capability. Those doubts were completely removed after the Monday's event.

I wonder how come you had had doubts? You kept telling us Tesla is the best.

Lidar + HD mapping can get you to a point, maybe faster, but it just can not do everything human can

In fact it can do more regarding sensing. Much better at night for example.

Using Lidar is added safety. Tesla is ok with some fatalities. Other car makers are not.

When you can select between several taxi services. Which one would you pick?

The minimum requirement for a self driving car is be able to do everything human drivers could. The Tesla vision based system is the only one at this moment can meet this requirement.

Human eyes are better than Tesla cams. Human brain is better than any of these NN machines.

You keep saying computers are better than humans, but then I have a very simple question to you. How come a human can learn something by looking at it 1 or 2 times while NN needs millions of repetitions?
How does an NN respond to an unknown object on the road and how does a human?

 

[CarlK]

@acoste Why are you still trolling here after people exposed your BMW and Porsche forum connections? Everything you've been saying about Tesla's vision and machine learning was proven wrong in Monday's presentation and demo. Go watch and cry!

I forgot about you but now you're on my ignore list where you belong too.

 

[electronblue]

@acoste Why are you still trolling here after people exposed your BMW and Porsche forum connections? Everything you've been saying about Tesla's vision and machine learning was proven wrong in Monday's presentation and demo. Go watch and cry!

I forgot about you but now you're on my ignore list where you belong too.

An ad hominem when you run out of arguments, really?

 

[electronblue]

The 2016 video was made after the mobileye breakup. IMO it was a proof of concept to reassure buyers that AP2 could do fsd, because it was being released with zero features (due to the breakup).

The same could be said of April 22nd though — a reassurance to a stock market this time that Tesla could do FSD and a way to take their eyes off a bit from apparent demand issues.

There is another correlation, too: As in 2016, it comes down to whether or not you believe Elon Musk’s statements.

 

[mongo]

The same could be said of April 22nd though — a reassurance to a stock market this time that Tesla could do FSD and a way to take their eyes off a bit from apparent demand issues.

There is another correlation, too: As in 2016, it comes down to whether or not you believe Elon Musk’s statements.

I was addressing the impact of the mobileye breakup and the use of ap1 code in the replied to comment. Specifically:

The 2016 FSD demo was done by porting AP1 software + some extra code unto the HW2 hardware. Tesla thought that if they take AP1 which was really good and combine it with the extra cameras of HW2 and write some extra code to handle turning at intersections and such, then they could do FSD. That strategy failed in part because of the Mobileye break up and without Mobileye, Tesla struggled to duplicate AP1

No AP1 HW was involved in that video.

You could look at the recent demos that way, I suppose. But the 2016 video was accompanied by the news that AP2 was going to have zero features to start with, whereas the HW3 demo was accompanied with short, medium, and long term timelines along with known current and demonstrated (at some level) future capabilities. On the company value side, the current order page working is also favorable towards revenue inclusion and the voulme of cars with AP is increasing.

As seen with every disengagement report, people were taking no news/ demo as bad bews/ demo. So, there was a PR IR side of things to address.

 

[electronblue]

@mongo Fair. And yes I agree the October 2016 video of course had nothing to do with MobilEye. Any suggestion otherwise is just revisionist history usually trying to give Tesla an ”out”. No, they knew what they announced in 2016 and it was all months after the MobilEye break-up...

 

[rnortman]

All it will need is to give the machine more miles of learning experiences.

Amazing! And the most amazing thing is that only Tesla has figured out that you just pump miles into machine learning and get a self-driving car out! All those idiots investing billions in developing this technology; if only they were as smart as Elon. Well I guess the cat is out of the bag now, right? And we're going to see Waymo do a complete about-face, ditch the LAME lidar, replace their nice cameras with cell phone cameras, and just go for MOAR MILEZ. I mean, Tesla just showed everybody how simple this is, right? You'd have to be an idiot not to get it.

 

[rnortman]

The minimum requirement for a self driving car is be able to do everything human drivers could. The Tesla vision based system is the only one at this moment can meet this requirement.

You do know that every company developing self-driving technology has cameras (usually with 360-degree coverage) and uses deep learning to process the camera data, right? Absolutely every company out there does this. Most of them have better cameras and more on-board compute power. If Tesla has an advantage, it's not HW3 or "machine learning" in general, it's the fabled "fleet data". This is not a sustainable advantage, if it's an advantage at all. If you can get L5 robotaxi-level autonomy with cheap cameras plus cheap GPU/TPU plus lots and lots of miles, then many companies will do exactly that, one way or another.

 

[rnortman]

When he started the project people had very little knowledge and experience with this technology. Elon made the best guess he could make, and as usual on the most optimistic side. We all know a lot more now. More accurate prediction is much easier to make this time around. Like American Airline can make much better prediction of cross Atlantic travel time than Christopher Columbus could.

When he started the project, he had very little knowledge and experience with this technology. Some people have been working on autonomous vehicles since the 80s, and even earlier. Deep learning wasn't new in 2016 either, and people were already using it for autonomous vehicles.

 

[brianman]

Tesla is ok with some fatalities. Other car makers are not.

Links to support your assertions?