Autonomy Investor Day - April 22 at 2pm ET

[diplomat33]

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.

No, just pumping a ton of data into machine learning blindly, will not just magically give you a self-driving car. The key is getting the right data, using that data correctly and what you do with the machine learning on the output end. Machine learning is kinda like sex, "it's not the size that matters, it's how you use it."

But the big data that Tesla has access to, is certainly helpful towards developing a self-driving car. As Tesla explained during Autonomy Investor Day, there are different helpful ways that they can leverage the big fleet. For example, they can collect very specific real world data like cars with bicycles on the rack and use that data to train the machine. Or, when Tesla has software that they think that can handle a specific situation, they can upload it into a lot of cars in "shadow mode" and collect feedback to see how it would perform in the real world. You can do these things with simulations of course, but having a large fleet of cars, means that you can do it in the real world, in real world conditions and you can do it faster and on a bigger scale. Again, that's going to be helpful.

So, Tesla's big fleet does not guarantee full self-driving but it is a very helpful tool.

 

[croman]

Human eyes are not better than Tesla's cameras. Lidar isn't better at night only complete darkness which doesn't exist in real life.

 

[CarlK]

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.

He started the project around the same time he founded OpenAI. No one knows everything but he knew way more about AI and machine and their impacts on self driving cars than anyone in the industry and way way more than those pundits. Look, even people like Anthony Levandowski now says Elon is wiser than me and he is right. Us much lessor people should not try to oursmart ourselves.

No, just pumping a ton of data into machine learning blindly, will not just magically give you a self-driving car. The key is getting the right data, using that data correctly and what you do with the machine learning on the output end. Machine learning is kinda like sex, "it's not the size that matters, it's how you use it."

But the big data that Tesla has access to, is certainly helpful towards developing a self-driving car. As Tesla explained during Autonomy Investor Day, there are different helpful ways that they can leverage the big fleet. For example, they can collect very specific real world data like cars with bicycles on the rack and use that data to train the machine. Or, when Tesla has software that they think that can handle a specific situation, they can upload it into a lot of cars in "shadow mode" and collect feedback to see how it would perform in the real world. You can do these things with simulations of course, but having a large fleet of cars, means that you can do it in the real world, in real world conditions and you can do it faster and on a bigger scale. Again, that's going to be helpful.

So, Tesla's big fleet does not guarantee full self-driving but it is a very helpful tool.

Tesla NN has enough data on all common cases already. It's no longer more data but more miles to cover enough tail cases, like ones you described, which you don't see often. The big fleet is the only way for you to get all those info. No other company has the capability to have it in the near future.

 

[diplomat33]

Tesla NN has enough data on all common cases already. It's no longer more data but more miles to cover enough tail cases, like ones you described, which you don't see often. The big fleet is the only way for you to get all those info.

I am not sure a big fleet is the only way to get all those special edge cases but it is certainly one way to do it, that really works. Certainly, the larger the fleet, the higher the odds that one of your cars encounters one of those tail cases that you are looking for, in a timely manner. And we see that confirmed during the AID event. When Tesla did their wide roll out of NOA with stalk confirm, they were able to collect over 9 million successful stalk confirmations! This many successful lane changes allowed Tesla to validate the software's reliability in just 6 months, to where they could remove the stalk confirmation. A smaller fleet or simulations can do it too but it will probably take longer.

 

[CarlK]

I am not sure a big fleet is the only way to get all those special edge cases but it is certainly one way to do it, that really works. Certainly, the larger the fleet, the higher the odds that one of your cars encounters one of those tail cases that you are looking for, in a timely manner. And we see that confirmed during the AID event. When Tesla did their wide roll out of NOA with stalk confirm, they were able to collect over 9 million successful stalk confirmations! This many successful lane changes allowed Tesla to validate the software's reliability in just 6 months, to where they could remove the stalk confirmation. A smaller fleet or simulations can do it too but it will probably take longer.

I was only qouting Elon and Karpathy but that's pretty easy to figure out. You see a common case every mile and billion times over a billion miles. On the other hand you see rare cases like a car or bike flies in front you only once millions of miles if not more. You still need to learn how to handle those situations. There is no short cut of getting enough data of any of these so the machine can be trained.

Elon also said simulation does not work. The self driving problem is already solved if we know how to set up the correct simulation that covers all tail cases. From what I understood Waymo no longer uses simulation because of that.

 

[Eno Deb]

Elon also said simulation does not work. The self driving problem is already solved if we know how to set up the correct simulation that covers all tail cases. From what I understood Waymo no longer uses simulation because of that.

You couldn't be more wrong. Waymo's Carcraft simulator is very much alive, and every other company in this field is also heavily investing in simulators. Simulation has a lot of advantages over fleet data. You can run specific scenarios in lots of "fuzzed" variations. If you rely on fleet data, you first have to hope that the cars actually encounter the scenario you need with a sufficient frequency (most of Tesla's "billions of miles" are probably monotonous highway driving), and then hope that you have the right triggers in place in the right cars to actually be able to harvest the data. I don't know anything about Tesla's simulator, but I doubt that it's anywhere near as advanced as Carcraft.

Here's an easy to understand blog post about the rationale behind Carcraft:

How simulation turns one flashing yellow light into thousands of hours of experience

 

[Dutchie]

The reason credibility of "feature complete" at end of this year is now much higher than a couple of months ago is that during the presentation three other guys (SME's) were presenting and not just Elon. I have not heard one of them contradicting what Elon said.

 

[Dutchie]

You couldn't be more wrong. Waymo's Carcraft simulator is very much alive, and every other company in this field is also heavily investing in simulators. Simulation has a lot of advantages over fleet data. You can run specific scenarios in lots of "fuzzed" variations. If you rely on fleet data, you first have to hope that the cars actually encounter the scenario you need with a sufficient frequency (most of Tesla's "billions of miles" are probably monotonous highway driving), and then hope that you have the right triggers in place in the right cars to actually be able to harvest the data. I don't know anything about Tesla's simulator, but I doubt that it's anywhere near as advanced as Carcraft.

Here's an easy to understand blog post about the rationale behind Carcraft:

How simulation turns one flashing yellow light into thousands of hours of experience

As I remember correctly is that Tesla also - in addition to - uses simulation

 

[Eno Deb]

As I remember correctly is that Tesla also - in addition to - uses simulation

Sure (if you read again you'll see that I mentioned it). But the previous poster implied that the industry was abandoning simulations. That's completely wrong.

Anyway, Musk's statement saying "if simulations worked we could prove that we live in a simulation" is one of the stupidest things he said at the event. He's obviously a smart guy, but sometimes he's just full of it. Reminds me of his "alien dreadnought" that was supposed to teach everyone else how to build cars ...

 

[Dutchie]

Sure (if you read again you'll see that I mentioned it). But the previous poster implied that the industry was abandoning simulations. That's completely wrong.

Anyway, Musk's statement saying "if simulations worked we could prove that we live in a simulation" is one of the stupidest things he said at the event. He's obviously a smart guy, but sometimes he's just full of it. Reminds me of his "alien dreadnought" that was supposed to teach everyone else how to build cars ...

I see it now. Read over that sentence..

 

[croman]

Let's not drink so deeply and to appreciate AP will still try to kill people in ordinary situations it makes harrowing. Just see the weekly submissions to Reddit where people post dashcam of their cars getting confused by aggressive driver's and getting flustered and going for gore points or concrete dividers or only lanes.

Thank God Tesla isn't as arrogant and willfully blind and some die hards on this forum. Let them work this out. It's been an embarrassingly long time since AP2 launched with outrageous lies about AP1 parity by December 31 2016 and we're finally starting to see some meaningful improvements.

 

[timewasted]

The reason credibility of "feature complete" at end of this year is now much higher than a couple of months ago is that during the presentation three other guys (SME's) were presenting and not just Elon. I have not heard one of them contradicting what Elon said.

Really though, that’s not exactly a high bar to clear. If you value being employed you don’t make a habit of discrediting your boss when he’s talking to a room full of investors.

 

[Dutchie]

Really though, that’s not exactly a high bar to clear. If you value being employed you don’t make a habit of discrediting your boss when he’s talking to a room full of investors.

At this level, I am not so sure. These guys are employable everywhere and I bet they have headhunters calling them weekly. The fact that they stay with Tesla for these years says a lot.

 

[croman]

Jim Keller is conspicuously absent from the discussion of the chip. Everyone is just going along with Pete Bannon without crediting the likely architect of HW3 who is now absent. The design of HW3 was finished right as Keller made his announcement to leave for Intel...

Don't think the remaining talent is anything but...

Karpathy is the only one that probably could leave and collapse the AP program. This is his dream job though, so other than Elon's poor temperament, I'm not sure he'd leave for anywhere else short of his own dream Google project.

 

[electronblue]

The reason credibility of "feature complete" at end of this year is now much higher than a couple of months ago is that during the presentation three other guys (SME's) were presenting and not just Elon. I have not heard one of them contradicting what Elon said.

They also, interenstingly, did not confirm any of Musk’s timeline claims.

If Karpathy had said all this will be feature complete by the end of the year, that certainly would have been noteworthy. But at every turn Elon was the one making bigger and bolder claims next to the technical guys reporting what was a much more down to Earth (and thus much more believable) report on current status.

 

[electronblue]

At this level, I am not so sure. These guys are employable everywhere and I bet they have headhunters calling them weekly. The fact that they stay with Tesla for these years says a lot.

I’m not sure that is an argument you want to make regarding Tesla’s AP team. The turnover has been horrific ever since Musk’s original claims.

 

[acoste]

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

Links to support your assertions?

This article explains it well:

Tesla has a self-driving strategy other companies abandoned years ago

Google/Waymo Stopped Testing Level 3 Self-Driving Tech After Testers Literally Fell Asleep While Using It, Switched To Full Autonomy

"Over the last three years, most other companies working on self-driving technology have followed Waymo's lead.
Tesla, meanwhile, has stubbornly pushed forward with its original strategy"


Volvo is now added to the list of carmakers pulling-away from deployment of Level 3 automated driving (at least for now) with the CEO characterizing the handover of vehicle control as unsafe.

Volvo's vision: "Our vision is that by 2020 no one should be killed or seriously injured in a new Volvo car."

Volvo Vision 2020


And Tesla/Musk is fully aware of the problem.

Not only did they skip the fact that the product was not ready:

"Weeks before the October 2015 release of Autopilot, an engineer who had worked on safety features warned Tesla that the product wasn’t ready, according to a resignation letter circulated to other employees and reviewed by the Journal.
Autopilot’s development was based on “reckless decision making that has potentially put customer lives at risk,” the engineer, Evan Nakano, wrote."

they also didn't care about the vigilance issue.

Musk himself said this that shows he is aware of it:

“When there is a serious accident it is almost always, in fact maybe always, the case that it is an experienced user, and the issue is more one of complacency,” Musk said on a quarterly earnings call earlier this month. “They just get too used to it. That tends to be more of an issue. It’s not a lack of understanding of what Autopilot can do. It’s [drivers] thinking they know more about Autopilot than they do.”

Tesla rejected more advanced driver monitoring features on its cars



All the fatal autopilot accidents and the Uber accident as well shows the vigilance issue. Yet there was no change in Tesla's approach.


.

 

[CarlK]

Anyway, Musk's statement saying "if simulations worked we could prove that we live in a simulation" is one of the stupidest things he said at the event. He's obviously a smart guy, but sometimes he's just full of it. Reminds me of his "alien dreadnought" that was supposed to teach everyone else how to build cars ...

He said that in jest. That's a favorite subject among people with too much excess brain power as physicists still can not figure out a way to prove that we are NOT simulation. The real meaningful comment he made just before that was if we can make correct simulation set up we have already solved the self driving problem. How could you simulate tail cases if you don't already know them?

At this level, I am not so sure. These guys are employable everywhere and I bet they have headhunters calling them weekly. The fact that they stay with Tesla for these years says a lot.

Yes smart people like to work with other smart people. Not so smart people, including bosses and coworkers, will play games to cover their inadequacy. Working for a boss that is not as smart as you is a torture. Smart people do not afraid to hire other smart people too, All these works favorably for Tesla.

Media make big deal every time when someone leaves the company. They don't care if the reason was that person is not good enough or can't stand the heat. In the end it's still who are working there matters. There are many signs that Tesla has a truly amazing team not just those few we know.

 

[Bladerskb]

He said that in jest. That's a favorite subject among people with too much excess brain power as physicists still can not figure out a way to prove that we are NOT simulation. The real meaningful comment he made just before that was if we can make correct simulation set up we have already solved the self driving problem. How could you simulate tail cases if you don't already know them?

You clearly don't understand how beneficial a good simulator is. You can take one instance you see in the Real world and create thousands of variations of it in a simulator. The problem is NOT simulating individual tail cases, the problem is simulating the way humans drive as a whole.

For example some of the pros of simulation:

• It's cheap -- no need to even have a car, let alone a road or safety drivers.
• It's fast -- you can drive a million miles in simulator in a night.
• It's safe -- you can test dangerous situations you could never try in the real world, like pedestrians running in front of you or being in the middle of accidents
• It's complex -- you can drive all your simulated miles in interesting and challenging conditions, not wasting time on simple and boring ordinary roads.
• It's repeatable -- you can do the same test again and again.
• It's variable -- you can try millions of tiny variations of any situation. You can try every type of weather and lighting. You can try every type of road and traffic.
• It's strange -- you can test extremely rare situations, like what your car might do in a tsunami, earthquake, fire, flood, tornado or dust storm.
• It's software -- you can create complex failures of your simulated hardware, mixed in with every situation, and know how you deal with them.

 

[Eno Deb]

He said that in jest. That's a favorite subject among people with too much excess brain power as physicists still can not figure out a way to prove that we are NOT simulation. The real meaningful comment he made just before that was if we can make correct simulation set up we have already solved the self driving problem.

That's BS too.

How could you simulate tail cases if you don't already know them?

For one, it's not only about "tail cases". So far Tesla is far from having mastered things like basic driving policy. And second, there are techniques like fuzzing and randomization that will absolutely produce unexpected outcomes, just like real world driving. It is also possible to feed fleet data into a simulation and multiply its usefulness.