Autonomy Investor Day - April 22 at 2pm ET

[mongo]

It's fast -- you can drive a million miles in simulator in a night.

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.

So your effective test rate is sub one mile a night?

'Every' where every is limited by your imagination, literally...

 

[CarlK]

That's BS too.

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.

C'mon. Whoever could have a computer to simulate all possible human behaviors he would have already owned the entire world.

 

[Eno Deb]

C'mon. Whoever could have a computer to simulate all possible human behaviors he would have already owned the entire world.

Yeah, that's exactly the kind of pseudointellectual BS that Elon uses when he has no solid results to show. Very common in Silicon Valley.

 

[CarlK]

Yeah, that's exactly the kind of pseudointellectual BS that Elon uses when he has no solid results to show. Very common in Silicon Valley.

Not interested in the pissing game that no one gains anything. Was trying to argue from technical point but apparently it will never work for you. You are a goner from my list.

Not so smart people, including bosses and coworkers, will play games to cover their inadequacy.

That includes forum posters too. You go enjoy yourself with people like you.

 

[electronblue]

I find it truly fascinating that @CarlK has started ignore-listing half this forum — and especially that he chooses to announce it in public.

Internet never disappoints.

 

[timewasted]

I find it truly fascinating that @CarlK has started ignore-listing half this forum — and especially that he chooses to announce it in public.

Internet never disappoints.

A challenger has appeared, threatening to take the mantle of "most willing to live in an echo chamber" away from @strangecosmos !

 

[timewasted]

You know what bugs me most about the people who hang onto Elon's every word as if it's the gospel truth? Lets say for the sake of argument that this "Autonomy Investor Day" wiped the slate clean of all of Elon's past broken promises (it didn't, but lets pretend it did).

Go take a look at the autopilot page for the Model 3 order process. The page currently says, in part, "Summon: your parked car will come find you anywhere in a parking lot. Really." This is listed as a current feature that you get today when you spend $5000. It has said this for the past two months (since February 28th of this year). This feature still does not exist. Yes, if you're one of the ~200 or so people in the beta group you get something like this, but it's not this because:

A) It is advertised as being available today without any qualifiers, when it does have qualifiers that restrict access to a very small fraction of the fleet.
B) Even the beta group does not have access to this feature, because it's hardly "anywhere in the parking lot" with the laundry list of restrictions on it (you must be within X feet, maintain constant visual supervision, and be aware that it may fail for any of Y reasons).

This is not some feature that is listed as "pending validation", or "waiting on regulatory approval", or "coming later this year". They are actively selling something that does not exist, and they're doing it today and not back in 2016. Why the hell does this not piss off more people?!

I'm really not asking anyone to abandon whatever faith they might have in Elon/Tesla. All I'm asking for is for people to look at the facts that are in plain sight and not just dismiss them because Elon told you that he means it for realsies this time.

 

[CarlK]

Your issues are liget but be patient if you still have faith in this. I don't have reason not to and willing to take a risk. People who don't trust this guy could just leave or not to become a customer. It's just that simple. He's trying to create a better world for us to enjoy and has already succeeded in many ways. With him or not but all those bashing seems to be pretty self serving whatever your agenda it.

I put a few on ignore when they are trying to turn a technical discussion into a pissing game when they don't have anything meaningful to argue. Few, if any, of us here are intelligent enough to know if Elon's right or wrong but on the AI front even people like Steven Hawkin and Neil deGrasse Tyson agree with almost all things he said. Many knowledgeable people on the self driving front are starting to go to his side where he was all alone before also. So whoever can only use "He can only BS" to create new physics is either super super smart or super dumb. I do have a Ph.D in physics from a top university and have published numerous peer reviewed papers, not in the AI area though. At least I likely am not super dumb. I have no interest in having this kind of "discussion" with a super dumb person. And I won't get any satisfaction out that this kind of debate win or lose. Anyone who thinks he can tell Elon you're only BS please let everyone know what are your life's achievements to prove BS'ing is not the only thing you can do.

Sorry to see what could be a good discussong board turned into a political fighting ground. Everyone loses.

 

[electronblue]

People who don't trust this guy could just leave or not to become a customer. It's just that simple.

It isn’t quite that simple. The cars purchased have been major investments into promised future features. Leaving would not just mean ”not trusting Elon”, it would mean eating depreciation on the car (ever increasing in the Tesla world of price cuts) and missing out on any chance on the promised features.

Personally I’ve simply decided to cut my losses regarding future Tesla purchases and am staying put with the current current car. You may dislike it but I have legitimate grievances and am not going anywhere.

 

[ZeApelido]

Holy crap, you guys really underestimate the data advantage Tesla has. Amnon (errr Blader) says any company can get enough test data. I say no way. The benefit of Tesla's crowdsourced fleet is the ability to collect data all over the world in different conditions. You know, not just in Mountain View, over, and over again...

Simulations are nice, they are good for augmenting your data set. But as Karpathy alluded to, if your simulation was so great it could generate all the possible conditions you will need, it means you already have a good enough model for robust driving. The simulation models, while they seem really cool, still are limited to the data use to build them. They create new scenes that are essentially interpolations.

The tail-end of test and error distributions will be filled with odd cases that are rarely encountered. To assume that a limited set of those cases + simulator is sufficient is mind-numbingly dumb and I can't believe we're having this discussion.

That's not to say Tesla's test data advantage is all that matters. But to downplay it is a joke.

 

[brianman]

if your simulation was so great it could generate all the possible conditions you will need, it means you already have a good enough model for robust driving.

This is incorrect. Describing a problem fully does not mean that you have a solution in hand. Describing a problem is easier than solving it.

 

[Eno Deb]

Simulations are nice, they are good for augmenting your data set. But as Karpathy alluded to, if your simulation was so great it could generate all the possible conditions you will need, it means you already have a good enough model for robust driving.

It was Elon who said that, and it's a fallacy. For a simulation to be useful the agents in the simulation don't have to have a self-driving stack with computer vision, perception, elaborate driving policy etc.

What some people don't get is that this was primarily a sales event aiming to convince investors and the markets that Tesla is competitive in autonomous driving after numerous missed targets. That's why he (literally) trash talked things like Lidar and simulations where others currently lead ("they are doomed, doomed I tell you" ).

The tail-end of test and error distributions will be filled with odd cases that are rarely encountered.

The thing is that capturing unexpected edge cases from a fleet doesn't just happen either. Tesla can't and doesn't collect all telemetry form all cars at all times. They install triggers in a subset of the fleet that allow them to capture specific types of events. But this also means that you have to already know what you are looking for in most cases, which means you can probably also integrate it in a simulation in many variants.

 

[CarlK]

He meant if you could bulid one rare case in the simulation you already know about that rare case (and consequently solves that problem). The problem with simulation is you don't know all those rare cases until you find out about them in the real world.

 

[ZeApelido]

This is incorrect. Describing a problem fully does not mean that you have a solution in hand. Describing a problem is easier than solving it.

That's fine if you believe that, I'm not too strong on that position. But I do believe that the simulations will not describe the problem fully.

 

[ZeApelido]

It was Elon who said that, and it's a fallacy. For a simulation to be useful the agents in the simulation don't have to have a self-driving stack with computer vision, perception, elaborate driving policy etc.

What some people don't get is that this was primarily a sales event aiming to convince investors and the markets that Tesla is competitive in autonomous driving after numerous missed targets. That's why he (literally) trash talked things like Lidar and simulations where others currently lead ("they are doomed, doomed I tell you" ).
The thing is that capturing unexpected edge cases from a fleet doesn't just happen either. Tesla can't and doesn't collect all telemetry form all cars at all times. They install triggers in a subset of the fleet that allow them to capture specific types of events. But this also means that you have to already know what you are looking for in most cases, which means you can probably also integrate it in a simulation in many variants.

I didn't say the simulations aren't useful. I said they can augment the real data set. But the simulations alone will not be enough. However, in theory enough raw data could be enough on its own (but may take a while to get).

Of course Tesla doesn't collect all the data. Why on earth would they?

Now you are downplaying their ability to collect edge cases. As I've said before, that's not too difficult when you have an L2 system in place to help you look at where disengagements occur. Even though the end result requires a manual step, having someone go through a series of 6 second video downloaded from automated triggers, and identifying the important ones is not that laborious.

Waymo data than Waymo.

 

[Eno Deb]

I didn't say the simulations aren't useful.

Well, you paraphrased Elon's claim that if you had a great simulator, you'd already have solved the self-driving problem. I think that's complete nonsense.

I said they can augment the real data set. But the simulations alone will not be enough. However, in theory enough raw data could be enough on its own (but may take a while to get).

Every single company active in this field uses both.

Waymo data than Waymo.

And yet, Tesla has not demonstrated that they can replicate what Waymo does (or Cruise, for that matter). Now we have Elon hyping up things like traffic light recognition as the next big step toward "feature completeness", even though others have been doing it for years. I think there are good reasons to be sceptical about their claims. Mind you, as a Model 3 owner I hope I'll be proven wrong.

 

[CarlK]

I didn't say the simulations aren't useful. I said they can augment the real data set. But the simulations alone will not be enough. However, in theory enough raw data could be enough on its own (but may take a while to get).

Of course Tesla doesn't collect all the data. Why on earth would they?

Now you are downplaying their ability to collect edge cases. As I've said before, that's not too difficult when you have an L2 system in place to help you look at where disengagements occur. Even though the end result requires a manual step, having someone go through a series of 6 second video downloaded from automated triggers, and identifying the important ones is not that laborious.

Waymo data than Waymo.

Lidar and simulation could get you to 99% relatively quickly, That's why those dozen or so start ups love it. You can set up a nice demo to impress investors or auto companies to invest or to buy you out. Then those edge cases you see only once is a long while will stop you. Waymo's current thinking is less clear but Waymo does need Waymo data it still does not have a mean to get.

 

[Daniel in SD]

Lidar and simulation could get you to 99% relatively quickly, That's why those dozen or so start ups love it. You can set up a nice demo to impress investors or auto companies to invest or to buy you out. Then those edge cases you see only once is a long while will stop you. Waymo's current thinking is less clear but Waymo does need Waymo data it still does not have a mean to get.

Tesla probably has more images of highway overpasses and the sides of semi trucks than anyone else and yet...
I think it remains to be seen how much a huge data set will help. I do agree that it's their biggest advantage.

 

[Eno Deb]

Tesla probably has more images of highway overpasses and the sides of semi trucks than anyone else and yet...

I'm not so sure if that's even true as far as image data is concerned. Don't forget that Google has an absolutely ginormous base of image data from various sources (including Streetview cars that have mapped and imaged pretty much every corner of the country). There are also dozens of companies that can provide labeled training datasets of any kind (some of which Tesla is probably using to train their vision system). The value of Tesla's fleet as a datasource, if it is indeed as big as they claim, would probably be more in targeted extraction in connection with specific driving events. In addition they are certainly using non-image data such as disengagements, but it's unclear how useful they actually are (given that drivers disengage for a variety of reasons) and how much effort it takes to establish context.

 

[CarlK]

That's AP1. Fingers were pointed at each other but image did come from Mobileye chip. Regardless EAP/FSD is a totally different technology.

Big data is essential for any deep machine learning not just self driving car. One thing we see AI do everyday is character recognition. It does a very good job now but still not perfect. That's why character verification can still be used to verify you're not a robot. However you probably have noticed those verification characters are harder and harder to recognize every day. That's what happens when you feed more and more data to the neural net. The same will happen to FSD.