Musk: All modes AP by end 2019

[Compton]

This is completely false. Mobileye EyeQ3 had dozens of NN models and that was released in 2014 (and is used for L2).
Mobileye EyeQ4 on the other hand has hundreds of NN models and it was released late 2017 (and is used for L3,L4, and L5 self driving cars).

Now ask yourselves this question.
How in the world is a company that is already pencils down, that already has their product in production, how are they late and behind a company that is scurrying around trying to get their NN to match L2 driver assistance competency??

You ppl and your tesla logic are amazing!

By the way, supercruise which uses eyeq3 (The same company who you say are late) is for sale and surpass AP1 and 2.

The confusion perhaps stems from the difference between deep learning neural networks and "normal" neural networks. Mobileye has historically used the latter method whereas Tesla is using the former approach. Deep learning networks can be trained both in an unsupervised or supervised manner. "Normal" neural networks are used for supervised prediction.

So saying that "ME is late to the NN game" might be true if the expression was modified slightly to "ME is late to the deep learning NN game" or "ME does not participate in the deep learning NN game". Whether or not deep learning is the key to affordable autonomous driving remains to be seen but Tesla's latest firmware update (released today) seems like an encouraging step in the right direction. The recent lack of progress was probably due to rebuilding of some basic groundwork of the deep learning system under the direction of Andrej Karpathy.

MobileEye's CEO was not a fan of deep learning in 2016 but I don't know his current stance; Mobileye Bullish on Full Automation, but Pooh-Poohs Deep-Learning AI for Robocars

[expecting a characteristically snarky reply from Bladerskb, possibly spiced up with some name-calling]

 

[Bladerskb]

The confusion perhaps stems from the difference between deep learning neural networks and "normal" neural networks. Mobileye has historically used the latter method whereas Tesla is using the former approach. Deep learning networks can be trained both in an unsupervised or supervised manner. "Normal" neural networks are used for supervised prediction.

So saying that "ME is late to the NN game" might be true if the expression was modified slightly to "ME is late to the deep learning NN game" or "ME does not participate in the deep learning NN game". Whether or not deep learning is the key to affordable autonomous driving remains to be seen but Tesla's latest firmware update (released today) seems like an encouraging step in the right direction. The recent lack of progress was probably due to rebuilding of some basic groundwork of the deep learning system under the direction of Andrej Karpathy.

MobileEye's CEO was not a fan of deep learning in 2016 but I don't know his current stance; Mobileye Bullish on Full Automation, but Pooh-Poohs Deep-Learning AI for Robocars

This article you cited has several wrong information on it. First of all the talk that the article is based on is here and Amon was talking against end to end driving which was being promoted by Nvidia by the demo they were doing.

First of all, I repeat NO self driving car company actually uses end to end NN. Not even Tesla!

start from 7:25

One erroneous claim by the article is that Google changed their approach based on AlphaGo, but this is complete fabrication by the author. Google still uses traditional programming approach for their planning and driving policy. Go watch any Waymo presentation to see the method they use.

What Ammon was against was end to end self driving which no one is doing. No self driving company uses an end to end NN to drive a car. Nvidia did a demonstration that was simply a demo.

The confusion perhaps stems from the difference between deep learning neural networks and "normal" neural networks. Mobileye has historically used the latter method whereas Tesla is using the former approach. Deep learning networks can be trained both in an unsupervised or supervised manner. "Normal" neural networks are used for supervised prediction.

You have the typical Tesla fan misconception when it comes to NN. I think all tesla fans should actually take time to research these things before jumping into the discussion.

The difference between artificial neural network and deep neural network is LAYERS not supervised vs unsupervised.
DNN has many hidden layers hence the word "deep". That's the difference between DNN vs NN.

It has nothing to do with Supervised or unsupervised learning

The difference between supervised versus unsupervised learning is that with supervised you deal with labeled data but with unsupervised you deal with unlabeled data.

So say you want to develop a classification network that tells the difference between a cat and a dog.

With supervised learning, you collect a bunch of pictures of cats and dogs and label them accurately and say "this is a cat" and "this is a dog", etc.
With unsupervised learning, you just collect random pictures and just feed it into the network without telling it what the data are.

(you can also use an existing pre-trained model to classify objects, ex: a model trained with ImageNet, VGG 16)

Whether or not deep learning is the key to affordable autonomous driving remains to be seen but Tesla's latest firmware update (released today) seems like an encouraging step in the right direction. The recent lack of progress was probably due to rebuilding of some basic groundwork of the deep learning system under the direction of Andrej Karpathy.

Again you are misinformed. Everyone does supervised deep neural network for their perception and object detection/classification.

Mobileye used 600+ people to labeled data for them.
Tesla outsources to an outside company to label data for them.

Everyone uses DNN. EVERYONE!

The fact that Tesla are still "rebuilding of some basic groundwork of the deep learning system" shows you how behind they are.

 

[ChrML]

One easy way to say this: There is no way a company can create a reliable end-to-end, self-driving car with today's narrow AI.

Down the road with general AI, it will be possible. You could put the ECU at a driver's school and tell it how to drive like a human. But general AI is still in early research and nobody has succeeded yet, so not likely a mainstream technology before many years.

So current solution: Lots of small chunk DNN's doing isolated tasks, connected together by conventional programming. Easy swappable modules in your code, easy to debug and the NN black boxes have well defined behaviors.

 

[Compton]

This article you cited has several wrong information on it. First of all the talk that the article is based on is here and Amon was talking against end to end driving which was being promoted by Nvidia by the demo they were doing.

First of all, I repeat NO self driving car company actually uses end to end NN. Not even Tesla!

start from 7:25

One erroneous claim by the article is that Google changed their approach based on AlphaGo, but this is complete fabrication by the author. Google still uses traditional programming approach for their planning and driving policy. Go watch any Waymo presentation to see the method they use.

What Ammon was against was end to end self driving which no one is doing. No self driving company uses an end to end NN to drive a car. Nvidia did a demonstration that was simply a demo.




You have the typical Tesla fan misconception when it comes to NN. I think all tesla fans should actually take time to research these things before jumping into the discussion.

The difference between artificial neural network and deep neural network is LAYERS not supervised vs unsupervised.
DNN has many hidden layers hence the word "deep". That's the difference between DNN vs NN.

It has nothing to do with Supervised or unsupervised learning

The difference between supervised versus unsupervised learning is that with supervised you deal with labeled data but with unsupervised you deal with unlabeled data.

So say you want to develop a classification network that tells the difference between a cat and a dog.

With supervised learning, you collect a bunch of pictures of cats and dogs and label them accurately and say "this is a cat" and "this is a dog", etc.
With unsupervised learning, you just collect the pictures and just feed it into the network without telling it what the data are.

(you can also use an existing pre-trained model to classify objects, ex: a model trained with ImageNet, VGG 16)

The reason people use supervised learning over unsupervised for object detection/classification is because of accuracy (which is not even comparable).



Again you are misinformed. Everyone does supervised deep neural network for their perception and object detection/classification.

Mobileye used 600+ people to labeled data for them.
Tesla outsources to an outside company to label data for them.

Everyone uses DNN. EVERYONE!

The fact that Tesla are still "rebuilding of some basic groundwork of the deep learning system" shows you how behind they are.

I stand corrected about my understanding of DNN and the name-calling. My neural network predicted only the snarkiness and off-putting writing style correctly.

My NN's next prediction is that you will continue to loathe Tesla and autopilot even if they manage to develop a well-functioning system.

 

[buttershrimp]

This is completely false. Mobileye EyeQ3 had dozens of NN models and that was released in 2014 (and is used for L2).
Mobileye EyeQ4 on the other hand has hundreds of NN models and it was released late 2017 (and is used for L3,L4, and L5 self driving cars).

Now ask yourselves this question.
How in the world is a company that is already pencils down, that already has their product in production, how are they late and behind a company that is scurrying around trying to get their NN to match L2 driver assistance competency??

You ppl and your tesla logic are amazing!

By the way, supercruise which uses eyeq3 (The same company who you say are late) is for sale and surpass AP1 and 2.

You’re right Blader.... I’m mid download of the next firmware update, but why even let it finish, I’m gonna cancel it and sell my car. I mean, Bob Lutz has created a masterpiece and I am incomplete until I have experienced the ecstasy of the supercruise Volt.

 

[croman]

This is completely false. Mobileye EyeQ3 had dozens of NN models and that was released in 2014 (and is used for L2).
Mobileye EyeQ4 on the other hand has hundreds of NN models and it was released late 2017 (and is used for L3,L4, and L5 self driving cars).

Embedded NN models. Yawn. Static and easily outpaced by something that can be updated OTA (like 2018.10.4). Karpathy will keep innovating. Tesla's AP2 hardware will continue to improve whereas any ME equipped car will need to be traded in before you can get EyeQ5 (which will actually do L5 and actually power a deep learning NN). Intel's manufacturing and chipset design allows them to attempt 7nm FinFet with EyeQ5. So maybe 2020 ME will come out with something that will actually do what Elon believes he can do. We'll see if in 2 years Elon can actually achieve his vision now that he has a competent Head of AI.

[/QUOTE] By the way, supercruise which uses eyeq3 (The same company who you say are late) is for sale and surpass AP1 and 2.[/QUOTE]

No, try again. I can use my AP anywhere and anytime. Can Supercruise do that? No, it must be lidar mapped and there must be no sun EVER AND I've got to have this ****ing machine staring me in the eye. No god damn toaster is ever looking me in the eye and constantly monitoring me.

 

[Xminus6]

Segway'ing off wipers & back to AP2 ~ has anybody else noticed the complete loss of lane change capability on freeways?? regardless how empty Freeway Lanes might be? We tried using the feature on the way home a couple times every minute, off & on for over 5 minutes and . . . . . . nothing. It's been spotty at best over the last year+ anyway. But it's not even spotty now. We just wanted to know whether it was even worth bringing up at the SC tomorrow when we go in for other nickel-&-dime stuff, again.
.

Yes. Happened on my X when it was working well before. Oddly enough, I just took my X in for service and was given an S75 and it also does not change lanes under AP either.

 

[Bladerskb]

I stand corrected about my understanding of DNN and the name-calling. My neural network predicted only the snarkiness and off-putting writing style correctly.

My NN's next prediction is that you will continue to loathe Tesla and autopilot even if they manage to develop a well-functioning system.

I don't think i used any term that could be construed as name-calling. Referring to your statement as misinformed is about as polite as i could be and calling you a Tesla fan would be like you calling me a warriors basketball fan, which would be accurate. No malicious attempt here, sorry.

You’re right Blader.... I’m mid download of the next firmware update, but why even let it finish, I’m gonna cancel it and sell my car.

I finally turned a red coat, my life is fulfilled!

 

[Bladerskb]

Embedded NN models. Yawn. Static and... EyeQ5 (which will actually do L5 and actually power a deep learning NN).

You my brother are very mistaken.

 

[croman]

I don't think i used any term that could be construed as name-calling. Referring to your statement as misinformed is about as polite as i could be and calling you a Tesla fan would be like you calling me a warriors basketball fan, which would be accurate. No malicious attempt here, sorry.

That poster isn't saying you name called, but rather that they were incorrect about the technical nomenclature they used in their post and that you were correct.

And I might be mistaken. At least I acknowledge that possibility and so I am able to improve (like AP2) whereas some are static and believe they are the end all be all and therefore fail to better themselves.

 

[Yuri_G]

Current estimate for FSD seems a bit more accurate with 2018.10.4. The videos I've seen shows Autopilot handling roads that both AP1 and AP2 (previous versions) would fail miserably at.

 

[Bladerskb]

And I might be mistaken. At least I acknowledge that possibility and so I am able to improve (like AP2) whereas some are static and believe they are the end all be all and therefore fail to better themselves.

All i'm trying to get across is that the idea that Tesla is the only company running deep neural networks is mistaken. Elon is the only one who blabs constantly about it though. Which leads to his fans making erroneous statements.

In referring to your comments, for example: deep learning neural network are simply a buzz word for neural network with many hidden layers. The most popular network today which is CNN was developed in the 90s.

So its wrong to say that eyeq4 is running nn models and eye5 is running deep learning neural networks because they are the exact same thing. Also the word model in this context is interchangeable with network. You can either call it a network or a model which is short for neural network model.

In short to reply to your statement, yes eyeq4 runs deep learning neural networks.
And no you won't need eyeq5 for Level 5, that's just marketing by mobileye. The only difference between eyeq5 and eyeq4 is that 5 has a boat load of TFLOPS for OEM to run any algorithm and even their own networks with.

All you need depending on your sensor size (some oem have like 10+ radars, 10+ cameras, 5+ lidar configuration) is two eyeq4.
And then a second intel/nvidia chip to run all your custom algorithm and driving policy if you wanted. Its similar to what Audi did with the L3 A8, they use an eyeq3 for all the vision stuff and then they use another chip to run their own custom algorithms for control actuation.

While you can run driving policy and your own custom algorithm in eyeq4 and that's what its made for. Some companies like waymo use intel chips with 50 TFLOP processing power.

Its kinda interesting how people would watch the above video and would rather take a project in its infancy rather than a product that is already finished. We know that Tesla is in the early stages of building their networks. Early early stages.

But because Elon has convinced alot of people that their cars are actually getting better even though when you look at the overall picture. You find out that you purchased a car with half its standard features missing which is being installed 2-3 years later through OTA. That's not a benefit that's a negative.

Saying that AP2 will keep improving and eyeq4 is static is not a positive to Tesla.
Its a huge negative. Its like having two people take a test and one finished first and got a 99.99% and the other is still taking the test and is just 20% through it. A bystander was asked to place bets on who would score better and he says the person whose not finished yet because he is able to improve but the guy whose finished will always be stuck at 99.99%.

 

[Yuri_G]

All i'm trying to get across is that the idea that Tesla is the only company running deep neural networks is mistaken. Elon is the only one who blabs constantly about it though. Which leads to his fans making erroneous statements.

In referring to your comments, for example: deep learning neural network are simply a buzz word for neural network with many hidden layers. The most popular network today which is CNN was developed in the 90s.

So its wrong to say that eyeq4 is running nn models and eye5 is running deep learning neural networks because they are the exact same thing. Also the word model in this context is interchangeable with network. You can either call it a network or a model which is short for neural network model.

In short to reply to your statement, yes eyeq4 runs deep learning neural networks.
And no you won't need eyeq5 for Level 5, that's just marketing by mobileye. The only difference between eyeq5 and eyeq4 is that 5 has a boat load of TFLOPS for OEM to run any algorithm and even their own networks with.

All you need depending on your sensor size (some oem have like 10+ radars, 10+ cameras, 5+ lidar configuration) is two eyeq4.
And then a second intel/nvidia chip to run all your custom algorithm and driving policy if you wanted. Its similar to what Audi did with the L3 A8, they use an eyeq3 for all the vision stuff and then they use another chip to run their own custom algorithms for control actuation.

While you can run driving policy and your own custom algorithm in eyeq4 and that's what its made for. Some companies like waymo use intel chips with 50 TFLOP processing power.

Its kinda interesting how people would watch the above video and would rather take a project in its infancy rather than a product that is already finished. We know that Tesla is in the early stages of building their networks. Early early stages.

But because Elon has convinced alot of people that their cars are actually getting better even though when you look at the overall picture. You find out that you purchased a car with half its standard features missing which is being installed 2-3 years later through OTA. That's not a benefit that's a negative.

Saying that AP2 will keep improving and eyeq4 is static is not a positive to Tesla.
Its a huge negative. Its like having two people take a test and one finished first and got a 99.99% and the other is still taking the test and is 20% through it and a guy was asked who he would place bet on to score better and he says the guy whose not finished because he is able to improve but the guy whose finished will always be stuck at 99.99%.

All that matters is what can be purchased and driven right now. As far as I can tell Supercruise is the only software that comes close and can't be driven on two lane roads and a select number of freeways. Tesla's new update just made your argument moot.

 

[Bladerskb]

All that matters is what can be purchased and driven right now. As far as I can tell Supercruise is the only software that comes close and can't be driven on two lane roads and a select number of freeways. Tesla's new update just made your argument moot.

Supercruise has been acclaimed to be the best L2 driver assistance in production by basically every journalist who did an in-depth review against AP. secondly Mobileye's eyeq4 is already in production since late 2017 and will show up in several car models this year and next.

Lastly all that matters ISN'T what can be driven today.

 

[cbardeen]

Why is it so hard for you to understand people wanting/hoping for Tesla to deliver on AP? No company provides a complete solution right now, but people interested in Tesla and AP have bought into Tesla’s vision for both electric cars and for self driving. If you are happy with Supercruise, great buy it. If you prefer a ME’s product plan, great buy a car from one of their partners when available. If you think Tesla is vaporware, then don’t buy one. There are obviously lots of paths to self driving and hopefully they will all succeed. I appreciate your comments on technology, but not the tone of your comments. Why does progress by one group have to be a put down directed toward another? Tesla has been pretty opaque about their AP activities, so people are grasping at whatever they can get to try to identify progress. I don’t really care what MobileEye does except as evidence of advancement of the technology, technology that won’t be unique to one manufacturer and that I want to eventually make it into the Tesla cars.

Did Tesla fall behind after the ME split? Sure. Are they taking longer than many expected to recover? Yes. Was this another example of Elon over promising? Sure. Will they ever develop a system that meets the claims that Elon has made? Who knows. However, the bottom line for me is that of all the car makers out there Tesla is the most willing and committed to pushing the envelope in terms of electric cars, supercharger networks, elegant design, sporty performance, lower costs, and using technology including self driving. I don’t know if they will succeed, but I have faith that they will try harder than most and are currently more committed to that path than anyone else. I do feel like that they now have the AI expertise in house to give themselves a chance to succeed with AP. They won’t be settling for compliance cars, econo boxes, high priced self-driving taxis, limited AP capabilities, ... Will someone come along come along with a better concept or a better product? Probably, but to date no one has matched the progress that Tesla has made for the combined vision of which AP is just one piece.

P.S. Personally, I wouldn’t expect or trust FSD from anyone anytime soon, but I want to push in that direction and am willing to provide driver oversight to have more capabilities in more situations. Supercruise seems to be taking the other approach of only being willing to to provide self driving features where they think they can do it without driver oversight. That may be preferable to some, but isn’t as interesting to me. Happy to have both options out there and let people pick.

 

[bnsfengineer]

From 3 months maybe....6 months definitely!! (TWICE NOW) to 18. Months from now?

Bring out the 20 gallon coolers of KoolAide please. We all need a cup or two. Pass it around.

So Why Why did you pre-pay for the Kool-Aid

 

[nowtleft]

Here we go again.

Tesla CEO Elon Musk: ‘self-driving will encompass all modes of driving by the end of next year’

It's like that scene in The Lego Movie where Batman tries repeatedly to hit a button....

Tesla can't even get me 4 wheels after 6 months so no breath held here for anything more technical!

 

[boonedocks]

So Why Why did you pre-pay for the Kool-Aid

Because I actually believed the rhetoric in the beginning. I always take people for THEIR word until THEY prove themselves wrong. I don't expect anything from Tesla EXCEPT what they SAY!! That includes public speeches and Tweets from the CEO. If that isn't the way Tesla wants plans and features and products announced then they, Tesla, should reel that in. ¯\_(ツ)_/¯

 

[bnsfengineer]

Because I actually believed the rhetoric in the beginning. I always take people for THEIR word until THEY prove themselves wrong. I don't expect anything from Tesla EXCEPT what they SAY!! That includes public speeches and Tweets from the CEO. If that isn't the way Tesla wants plans and features and products announced then they, Tesla, should reel that in. ¯\_(ツ)_/¯

I'm right there with you! Only difference is I poor otherwise I'd have a large cup too

 

[gluu]

All i'm trying to get across is that the idea that Tesla is the only company running deep neural networks is mistaken. Elon is the only one who blabs constantly about it though. Which leads to his fans making erroneous statements.

In referring to your comments, for example: deep learning neural network are simply a buzz word for neural network with many hidden layers. The most popular network today which is CNN was developed in the 90s.

So its wrong to say that eyeq4 is running nn models and eye5 is running deep learning neural networks because they are the exact same thing. Also the word model in this context is interchangeable with network. You can either call it a network or a model which is short for neural network model.

In short to reply to your statement, yes eyeq4 runs deep learning neural networks.
And no you won't need eyeq5 for Level 5, that's just marketing by mobileye. The only difference between eyeq5 and eyeq4 is that 5 has a boat load of TFLOPS for OEM to run any algorithm and even their own networks with.

All you need depending on your sensor size (some oem have like 10+ radars, 10+ cameras, 5+ lidar configuration) is two eyeq4.
And then a second intel/nvidia chip to run all your custom algorithm and driving policy if you wanted. Its similar to what Audi did with the L3 A8, they use an eyeq3 for all the vision stuff and then they use another chip to run their own custom algorithms for control actuation.

While you can run driving policy and your own custom algorithm in eyeq4 and that's what its made for. Some companies like waymo use intel chips with 50 TFLOP processing power.

Its kinda interesting how people would watch the above video and would rather take a project in its infancy rather than a product that is already finished. We know that Tesla is in the early stages of building their networks. Early early stages.

But because Elon has convinced alot of people that their cars are actually getting better even though when you look at the overall picture. You find out that you purchased a car with half its standard features missing which is being installed 2-3 years later through OTA. That's not a benefit that's a negative.

Saying that AP2 will keep improving and eyeq4 is static is not a positive to Tesla.
Its a huge negative. Its like having two people take a test and one finished first and got a 99.99% and the other is still taking the test and is just 20% through it. A bystander was asked to place bets on who would score better and he says the person whose not finished yet because he is able to improve but the guy whose finished will always be stuck at 99.99%.

Mobileye will keep improving their system as well as Tesla, so if we want to predict the winner of this race, we need to consider their current positions and their current velocity and acceleration. Currently, because Tesla is vertically integrated, have cars with OTA updates and hardware for FSD already out in production, they are able to deploy any breakthroughs in their software to an entire fleet in the span of a week and get production validation of the performance of their work.

Mobileye's ultimate flaw despite their current positional lead is that they can't get to production and do engineering validation quickly, their hands are tied because of how slow and cautious the OEMs are. Think about their development cycle, let's say Mobileye has FSD right now, only they do not know it yet because they haven't done production validation yet. How long would it take for them to get to market? The OEMs are a huge obstacle for them, and only Tesla was willing to push so aggressively in the past.