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

[mongo]

There really isn’t any proof of that either beyond some limited trigger collection.

However, they are crowd sourcing the validation...

 

[electronblue]

However, they are crowd sourcing the validation...

I fully agree the AP2+ fleet gives Tesla a validation and deployment advantage.

 

[CarlK]

There really isn’t any proof of that either beyond some limited trigger collection.

Oh yeah there are many. This is what Chris Lattner said after he left Tesla. He's the one who's most familiar with the project at the time. He had no reason to cover for Tesla either.
"One of Tesla's huge advantages in the autonomous driving space is that it has thousands of cars already on the road. We built infrastructure to take advantage of this, allowing the collection of image and video data from this fleet, as well as building big data infrastructure in the cloud to process and use it."

I don't know why you want to make fleet machine learning sounds like such an exotic thing. Because Tesla is doing better than anyone else on that?

 

[Bladerskb]

Oh yeah there are many. This is what Chris Lattner said after he left Tesla. He's the one who's most familiar with the project at the time. He had no reason to cover for Tesla either.
"One of Tesla's huge advantages in the autonomous driving space is that it has thousands of cars already on the road. We built infrastructure to take advantage of this, allowing the collection of image and video data from this fleet, as well as building big data infrastructure in the cloud to process and use it."

I don't know why you want to make fleet machine learning sounds like such an exotic thing. Because Tesla is doing better than anyone else on that?

We already know the type of data they collected and how much. About 0.01% of data is collected. Far cry from the "tesla has billions of miles of data" mantra.

 

[CarlK]

We already know the type of data they collected and how much. About 0.01% of data is collected. Far cry from the "tesla has billions of miles of data" mantra.

"We"?

 

[iTech]

"We"?

I'm guessing you've never seen any posts from @verygreen?

 

[stretchsje]

We already know the type of data they collected and how much. About 0.01% of data is collected. Far cry from the "tesla has billions of miles of data" mantra.

They only want the edge cases. The less they collect despite large amounts of miles being driven shows they are filtering out the already trained cases successfully. It's a very good thing.

 

[CarlK]

I'm guessing you've never seen any posts from @verygreen?

No although I'm sure that was taken out of context by Bladerskb.

They only want the edge cases. The less they collect despite large amounts of miles being driven shows they are filtering out the already trained cases successfully. It's a very good thing.

Makes sense.

 

[electronblue]

Oh yeah there are many. This is what Chris Lattner said after he left Tesla. He's the one who's most familiar with the project at the time. He had no reason to cover for Tesla either.
"One of Tesla's huge advantages in the autonomous driving space is that it has thousands of cars already on the road. We built infrastructure to take advantage of this, allowing the collection of image and video data from this fleet, as well as building big data infrastructure in the cloud to process and use it."

I don't know why you want to make fleet machine learning sounds like such an exotic thing. Because Tesla is doing better than anyone else on that?

I don’tt think it is exotic. I just disagree with you on the interpretation of what Chris Lattner is saying.

Tesla collects limited trigger data from its consumer fleet. This is useful for validation, but there is no NN training happening inside the consumer cars nor is there any proof the trigger data is used to train NNs to any significant degree anyway. They train their NNs just like everyone else on test vehicles and simulators. The fleet is not learning, that is the part I disagree with.

Tesla does have a validation and deployment advantage with their OTA consumer fleet. That is not insignificant but it is not the same thing as the consumer fleet out there learning.

Mind you I’m not saying this can’t change in the future either but this is what we know of the current situation.

 

[CarlK]

I don’tt think it is exotic. I just disagree with you on the interpretation of what Chris Lattner is saying.

Tesla collects limited trigger data from its consumer fleet. This is useful for validation, but there is no NN training happening inside the consumer cars nor is there any proof the trigger data is used to train NNs to any significant degree anyway. They train their NNs just like everyone else on test vehicles and simulators. The fleet is not learning, that is the part I disagree with.

Tesla does have a validation and deployment advantage with their OTA consumer fleet. That is not insignificant but it is not the same thing as the consumer fleet out there learning.

Mind you I’m not saying this can’t change in the future either but this is what we know of the current situation.

Very creative argument but very obviously wrong too. Even without the fact that it's against everything Tesla people said it defies common logic too. Where are Tesla's test cars? Have you ever heard them mentioned anywhere? Tesla will never be able to deploy FSD everywhere in the world if it relies on it's test vehicles "just like everyone else". That certainly fits your thesis that Tesla is behind everyone else but then again the earth is flat too, No?

 

[dgatwood]

From earlier information, the SoC is a Samsung Exynos device (the specifics of which are unknown), and the NN accelerators are PCI-express devices connected to it. Now EM has said there are two NN accelerators per SoC. Two SoC for redundancy, two NN PCI-e devices per SoC, or four total.

I wonder why they are using two NN accelerators per cluster rather than designing the NN accelerator from inception with twice the compute and memory bandwidth. I have some ideas but none of them are compelling.

Well, for one thing, there's a limit to how superscalar you can get before you start losing performance because of bus contention on the shared memory. The two TPUs (can we call them that, or is that term just reserved for Google's hardware?) probably each have their own internal RAM (TRAM?), which would enable higher parallelism. Just a gut feeling, though.

PCI express is a shared bus

I'm not sure what you meant by that. (Parallel) PCI was a shared bus. PCIe is a dedicated point-to-point connection.

No matter how you look at it, its still defined as two chips (SoC), with each having two dedicated NN hardware accelerators (microprocessors).

To be slightly pedantic, I'm not sure I'd call a tensor processing unit a microprocessor. They may not even have Turing complete instruction sets, for all we know. Massively parallel vector engines aren't necessarily capable of running code on their own. I mean, if you squint hard enough, I guess you could call almost anything a microprocessor these days, but it's right at the fringes of what would qualify, either way.

But if you're saying you think the hardware accelerators are inside the same chip as the SoC, I highly doubt it. It costs a decent chunk of change to roll a custom IC, and I wouldn't expect Samsung to do that for a company that probably only wants a few thousand units per week after the initial flood of replacements when they could send Tesla to TSMC or somebody to build their accelerator, then have them slap it onto the PCI bus, and not have to do anything custom for Tesla at all.

Besides, the whole point of using an off-the-shelf SoC is that it's a known entity, so you don't have to debug that part. A standard SoC is a box of snakes nailed shut; a custom SoC is a bag of snakes, wide open....

 

[electronblue]

Very creative argument but very obviously wrong too. Even without the fact that it's against everything Tesla people said it defies common logic too. Where are Tesla's test cars? Have you ever heard them mentioned anywhere? Tesla will never be able to deploy FSD everywhere in the world if it relies on it's test vehicles "just like everyone else". That certainly fits your thesis that Tesla is behind everyone else but then again the earth is flat too, No?

I belive @verygreen on this one, that is pretty much all there is to it.

 

[CarlK]

I belive @verygreen on this one, that is pretty much all there is to it.

Why you and blader who were so adamant of what you think will always refer to @Evergreen and without even a quote or link? Why you are even here to start the argument?

We don't even need to refer to anyone or anything. If by your thesis that Tesla is not ahead of others because it does not know to use its consumer fleet data to train its NN then either Tesla is super stupid or you are.

 

[electronblue]

Why you and blader who were so adamant of what you think will always refer to @Evergreen and without even a quote or link? Why you are even here to start the argument?

We don't even need to refer to anyone or anything. If by your thesis that Tesla is not ahead of others because it does not know to use its consumer fleet data to train its NN then either Tesla is super stupid or you are.

Tesla is ahead of others in consumer fleet validation and deployment. They just don’t use it that fleet to train their NNs to any significant degree — not yet anyway. Maybe one day they will? We’ll know then.

As for @verygreen ’s credentials and what Teslas currently send back to mothership I’m afraid that it is your ignorance on the topic that is more of an issue than credibility of the data we have. This is well known stuff.

 

[OPRCE]

Why you and blader who were so adamant of what you think will always refer to @Evergreen and without even a quote or link? Why you are even here to start the argument?

We don't even need to refer to anyone or anything. If by your thesis that Tesla is not ahead of others because it does not know to use its consumer fleet data to train its NN then either Tesla is super stupid or you are.

I belive @verygreen on this one, that is pretty much all there is to it.

@verygreen may believe very little to no data is ever transmitted back for learning, because he has not seen it happen on his vehicle(s), but there could be a myriad of reasons why he sees so little, e.g. Tesla has identified his VIN(s) and quietly cut him out of the loop, or Tesla targets vehicles in localised areas for periodic data-gathering, or it depends or specific usage patterns/mileage threshold, etc., etc., so, while I appreciate the facts that verygreen reports, without complete insight into Tesla's methods we just cannot ATM conclusively know the reason why he gets the results he does.

OTOH ex-insider Chris Lattner reports Tesla has built a significant fleet-learning infrastructure/system, so it is hard to believe that he just fabricated this story or that the system described is not actually being put to its intended use.

 

[electronblue]

OTOH ex-insider Chris Lattner reports Tesla has built a significant fleet-learning infrastructure/system, so it is hard to believe that he just fabricated this story or that the system described is not actually being put to its intended use.

I just don’t think Lattner’s comments mean what some people think they mean.

Look. This is my opinion based on what we know, the big tidbits and the smaller ones. I think the pieces make sense that say so far Tesla has been training locally and mainly validating (and deploying) globally. This, I believe, is an accurate overall description what so far has made Autopilot 2+ tick. I think it is misleading to say NNs are trained on the global fleet or that the global fleet is doing the learning. No, learning — at least in any major way — happens in California and is uploaded to the fleet. The triggers we know of many be used in some minor ways but mostly I’d call this validation and elevating it to major role in the process is likely misleading.

This may change in the future. Maybe HW3 starts doing something significantly different. Also others may also believe otherwise, fair enough. This is my view.

 

[Az_Rael]

Where are Tesla's test cars? Have you ever heard them mentioned anywhere?

Tesla’s test cars have been spotted many times. You occasionally see people post them on Reddit. They are often Tesla’s running around with lots of sensors taped to the outside of them and sometimes a big red buttton on the inside with manufacturer plates. People have also spotted the engineers driving around with a laptop hooked into the car in the passenger seat (again with manufacturer plates). They are probably just harder to spot in Fremont now since every 3rd car is a Tesla up there it seems like.

Here is an older example of what was probably an AP2 test car from before AP2 released: Tesla with big red stop-button spotted

 

[mongo]

Tesla’s test cars have been spotted many times. You occasionally see people post them on Reddit. They are often Tesla’s running around with lots of sensors taped to the outside of them and sometimes a big red buttton on the inside with manufacturer plates. People have also spotted the engineers driving around with a laptop hooked into the car in the passenger seat (again with manufacturer plates). They are probably just harder to spot in Fremont now since every 3rd car is a Tesla up there it seems like.

Here is an older example of what was probably an AP2 test car from before AP2 released: Tesla with big red stop-button spotted

The big red button versions are usually part of new vehicle platform testing. FSD testing vehicles should not require such things (brake guaranteed to stop car/ disengage AP).

 

[CarlK]

@verygreen may believe very little to no data is ever transmitted back for learning, because he has not seen it happen on his vehicle(s), but there could be a myriad of reasons why he sees so little, e.g. Tesla has identified his VIN(s) and quietly cut him out of the loop, or Tesla targets vehicles in localised areas for periodic data-gathering, or it depends or specific usage patterns/mileage threshold, etc., etc., so, while I appreciate the facts that verygreen reports, without complete insight into Tesla's methods we just cannot ATM conclusively know the reason why he gets the results he does.

OTOH ex-insider Chris Lattner reports Tesla has built a significant fleet-learning infrastructure/system, so it is hard to believe that he just fabricated this story or that the system described is not actually being put to its intended use.

Good point on that. Not to mention someone hacked into the car still wouldn't know what even those what he sees little data transimmitted to the mothership are used. I don't believe Chris Lattner or Elon who has most intimate knowledge of it has any reason to lie. The most convincing argument is why is Tesla so dumb not to use those data only available to them but others are dying for? My speculation from the begining was Tesla started the program with camera in every car for the simple reason of this data collection. A lot of people argue at that time it should use Lidar (that can only be put in limited number of test cars) but it looks that strategy is starting to pay off now.

Tesla’s test cars have been spotted many times. You occasionally see people post them on Reddit. They are often Tesla’s running around with lots of sensors taped to the outside of them and sometimes a big red buttton on the inside with manufacturer plates. People have also spotted the engineers driving around with a laptop hooked into the car in the passenger seat (again with manufacturer plates). They are probably just harder to spot in Fremont now since every 3rd car is a Tesla up there it seems like.

Here is an older example of what was probably an AP2 test car from before AP2 released: Tesla with big red stop-button spotted

Tesla certainly can have test cars for any purposes but as @mongo pointed out those likely are not FSD cars which can easily be disengaged without needing a big red button. One pretty clear evidence is Tesla reported zero autonomous test miles to Californiz DMV for the entire year last year. So it is definitely not doing things "like everyone else".

I just don’t think Lattner’s comments mean what some people think they mean.

Look. This is my opinion based on what we know, the big tidbits and the smaller ones. I think the pieces make sense that say so far Tesla has been training locally and mainly validating (and deploying) globally. This, I believe, is an accurate overall description what so far has made Autopilot 2+ tick. I think it is misleading to say NNs are trained on the global fleet or that the global fleet is doing the learning. No, learning — at least in any major way — happens in California and is uploaded to the fleet. The triggers we know of many be used in some minor ways but mostly I’d call this validation and elevating it to major role in the process is likely misleading.

Sounds like a wishful thinking from someone who has already made up his mind that Tesla will not succeed. But again Tesla can not be this dumb to not do the obvious thing. Your wish is not going to be true.

 

[electronblue]

One pretty clear evidence is Tesla reported zero autonomous test miles to Californiz DMV for the entire year last year. So it is definitely not doing things "like everyone else".

Those cars from others, that are reporting autonomous miles for, are validation cars — they are not training neural networks in those cars either. Sure they collect some data but really the training of the networks is done in a very different way and different type of driving is used for that (when actual driving is used). Of course one difference exists: with their own test fleet manufacturers have access to both the driving scenario and full sensor data so they could use that for training more readily.

I’m sorry to say @CarlK but you seem to lack the basic understanding of how neural networks actually are trained so you have unrealistic expectations of what Tesla is doing with their fleet. We know the triggers AP2/2.5 cars have and that limited trigger data gathered from those cars really isn’t sufficient to train the vision networks Tesla uses, even if sent out in volume (and it isn’t sent out in such volume anyway). They have been trained elsewhere and whatever data is gathered from the fleet is validation and sprinklings on top at best at this time.

Maybe HW3 changes things who knows. I’m talking the current situation.

Tesla’s greatest asset so far is the ability to iterate in lab — which is where they train the networks — and deploy those iterations fast to the fleet for use and gaining some testing data. That is not an insignificat asset of course but it is not a fleet of learning cars in consumer hands.