Yeah, they have special ui in the data gathering cars to mark situations for later training and debug and whatnot as they drive around.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Waymo’s “commercial” ride-hailing service is... not yet what I hoped
- Thread starter strangecosmos
- Start date
Yeah, they said it could do FSD. The video description they reference says "Take a ride in a Tesla with Full Self-Driving Hardware." and the blog post that references the video says "We are excited to announce that, as of today, all Tesla vehicles produced in our factory – including Model 3 – will have the hardware needed for full self-driving capability". Hence combining it all together they are saying that any car could do it after some short calibration.
Sure, you may know more than me, but this knowledge is useless to us all because you cannot share it. You are welcome to point where I am incorrect in my investigations, though, be as specific as you can, not just some vague "it's all wrong"
And where am I disagreeing with this?
This is a very nebulous claim. 10x faster doing what? "the nvidia chip" being which chip? (you do know Tesla uses 4 different NVidia chips in their mcu1 cars, and 2 different NVidia chips in their MCU2 cars, right?)
NVidia does not have any AI chips that I am aware of, they do general purpose GPU chips that happens to be good for NN stuff too. The specialized NN chips are coming from a bunch of other companies.
So off the top of your head (since you seem to know so much about this NN stuff) what AI chips does Intel currently do, what companies that do AI chips they purchased in the last several years and how do those compare to general purpose GPU setups from NVidia in various important NN benchmarks?
I think it is very simply because we have looked into it too and came to different conclusion about its applicability here.
AlphaGo was super impressive. The ImageNet solution rightfully started a revolution. The notion to just drive a lot with cameras and it will learn is inviting and enticing. We have all seen the Nvidia end to end NN driving demo I am sure.
Personally I just don’t believe it is that simple for general autonomous driving. I expect it to require a mix of technologies and approaches to get there which is why I consider Waymo the current leader and not say Tesla.
Not even Tesla runs end to end NNs in Autopilot 2. The steering wheel is not being turned by an NN.
CarlK
Active Member
@verygreen
What do you mean by discussed at length? You could very well be wrong no matter how long you have discussed it. Care to post your arguments here so I can see them?
What do you mean by discussed at length? You could very well be wrong no matter how long you have discussed it. Care to post your arguments here so I can see them?
CarlK
Active Member
@verygreen
So many
So many of what you said are wrong I don't even know where to start.
Why Tesla Dropped Nvidia's AI Platform For Self-Driving Cars And Built Its Own
BTW Nvidia is more classy than Mobileye in its reaction to Tesla's new chip. Instead of hurling insults it only said we welcome Tesla to come back if it does not work out for them..
So many
Yeah, they said it could do FSD. The video description they reference says "Take a ride in a Tesla with Full Self-Driving Hardware." and the blog post that references the video says "We are excited to announce that, as of today, all Tesla vehicles produced in our factory – including Model 3 – will have the hardware needed for full self-driving capability". Hence combining it all together they are saying that any car could do it after some short calibration.
Sure, you may know more than me, but this knowledge is useless to us all because you cannot share it. You are welcome to point where I am incorrect in my investigations, though, be as specific as you can, not just some vague "it's all wrong"
And where am I disagreeing with this?
This is a very nebulous claim. 10x faster doing what? "the nvidia chip" being which chip? (you do know Tesla uses 4 different NVidia chips in their mcu1 cars, and 2 different NVidia chips in their MCU2 cars, right?)
NVidia does not have any AI chips that I am aware of, they do general purpose GPU chips that happens to be good for NN stuff too. The specialized NN chips are coming from a bunch of other companies.
So off the top of your head (since you seem to know so much about this NN stuff) what AI chips does Intel currently do, what companies that do AI chips they purchased in the last several years and how do those compare to general purpose GPU setups from NVidia in various important NN benchmarks?
So many of what you said are wrong I don't even know where to start.
Why Tesla Dropped Nvidia's AI Platform For Self-Driving Cars And Built Its Own
BTW Nvidia is more classy than Mobileye in its reaction to Tesla's new chip. Instead of hurling insults it only said we welcome Tesla to come back if it does not work out for them..
Last edited:
CarlK
Active Member
I don't know if the steering wheel is turned by NN but it's all that better if it is. If you missed my how batter hit balls analog please think about it again.
If you just plan to rehash posts from uninformed journalists, that's not going to be a useful conversation, I am afraid. It's not like those articles were not seen and taken apart before. If you want to add insight, show us the code or some hard data points, or some original analysis, not some nebulous stuff rehashed to the Nths degrees by overly optimistic people that often don't even know what they are talking about.
Also in case you did not know, Mobileye does not exist anymore for quite some time. Personally I don't care about Tesla "new chip", and you know why Nvidia does not care? Because Tesla is still going to buy chips from them for the autopilot platform even for hw3, it's a bad business decision to insult your clients as you might imagine.
For some of the previous discussions see the thread here (I am linking one of my posts but you probably need to read some other ones in there): Tesla Autopilot miles estimated at over 1.2 billion
CarlK
Active Member
Tesla is not going to use Nvidia's PX2, up to now world's most powerful general purpose AI chip for self driving car availabe, when it puts its new AI chip in the new HW3 board. Tesla knew the Nvidia chip will probably not enough to achieve FSD even then.
Look inside Tesla’s onboard Nvidia supercomputer for self-driving
Mobileye, in case you did not know, still exists as an Intel subsidiary.
There is a thing called Google search. Why you don't even bother to use it to fact check before saying all those things you said?
I've seen a lot of opinions in that thread but not anything of substances to dispute Tesla's claim of machine learning. If MIT self driving car lab, which I assume is a knowledgeable and unbiased organization, did not think those Tesla machine learning miles are important why would it even bother to tabulate that data?
Please trust me when I say I understand you batter analogy. I understand why AlphaGo is great, why ImageNet was significant, why Nvidia’s end to end demo was enticing.
I simply do not believe deep learning or what you call NN machine learning alone will solve autonomous driving in the foreseeable future due to the more complex nature of the task. It will take a mix of technologies in my view.
I simply don’t believe training an NN to generically drive a car in all its variety is possible at this time. What’s more I don’t think anyone is trying either. Everyone uses a mix of technologies. Tesla too.
I think we probably could train NNs to replace the batter though.
Also I don’t believe in Tesla’s volume advantage on this in any case because I don’t believe they actually train NNs using consumer cars so your thesis is simply vastly different than mine and hence we disagree.
Tesla does have some benefits like ability to get testing or validation data from the consumer fleet and deploy updates fast to the consumer fleet.
CarlK
Active Member
I don't think we have that much disagreements. I don't thinking machine learning is used to the level of directly "driving" the car although nothing says it could not do that either. "Driving" a car is actually not that a hard thing to do. The machine learning is only used to make decisions of where to go and where not to go and how fast it should go, and to make better decisions as it acquires more data. The only place we still have disagreement is I do believe Tesla uses data from "consumer cars" to do the learning. Everyone says that including many insiders. What are reasons you think it does not or can not do that?
Last edited:
You don't even know what you are talking about, do you? I recommend you to read up on what NVidia PX2 is at Drive PX-series - Wikipedia hint: it has multiple configurations with multiple possible chips. Moreover, you will learn that what it really is is just "general purpose GPU" married to some arm cores.
I see you did not work at an "Intel subsidiary" to know what "Intel subsidiary" means too?
Yeah, there's this wonderful google search thingie that you are neglecting to use before bringing in your theories. Really strange since you insist on others to use it first.
I have precise examples of what a snapshot is like, but of course it's nothing of substance because oh, a research center that can only speculate how the system works inside said something. There's zero actual learning in the car (there is inference using a pre-trained model, it's used for vision in particular, but it's not learning in car).
I don't know why MIT people think those miles are important and I don't agree with them and I have the data to back it out and they don't.
It is a sum of too many things read and learned to list — it really is easier to just disagree — but common sense does guide me that way too. Gathering sufficient data to actually train an NN from the car, store in car and transfer over networks does seem prohibitive size wise.
It has always seemed more likely to me Tesla would use remote data for validation not NN training and that is how Elon Musk originally portrayed it as well. We know what the cars do send back home and that again supports this concept — it is not really NN training type of data.
In theory anything can change over time but I do not believe training NNs from customer car data is Tesla’s approach.
CarlK
Active Member
This is one of the earliest articles about Tesla machine learning. Answered a lot of questions you have.
How Tesla is ushering in the age of the learning car
Actually they might use customer data for training. For example imagine they are debugging detection of gore lanes and ask some customer cars to provide a picture when it thinks it's in the gore lane. Some cars trigger the condition and upload the resulting pictures. Some of them are correct and some of them return something like below:
Could they just place the false positives into their training set marked appropriately to avoid this false positive in the future? Absolutely! (assuming somebody actually reviews those snapshot uploads and actually vets them).
Can they see something interesting not in the false positive images and also include those into the training sets to further reinforce something? Yup!
But what cannot happen is they cannot just have the car itself decide "hey, I think I misdetected something" or "there's a bicycle here, but nothing like that in my visual detections feed" or other such unknowables and report it to the mothership.
Of course stuff like "hey, I would turn left here but the driver goes straight" or "I would not brake here but the driver did" is even more out of the realms of possible with current firmware.
I am not sure you understand me. I don’t have any questions about this and I know all that history.
I simply do not believe it means what you think it means. I do not believe Tesla trains NNs on consumer cars and I do not believe that is their plan.
Validation? Yes. Whitelisting or HD mapping in the future? Yes. NN training? No.
So I asked, the way the project started is that Google bought 2 companies and merged them. One of them was already developing a lidar based solution. At this point Google and Nvidia shared the opinion that Lidar will be a better solution so they went with that.
lunitiks
Cool James & Black Teacher
CarlK
Active Member
You think Tesla is collecting big data from consumer cars but not using them for NN training? That does not make sense to me. Unlike others Tesla has very few test cars if any. All its data are from consumer cars.
Last edited:
Yes that is generally what I believe.
I believe they plan to make use of the consumer data for many things validation, white listing, mapping and such. Validation alone is a big benefit for Tesla for sure.
Definitions matter of course. @verygreen is right that Tesla could feed false positive snapshots into image recognition labelling or something like that and that is what I put under validation.
What I do not think happens is consumer cars recording driver inputs and camera inputs and then using those to train NNs to drive or even positively see. I don’t think the cars learn as such.
Such NN training — to the extent NNs do it — is done by engineering cars and simulators and this is why I feel Tesla does not have a volume benefit over others in the area of NN training.
The actual static NNs shipping in Autopilot 2 cars were trained by engineering cars and simulation and they do not change while in the car beyond software updates.
Tesla’s main volume benefit from their fleet in my view is in validation and software deployment speed. These are not insignificant benefits but they are not direct NN training benefits.
Last edited:
lunitiks
Cool James & Black Teacher
Similar threads
- Article
