Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.
Register

Coast to coast drive happening this year for all FSD Teslas!

This site may earn commission on affiliate links.
Do you have a link to his statement, I'm curious about the context in which it was said.

Yes. From the Tesla 2018 Q3 conference call:

Andrej Karpathy -- Director of AI & Autopilot Vision
Yeah, certainly. Hi, everyone. My name is Andrej Karpathy. I'm the Director of AI here at Tesla. And my team trends all the neural networks that analyze the images streaming in from all the cameras for the Autopilot. For example, these neural networks identify cars, lane lines, traffic signs and so on. The team is incredibly excited about the upcoming upgrade for the Autopilot computer, which Pete briefly talked about. This upgrade allows us to not just run the current neural networks faster, but more importantly, it will allow us to deploy much larger computationally more expensive networks to the fleet.

The reason this is important is that it is a common finding in the industry and that we see this as well is that as you make the networks bigger by adding more neurons, the accuracy of all their predictions increases with the added capacity. So, in other words, we are currently at a place where we've trained large neural networks data work very well, but we are not able to deploy them to the fleet due to computational constraints. So all of this will change with an acceleration of the hardware and it's a massive step improvement in the compute capability and the team is incredibly excited to get these networks out there.
Tesla Motors Inc (TSLA) Q3 2018 Earnings Conference Call Transcript -- The Motley Fool
 
No. Even what Tesla originally sold as "full self driving" didn't require a human driver.

FSD will require a safety driver at first but later, when it gets safe enough to do so, the driver will not be required. In fact, eventually, the steering wheel will be removed completely. So I don't see any contradiction: Tesla sold FSD as not requiring a driver and that is exactly what FSD will be once it is reliable enough.

This is not a hard concept to understand. FSD is not something that Tesla can just work on for awhile and test for a long time and then release one big L5 software update where we instantly have a robotaxi on day 1. FSD is a complex process. Just look at Waymo that also uses safety drivers but is gradually removing them as the system gets more reliable. That's how it works. In the beginning, for safety, FSD will require a safety driver, and then as it gets better, the driver can be removed. So in the end, yes, FSD will not require a driver at all. But that's the end, not the beginning.
 
Last edited:
  • Disagree
Reactions: DrDabbles
Yes. From the Tesla 2018 Q3 conference call:

Andrej Karpathy -- Director of AI & Autopilot Vision
The reason this is important is that it is a common finding in the industry and that we see this as well is that as you make the networks bigger by adding more neurons, the accuracy of all their predictions increases with the added capacity. So, in other words, we are currently at a place where we've trained large neural networks data work very well, but we are not able to deploy them to the fleet due to computational constraints. So all of this will change with an acceleration of the hardware and it's a massive step improvement in the compute capability and the team is incredibly excited to get these networks out there.
Tesla Motors Inc (TSLA) Q3 2018 Earnings Conference Call Transcript -- The Motley Fool

That's great thanks.

Do we know how Tesla curated this training with the improved input data back in Q3 2018? Through simulations or had they deployed a fleet of early AP3 cars in 2017?
 
I doubt that Tesla had any AP3 cars back in 2017. At least, we never heard of any. My guess is Tesla used simulations.

Would they need AP3 vehicles to train AP3 networks? AP2.5 may have been able to collect full resolution training data for AP3 networks for the past couple of years. AP3 hardware is only required to run the result of the network in the vehicles, only AP2.5 would have been required to train.
 
Would they need AP3 vehicles to train AP3 networks? AP2.5 may have been able to collect full resolution training data for AP3 networks for the past couple of years. AP3 hardware is only required to run the result of the network in the vehicles, only AP2.5 would have been required to train.

Yes, Tesla used AP2 to collect the data to train the AP3 NN. They just could not deploy the NN in AP2.
 
  • Like
Reactions: willow_hiller
Yes. From the Tesla 2018 Q3 conference call:

Andrej Karpathy -- Director of AI & Autopilot Vision
Yeah, certainly. Hi, everyone. My name is Andrej Karpathy. I'm the Director of AI here at Tesla. And my team trends all the neural networks that analyze the images streaming in from all the cameras for the Autopilot. For example, these neural networks identify cars, lane lines, traffic signs and so on. The team is incredibly excited about the upcoming upgrade for the Autopilot computer, which Pete briefly talked about. This upgrade allows us to not just run the current neural networks faster, but more importantly, it will allow us to deploy much larger computationally more expensive networks to the fleet.

The reason this is important is that it is a common finding in the industry and that we see this as well is that as you make the networks bigger by adding more neurons, the accuracy of all their predictions increases with the added capacity. So, in other words, we are currently at a place where we've trained large neural networks data work very well, but we are not able to deploy them to the fleet due to computational constraints. So all of this will change with an acceleration of the hardware and it's a massive step improvement in the compute capability and the team is incredibly excited to get these networks out there.
Tesla Motors Inc (TSLA) Q3 2018 Earnings Conference Call Transcript -- The Motley Fool
If they already had neural networks that work very well they could demo them by just filling the trunk with computers. If they could demo it I'd think they would simply to sell more FSD cars and support the stock price even if the hardware was a year away.

They haven't done anything like that so I'm very skeptical they have anything like FSD.
 
If they already had neural networks that work very well they could demo them by just filling the trunk with computers. If they could demo it I'd think they would simply to sell more FSD cars and support the stock price even if the hardware was a year away.

They haven't done anything like that so I'm very skeptical they have anything like FSD.

Tesla did test drives on Autonomy Investor Day that did show that they have alpha FSD that works. So they did demo it.
 
Tesla did test drives on Autonomy Investor Day that did show that they have alpha FSD that works. So they did demo it.

Yeah, strange how they wouldn't let anyone record their demo rides though. All we got was that one Telsa produced video which for all we know could have been run 97 where it finally managed to go the whole loop without crapping out.

If I were a less trusting person I might think they were trying to hide how far away from FSD they really are.