Pegasus chip

[Chickenlittle]

so i will admit confusion about how AI will be used. I thought data transmitted to tesla servers for their AI to learn but Pegasus has AI on chip so does that mean each car teaches itself? I cannot imagine each car learning on own. That would result in different cars doing things differently. Educate me please

 

[ZeApelido]

Each car still needs to do object detection / inference / classification based on incoming sensor data (in real time). This will include forward passing data through several massive neural networks that leverage parallel processing on a GPU to perform enough computations in real time.

Better GPUs mean more computations can be done. Which can mean:

1) More computation frames / second
2) More sensors (i.e. can add lidar to cameras / radar)
3) More resolution to sensors (Camera resolution used can go up --> better object detection).

 

[animorph]

There's a difference between acting as an AI and training the AI. Training can take a lot more computer and data input resources, and/or time. The training results in a data set that can be downloaded to "reasonably" sized computer where it can act as an AI. There could be minor "learning" and calibration functions locally in the car that could make cars behave a little differently. But the main AI functions are trained outside the car and downloaded as part of the firmware.

It is also possible that local map data could be loaded in real-time and it could contain updated info or guidance that could make the car behave differently than before, even without a full firmware update. Something like "HD" mapping of each lane of a roadway might be updated to follow the lanes better than before.

I have no detailed knowledge of what Tesla is doing specifically.

 

[Tam]

...Pegasus has AI on chip so does that mean each car teaches itself...

I don't see how this teaching oneself would exclude fleet learning.

For example, one car may never see a traffic cone yet while another has been trying to run over traffic cones all the time.

One car does not know about traffic cone while the other one would learn on its own to avoid hitting traffic cones.

The experienced one would then share that learning to the rest of the fleet so that those who never encounter a traffic cones will know when they detect this kind of shape, they should avoid running over them.

 

[Chickenlittle]

There's a difference between acting as an AI and training the AI. Training can take a lot more computer and data input resources, and/or time. The training results in a data set that can be downloaded to "reasonably" sized computer where it can act as an AI. There could be minor "learning" and calibration functions locally in the car that could make cars behave a little differently. But the main AI functions are trained outside the car and downloaded as part of the firmware.

It is also possible that local map data could be loaded in real-time and it could contain updated info or guidance that could make the car behave differently than before, even without a full firmware update. Something like "HD" mapping of each lane of a roadway might be updated to follow the lanes better than before.

I have no detailed knowledge of what Tesla is doing specifically.

What would bother me is that when you get in a car that has its “own” AI experience then that would imply that some cars maybe better drivers in the same location. Will there be ratings based on safety/accident or near misses on each car? Reliability would be expecting each car to react the same in identical situations which would not be consistent with each car having its own AI chip. Would there be a rating situation when hiring a ride for those cars with higher mileage (hence higher teaching lesson) would be more favorable selection?