Ask HN: Is Tesla actually using customer data for training?

3 pointssnrji6y ago2 comments

I've read many times something along the lines of "Tesla has an advantage in autonomous cars because they have the data of thousands of customers, which implies millions of miles".

However, I can't find any source detailing whether Tesla is actually using customer data for training (and not only for debugging).

If so, where does the computation happen? Locally? Then, are Tesla chips ready for training? Did they mention that when unveiling FSD hardware)? How do they sync/integrate/debug that data/training? Otherwise, if the training is done at Tesla, how they transfer gigabytes of data?

Also, wouldn't that be a perfect case for imitation learning?

And is it legal to use customer data that way?

Is it possible that FSD is "secretly" ready for training?

Thanks!

2 comments

kjksf6y ago

This is explained in detail in this talk: https://youtu.be/Ucp0TTmvqOE?t=6662

Calling it "customer data" is a stretch.

The software in Tesla cars can do the following:

- send pictures of the road taken at random intervals - record and upload clips of scenes / objects that match criteria pre-programmed by Tesla - record and upload clips during error events (e.g. when a driver over-rides the software, which implies that the software made a bad call) - record and upload clips when software is running in shadow mode and detects that it mis-predicted the behavior

And yes, this data is added to training set.

Uploaded data consists of short, compressed video clips during abnormal events. It's not full feed therefore not "gigabytes of data".

The user agrees to this in the software, so yes, it's legal.

snrjiOP6y ago

Thank you very much for your explanation! That's exactly what I was asking for.

j / k navigate · click thread line to collapse