[Jimbo]

It's complicated but there are some simple principles which help:

The car has many neural nets (NNs) running all the time. This is the "hydra" architecture that Andrej discussed. Imagine each "head" of the hydra has a different speciality, like recognising pedestrians, cars, figuring out drivable space, is it raining, ⁹etc.

The in-car NNs are playback only. They don't learn.

The "training" is done in-house. Dojo will be used for this soon.

Driving policy is traditional C++ code at the moment

 

[Jimbo]

Because the Net presented at AI day was enormous and probably ran into millions of dollars.

Eh?

What was presented at AI day was a system for training NNs and the end result, which is a bunch of NNs which run in the car.

Because they don't "learn" independently the NNs running in the car will never diverge.