PJ Vogt

In the other seat was their partner, watching the monitor displaying a graphical interface designed by Dmitry Dolgov.

The people watching the screen would call out problems ahead, some discrepancy between what the sensors were seeing and what was actually in the road.

This is what teaching a car to drive actually looked like.

Two-person teams manning the cars, logging errors, going back to the office to troubleshoot, and then updating the code.

I asked Don Burnett about this era.

And while you're doing this and then like you leave work and you get in your car that you drive as a human, did you find yourself thinking more carefully like, how do I know what I know when I'm driving?

Like you're trying to teach a machine by day.

Did it affect how you thought about human driving by night?

Almost obnoxiously so to any passengers in the car with me.

I was obsessed with one big question, which is why do humans drive the way they drive?

And it turns out there were no good answers.

And I still think they're not great answers.

And instead of actually answering that question, we've just turned to machine learning to infer the deep truths behind why humans do what they do.

And so there's some basic principles that you can understand.

Like we try to minimize lateral acceleration, meaning you don't want to be thrown to the outside of your car when you're making a turn.

So you're going to slow down.

But how much do you slow down, right?

And it turns out that's contextual.

Don gave me an example.

So you're trying to figure out the right speed and angle for the car on one of those tight, curvy on-ramps onto the highway.