Elon Musk: You can't solve self-driving unless you have millions of cars on the road.
“We are no longer compute-constrained for training.
I checked in with the team, is there anything we could do to improve the pace of progress with respect to training and inference?
And currently, that is not the limiting factor.
In fact, the limiting factor right now is that the amount of miles between interventions is so long that it takes quite a while to figure out which version is better than the other version because none of them are requiring any interventions.
You start getting to thousands of miles between interventions, you need 10,000 miles to get an intervention. The average person only drives about 10,000 miles in a year.
You start getting to thousands of miles between interventions, you need 10,000 miles to get an intervention. The average person only drives about 10,000 miles in a year.
If it's in an urban environment, the average speed is 20 mph. Our professional test drivers get pretty bored, frankly. They're like, okay, I drove all week and there was no intervention. The highlight of the week would be like, yes, an intervention, finally! It's getting to that point.
So, this is where, actually, having a giant fleet is extremely important, because we can deploy a new FSD model and run it in shadow mode and see how well it performs, compare how the human drives the car versus the new self-driving build, and then, analyze that delta in shadow mode, like the shadow knows or doesn't know as the case may be, and then, be able to assess by getting billions of miles very quickly with the giant fleet.
You start getting to thousands of miles between interventions, you need 10,000 miles to get an intervention. The average person only drives about 10,000 miles in a year.
If it's in an urban environment, the average speed is 20 mph. Our professional test drivers get pretty bored, frankly. They're like, okay, I drove all week and there was no intervention. The highlight of the week would be like, yes, an intervention, finally! It's getting to that point.
So, this is where, actually, having a giant fleet is extremely important, because we can deploy a new FSD model and run it in shadow mode and see how well it performs, compare how the human drives the car versus the new self-driving build, and then, analyze that delta in shadow mode, like the shadow knows or doesn't know as the case may be, and then, be able to assess by getting billions of miles very quickly with the giant fleet.
