As of the end of 2021, Tesla has 2.2 million vehicles on the road either already equipped with its Full Self-Driving computer or capable of upgrading to it. The most recent comparable figure for Alphabet's self-driving car project, Waymo, is 600 vehicles. What do these numbers mean in practice?

First, we can pare down Tesla’s 2.2 million vehicles to the number who actually purchase the Full Self-Driving (FSD) option, which is about 10%. That makes the relevant figure 220,000 FSD-capable cars or 367 times more than in Waymo’s fleet.

However, this is not yet apples-to-apples. To account for the fact that each Waymo vehicle is likely driven approximately ten times more than each Tesla vehicle, we ought to adjust the figure to 22,000 Waymo-equivalent cars for Tesla. Then Tesla’s data advantage is 36.7x. 

Baidu helpfully published a paper that allows us to quantify what impact differences like this have on AI performance. Using Baidu’s scaling rate for computer vision, we can calculate that a 36.7x data advantage results in about a 50% AI performance advantage. From this simple method, we should expect, then, a minimum of a 50% performance advantage for Tesla over Waymo.

I say a minimum of 50% because, in one sense, the Baidu paper tests one of the worst case scenarios for Tesla. The Baidu paper uses homogeneous data, i.e. more examples of the same object type, like a fish or a cat or a flower. Part of the challenge of data gathering for autonomous cars is the extreme heterogeneity of the data; you’re looking for the first photo of a black bear standing on top of a car, not the billionth photo of a car. 

It is impossible, as far as I can tell, to rigorously estimate how far above 50% Tesla’s AI performance advantage is. This would require, I think, an extremely detailed knowledge of all the events that take place in the physical world, which no one has.