The Uber-Lyft Theory of Models

Think about the last time you opened a rideshare app. You probably opened two. You looked at both prices, saw one was three dollars cheaper, and took that one. You did not think about it. You did not feel loyalty. You have no opinion about either company.

That is where models are going, and I think it happens faster than most people expect.

For a while there was a real quality gap, and quality gaps create loyalty. If one model could do the thing and the others could not, you used that one and did not shop around. But the gap between the frontier and the very good has been closing steadily, and for the overwhelming majority of tasks the very good is indistinguishable from the frontier. Once that is true, you are choosing between two cars that will both get you to the airport.

And then you look at the price.

The interesting thing about the rideshare comparison is how small the switching trigger is. Not fifty percent cheaper. Fifty cents. A rounding error on a twenty dollar trip is enough to flip the decision, and the person flipping is not poor. They have money. They just cannot construct a reason to pay more for something identical.

That is the part people miss when they argue that enterprises will pay for quality. Of course they will pay for quality. What they will not do is pay for the label of quality when the output is the same. A finance team that has money in the wallet will still route a summarization job to whatever costs a third as much, because there is no story they can tell about why the expensive one was necessary.

This is a race to the bottom in the sense that prices grind toward the cost of compute. But calling it only that misses what is actually happening, which is that the choice is getting decomposed.

It is not one model per company. It is one model per task. The cheap fast one for classification, the good one for anything a customer reads, the reasoning one for the handful of problems that genuinely need it, and a router in between deciding which is which. Nobody is loyal because nobody is making a single choice.

The consequence is that the durable business is not in being the best model in some abstract sense. It is in the layer that decides, in the workflow the model is embedded in, in the data that makes the output specifically useful to you. Those things have switching costs. The model does not.

Rideshare drivers figured this out years ago. They run both apps.