Researchers prove you can serve millions of users with just a handful of customized AI models instead of one per person

A team developed a mathematical method to figure out the smallest set of AI models needed to satisfy wildly different user preferences without building a separate model for each person. Instead of maintaining millions of individual models, a company could keep a portfolio of maybe 10 or 20 specialized versions and route each user to the one closest to what they want.

Right now, personalizing AI to individual users is impossible at scale because you'd need to build and maintain a custom model for every single person, which would cost more than anyone can afford. This paper shows there's a hard limit to how many models you actually need — meaning the cost of personalization drops from astronomical to maybe practical. It turns out you can cover the entire landscape of human preferences with a small, fixed portfolio of models, which changes whether personalization is even economically feasible.