Retail, media, finance, science, industry, security and government increasingly depend on predictions produced through techniques such as machine learning. How is it that machine learning can promise to predict with great specificity what differences matter or what people want in many different settings? We need, I suggest, an account of its generalization if we are to understand the contemporary production of prediction. This article maps the principal forms of material action, narrative and problematization that run across algorithmic modelling techniques such as logistic regression, decision trees and Naive Bayes classifiers. It highlights several interlinked modes of generalization that engender increasingly vast data infrastructures and platforms, and intensified mathematical and statistical treatments of differences. Such an account also points to some key sites of instability or problematization inherent to the process of generalization. If movement through data is becoming a principal intersection of power relations, economic value and valid knowledge, an account of the production of prediction might also help us begin to ask how its generalization potentially gives rise to new forms of agency, experience or individuations.