CONCEPT
Maximum Likelihood (as AI Training Objective)
The Fisherian principle—find the model parameters that make the observed data most probable—that sits at the mathematical heart of virtually every AI system now in production, carrying all of its power and all of its warnings into the present.
When an engineer trains a language model, she is running an optimization that
Ronald Fisher named, justified, and understood more deeply than anyone before him. Maximum likelihood is the procedure of finding the configuration of a model's parameters that makes the observed training data as probable as possible under that model's assumptions. Fisher formalized it in the 1920s to estimate gene frequencies and crop responses; the engineers of 2026 use it to train systems that generate text, code, and decisions. The objective is identical: maximize the probability of the data you have, under the model you are building. When a
large language model is trained to predict the next token, it is minimizing cross-entropy loss, which is algebraically equivalent to maximizing the likelihood Fisher defined. The mathematical heart of the most powerful AI systems in history is a principle a man worked out with a mechanical calculator before the Second World War. Fisher