PERSON
Ray Solomonoff
The mathematician who defined the perfect predictor—then proved no machine could ever be it—founding algorithmic information theory, formalizing Occam’s razor as a number, and supplying the uncomputable ideal against which every real learning system, including today’s language models, can be precisely measured and found wanting.
Ray Solomonoff did not build the machines we now argue about. He did something stranger and more durable: he worked out, in the late 1950s and early 1960s, what the perfect version of those machines would even be. His central discovery—algorithmic probability, and the prediction method now called Solomonoff induction—is, by a precise mathematical argument, the best possible way to predict: the one that weights every consistent explanation by its simplicity, takes in all the evidence, and answers each question with the probability that an ideally rational agent would assign. It is also, by an equally precise argument, impossible to actually run. Both halves of that sentence matter, and the whole of modern artificial intelligence lives in the tension between them. The machines that predict the next token—the
large language models that draft our emails, write our code, and pass our examinations—are, whether their builders know it or not, crude computable