In this 2000 paper, Marcus Hutter combined sequential decision theory with Ray Solomonoff’s theory of universal induction to define a single agent he called AIXI. The construction is striking in its economy: an AIXI agent has no tunable parameters and no task-specific knowledge, yet Hutter argues it is the most intelligent unbiased agent possible. It works by considering every computable hypothesis about its environment, weighting each by a universal prior that favors simpler explanations, and then choosing actions that maximize expected future reward across all those hypotheses.
The paper shows that this one formalism subsumes a wide range of problems usually studied separately: sequence prediction, strategic game playing, function optimization, and both reinforcement and supervised learning all fall out as special cases. AIXI is in this sense a clean mathematical definition of what optimal general intelligence would mean if computation were free.
The catch, which Hutter states plainly, is that AIXI is uncomputable, because the universal prior it relies on cannot be calculated. To address this he constructs a time and space bounded variant called AIXI-tl, which remains provably more intelligent than any other agent constrained to the same compute budget, at the cost of running time on the order of t times 2 to the power l.
For a general reader, the value of AIXI is conceptual rather than practical: it gives a precise yardstick for what intelligence could be, which helps clarify debates about how far real systems are from a theoretical ceiling and why simplicity-favoring priors keep reappearing in machine learning.