Solomonoff Induction

In the [YOU] on AI Field Guide

The cycle that began with [YOU] on AI keeps returning to the question of what these systems actually know—whether they understand or merely predict, whether their fluency reflects competence or mimicry. Solomonoff induction reframes this debate. There is a correct way to learn from experience, he showed; it is the one that weights explanations by simplicity and updates on evidence by Bayes’ rule. Real systems are good to the extent that they track this ideal and bad to the extent that they depart from it. The question is no longer the vague “does it understand?” but the sharp “how close does this approximation come to the optimal predictor, and where does it fall short?”

The fall-short is where it gets interesting for the cycle’s concerns. Because a deployed model is a single fixed approximation rather than the ideal sum over all explanations, it has no access to the full space of hypotheses the ideal would weigh. It cannot represent the full range of uncertainty the way the ideal predictor would. This is part of why models can be fluent and confidently wrong at once: they have compressed data into one bounded model without the ideal’s built-in acknowledgment of every alternative explanation. A hallucination, in this light, is an approximation error wearing the costume of certainty—exactly what the cycle identifies as the signature hazard of large language models.

Solomonoff induction also clarifies the permanence of the AI capability gap. The optimal predictor is not merely difficult to build but proven impossible to build. Every actual system is forever at a distance from the ideal. This cuts against the most ambitious claims about imminent artificial general intelligence: optimality is not a destination that more compute will eventually reach but an asymptote that every system approaches without arriving. Progress is real; perfection is structurally excluded.

Origin

Solomonoff published the core of the idea in a 1960 technical report, “A Preliminary Report on a General Theory of Inductive Inference,” and developed it in two 1964 papers in Information and Control. He had circulated earlier versions as early as 1956 and 1957, including a paper titled “An Inductive Inference Machine” presented at a 1957 symposium. The construction fused four ideas: Occam’s razor (prefer simpler explanations); the principle of multiple explanations attributed to Epicurus (if several explanations are consistent with what you have seen, keep all of them); the universal Turing machine (which let “explanation” mean “program” and “simplicity” mean “program length”); and Bayes’ rule (which turned the weighted explanations into a genuine probability for the next observation).

The theory went largely unnoticed until the Soviet mathematician Andrey Kolmogorov independently developed the related notion of program-length complexity in the mid-1960s. Kolmogorov’s framing emphasized randomness rather than prediction, and the field—now called algorithmic information theory—became associated with his name as much as Solomonoff’s. Marcus Hutter extended the prediction framework to agents that act in his 2004 book Universal Artificial Intelligence, producing the AIXI model, which is to acting agents what Solomonoff induction is to passive predictors: the provably optimal policy, and provably uncomputable.

Key Ideas

The construction: all programs, weighted by length. The algorithmic probability of the next symbol is computed by running all programs consistent with the observed data and letting them vote, with shorter programs receiving exponentially more weight. Occam’s razor is not imposed from outside as a preference; it falls out of the mathematics as a consequence of counting programs by length. The resulting predictor is Bayesian in the most thoroughgoing sense: it considers every possible hypothesis and updates on all the evidence.

Optimality and convergence. For any data sequence with a computable generating process, Solomonoff induction converges on the correct predictions in the limit. No computable method can systematically outperform it: any method that does better on some sequences will do worse on others, and Solomonoff’s method is optimal in expectation over the universal prior. This is the strongest possible statement about what it means to learn correctly from experience.

Uncomputability: the permanent gap. The sum over all programs is uncomputable: it ranges over programs that never halt, and the halting problem guarantees you cannot tell in advance which do. Every real system approximates Solomonoff induction by restricting to a tractable subset of programs—the neural network architectures and training procedures that constitute modern deep learning. The gap from these approximations to the ideal is not temporary but proven permanent. The interesting engineering questions are about how the approximation spends its finite budget: which regularities it captures cheaply and which it misses.

The prediction-compression identity. Solomonoff showed that predicting the next symbol well is mathematically equivalent to compressing the data, which is mathematically equivalent to finding its regularities. These are one task seen from three angles. The implication for modern AI: training a model to minimize prediction error is training it to compress, which is training it to find structure. The compression achieved by a large language model is the bounded version of the compression Solomonoff defined as the whole of inductive intelligence.