Book

What if intelligence is not a matter of knowing more, but of needing to know less? Almost every account of machine intelligence assumes that a better agent models the world more faithfully. This book develops the opposite position, and makes it precise. Its central claim is that intelligent computation is the selection of a state transition on the coarsest abstraction a goal permits - and that this abstraction is not a modelling choice but a computable consequence of the goal itself. The designer has no say in it.What follows can be measured in bits. Every task turns out to carry an intrinsic bandwidth. The cost of coordinating several agents is set by that bandwidth and is provably indifferent to the size and the entropy of the physical world behind it. Reliability, not accuracy, fixes the price. And an ontology may drift arbitrarily fast at no cost whatever - provided the drift is a symmetry the abstraction quietly absorbs.Unusually for a work of theory, the book also records where it broke: the predictions that were falsified, the constants that proved wrong, and what replaced them. For readers in information theory, reinforcement learning, multi-agent systems, and the foundations of AI.