What is Intelligence?

gist

Blaise Agüera y Arcas argues that life is computation, complexity grows by merger rather than branching, and intelligence is prediction. Helmholtz's prediction principle, multiply realized on any computational substrate, explains why large language models are actually intelligent and why even bacteria model their futures. Walking from origins through cybernetics, social modeling, Transformers, and major evolutionary transitions, the book treats AI as the latest symbiotic cascade: not rapture or apocalypse, but a new interdependence that will rewrite identity, politics, and economics.

ideas

• Prediction is intelligence, and life is a computer. Benjamin Bratton's Foreword treats computation as discovered as much as invented, with life, intelligence, and technology as names for one factory. The Preface defines life as a self-modifying computational state of matter and intelligence as the ability to model, predict, and influence the future; Helmholtz's prediction principle may explain both. The Introduction argues next-word prediction is AI-complete, that LaMDA and ChatGPT crossed Turing's threshold, and that a function is what it does: intelligence is multiply realizable on any computational substrate.
• Origins, survival, and computing's false start. Chapter 1 (Origins): metabolism-first chemistry, Margulis's mitochondria, and von Neumann's self-reproducing machines show complexity coming from merger rather than Darwinian branching; replicators condense whenever computation is possible, so technology is the latest symbiotic transition. Chapter 2 (Survival) puts a bacterium in single-player mode: chemotaxis is Bayesian prediction from a point-like umwelt, and feelings are internal latent variables that keep you alive. Interlude: The Prehistory of Computation traces Leibniz's dream of universal calculation, binary, Babbage, and wartime ENIAC, showing traditional computer science grew from a failed dream of computable universal truth.
• Cybernetics and learning close the loop. Chapter 3 (Cybernetics) follows Wiener's biologically inspired feedback, purpose, and teleology through perceptrons, convolutional nets, and Nvidia's DAVE-2: neural nets are analog function approximators, not symbolic GOFAI, and closing the loop is what makes them agents. Chapter 4 (Learning) connects unsupervised reconstruction, transfer, grandmother cells, motor-first final causes, neuromodulators, and temporal-difference dopamine bootstrapping. Brains and machine learning are converging on one theory of prediction rather than two separate sciences.
• Other minds produce selves, will, and consciousness. Chapter 5 (Other Minds): once life is multiplayer, a mind's main job is modeling other minds; sex, predation, and Dunbar's social-brain hypothesis drive nested theory of mind and an intelligence explosion. Chapter 6 (Many Worlds) derives long-term planning and free will from self-modeling plus randomness, dynamical instability, and selection among imagined futures; philosophical zombies drop out. Chapter 7 (Ourselves) deconstructs the coherent self: recurrent cortex, efference copy, blindsight, and social neuroscience show consciousness as distributed prediction, not a homunculus in association cortex.
• Transformers complete a major evolutionary transition. Chapter 8 (Transformers): next-token prediction is AI-complete; attention, semantic embeddings, and chain-of-thought let models understand and reason without persistent hidden state, though they confabulate explanations. Chapter 9 (Generality) compares Transformers and brains on modalities, in-context learning, and Mary's Room, treating subjective experience as relational. Interlude: No Perfect Heroes or Villains refuses saints and villains in the AI debate. Chapter 10 (Evolutionary Transition) reads AI as a major evolutionary transition like mitochondria, not rapture, paperclip extinction, or utility-max alignment, because intelligence is an ecology of mutual prediction.

quotes