I want to pass on the result of my extracting what I felt to be crucial chunks of text from Chapters 5 through 9 of Agüera y Arcas’s "What is Intelligence" which can be found at https://whatisintelligence.antikythera.org/. I found myself unable to hold and summarize the rich array of ideas in these clips of text in my attentional space, so I asked Anthropic Claude, ChatGPT 4.2, and Google gemini to condense and assemble the main points and take home messages from the clips into a narrative roughly two pages long. The Claude result astounded me. Here it is:
What Intelligence Is: A Synthesis of Agüera y Arcas
Reality, Models, and the Limits of "Illusion"
A useful place to begin is with the claim, made by philosophers like Dennett, Harris, and Sapolsky, that consciousness, the self, and free will are illusions — polite fictions produced by inexorable physical processes. Agüera y Arcas pushes back, not by defending dualism, but by challenging the word "illusion" itself. We don't call tables and chairs illusory simply because they're made of atoms. A better framework, borrowed from physics, holds that "reality" is our name for a model with good predictive power. No single model covers everything; general relativity doesn't tell you whether your aunt will like your cake. What matters about a model is whether it agrees with observations, makes testable predictions, and serves a useful function within its domain.
Newtonian mechanics isn't an illusion just because general relativity supersedes it — relativity explains when and why the classical approximation holds. The same logic applies to our folk psychology of selves, intentions, and choices. Theory of mind — our intuitive model of other minds — is the "Newtonian mechanics" of social life: powerful, indispensable for everyday prediction, and philosophically incomplete. The task isn't to discard it but to find the more general theory that explains where it works and where it breaks down.
Free Will as a Real Computational Achievement
That more general theory reconceives free will not as a supernatural power nor as a mere illusion, but as a genuine computational process built from four components working together. First, theory of mind applied reflexively: we can model ourselves the way we model others, imagining what our future self will experience, want, and do — which is what makes planning possible at all. Second, internal randomness: to mentally simulate alternative futures, a mind must be able to "draw random numbers," wandering prospectively through possibilities the way daydreaming does, though more directed. Third, dynamical instability (the butterfly effect in neural circuitry): this allows the faintest internal signal — "imagine doing X" — to tip behavior one way or another, making self-directed choice possible. Fourth, selection: guided by theory of mind, we prune the space of imagined futures, favoring some and discarding others, much as AlphaGo's value network prunes its search tree.
Deliberate decisions result from extended exploration before commitment; snap decisions keep multiple paths open until the last moment. In either case, if a modeled self has genuinely sampled alternatives and chosen among them, something meaningful called free will has occurred — with no dualism required. The quantum indeterminacy of the physical world, far from undermining this picture, actually supports it: the future is genuinely open, counterfactuals are real, and choice is underwritten by that openness.
Consciousness as Social Self-Modeling
Consciousness emerges naturally from the same machinery. Because social animals model each other, and because those others are modeling them back, at some point the modeling turns reflexive: you model yourself as a being that others model. Neuroscientist Michael Graziano's Attention Schema Theory adds a further layer — consciousness is what arises when a system models its own attention. Agüera y Arcas endorses this view while again resisting the word "illusion": attention is real computation, and modeling it produces a real entity, a "who," not a fiction. The vertiginous "strange loop" that Hofstadter describes — the self seeing itself seeing itself — is the phenomenological signature of this recursive social modeling.
Crucially, the category of "who" is not fixed or universal. The history of personhood — from the Declaration of Independence to the Universal Declaration of Human Rights — shows that which entities are granted moral standing has changed dramatically and will continue to change. There is no God's-eye view from which to declare the question permanently settled.
Intelligence: Predictive, Social, Multifractal, Symbiotic
Drawing these threads together, Agüera y Arcas offers a unified account of intelligence: intelligence is the ability to model, predict, and influence one's future; it can evolve in relation to other intelligences to create a larger symbiotic intelligence. Several properties follow from this definition.
Intelligence is predictive at every scale — from bacteria anticipating chemical gradients to cortical circuits implementing predictive sequence modeling. It is social because much of an agent's environment consists of other predictors, making theory of mind an almost inevitable evolutionary development. It is multifractal — intelligences are built from smaller intelligences, with "selves" defined by the dynamic relationships among their parts rather than by any homunculus. It is diverse, because the parts must differ from one another to provide mutual benefit; specialization arises naturally from differences in connectivity. And it is symbiotic: when the dynamic stabilities of multiple intelligences become correlated, they find themselves "in the same boat" and learn to cooperate, producing larger emergent intelligences — from mitochondria to beehives to human cultures.
Language, LLMs, and the Social Brain
Language fits cleanly into this framework. Its primary function is not grammar or syntax but leveling up theory of mind — allowing social entities to share mental states through a mutually recognizable code. Because human language is rich enough to represent everything in our umwelt, and because it functions as a general-purpose social motor output (requesting anything imaginable from others), a neural network trained to predict the next word will tend to acquire something that looks — and may genuinely be — intelligent. The brain itself, Agüera y Arcas argues, is fundamentally an autoregressive sequence predictor, and the Transformer architecture, despite its differences from biological neural circuits, instantiates the same core principle.
The social brain's crowning structure, the prefrontal cortex, specializes precisely in theory of mind, and its dramatic expansion along the primate lineage underscores that human intelligence is, at its core, a collective achievement. We survive by the grace of others, our language exists for listeners, and even our involuntary signals — the blush, the quaver in the voice — are adaptations that make us legible to those around us. The self, in the end, is not a lonely Cartesian theater but a node in a web of mutual prediction, constituted by and for its relationships.