glossary

Frank Jackson's thought experiment, dissolved: to see red for the first time you need a model of redness, not just a retina. Sense organs are convenient prompts to existing internal models, not the seat of qualia. Charles Bonnet syndrome (vivid hallucinations as vision fails) and "prisoner's cinema" show the visual cortex's hallucination keeps running when sensory error-correction is removed: perception was the model all along. AI systems can have noses (and eyes, and ears) via textual or multimodal inputs.

Chalmers/Schneider thought experiment: replace neurons with computational equivalents one at a time, would you notice? Functional answer: no, because identity is the dynamical pattern, not the substrate. Concrete vehicle for substrate independence; forces explicit commitment to functionalism.

The language-specialized brain region (typically left hemisphere) whose job is real-time autocompletion of a unified self-narrative, including post-hoc explanations for actions it did not initiate. Functionally equivalent to a fluent confabulator. In one classic split-brain experiment, the left hemisphere invents reasons ("the chicken claw goes with the chicken; you need a shovel to clean out the chicken shed") for actions cued in the disconnected right hemisphere.

The empirical phenomenon (Johansson) that ~80% of subjects fail to notice when their stated preference is surreptitiously swapped, and their justifications for the swapped choice are statistically indistinguishable from genuine ones. Effect persists across sex, age, response time. Even in political polling: 92% of respondents accept altered survey responses, and 48% become willing to switch coalitions.

The felt-unity-from-within of a tightly coordinated mutual-prediction system. In a rowing crew, swing has both an external face (measurable unified motion) and an internal face (each rower's experience of unison). AyA's structural analogue for what consciousness is to a recursive mutual-prediction loop: real, functional, subjective, distributed, not localized. Swing is a verb, not a thing; so is the self.

The model an agent builds of another agent's P(X,H,O), recursively including the other's model of the modeler. AyA's strong claim: theory of mind is mind, both in social space (between brains) and within a single brain (between cortical columns or hemispheres). Promotes ToM from a special-purpose social skill to the general operation that produces selves at every scale, dissolving the boundary between "social" and "general" intelligence.

The observer-relative property of being fully predictable to a modeler. Hofstadter's coinage, sharpened: when an agent becomes perfectly predictable to you, it stops registering as agential to you. Not a property of the agent in itself, but of the relationship between the agent and its observer.

The brain's model of its own attention. AyA extends Graziano: consciousness is what it is like to model your own modeling of attention, hence personhood is "that which can pay attention and can model that attention." Operational definition that survives both the puppet case (ventriloquism conjures a "who" via attention modeling) and the AI case.

The phenomenon (Humphrey, Weiskrantz) in which subjects with destroyed primary visual cortex perform visual tasks reliably while reporting subjective blindness. Reveals not "competence without consciousness" but that the interpreter is connected to one cortical region's model and not to the (still-functioning) subcortical visual pathway. "I am unaware of X" is a claim from one brain region, not a global fact about the system.

Bach's developmental ladder of self-reference, extended by the framework to four degrees:

1. Perception: content is present; the system models reality, finds patterns.
2. Consciousness (the observer): the system notices that something is noticing.
3. Self-model: "I am the thing making these models."
4. Meta-self (framework extension): the "I" itself is recognized as a construction produced by the same process that produces everything else in experience.

Not merely a developmental sequence: distinct computational regimes that systems can stably occupy.

The upstream organism-level evaluation system that generates emotions and tells the conscious self how to feel about its relationship to the world. Distinct from the conscious self that experiences the resulting affects.

Not a story about a supernatural being creating a physical universe, but a phenomenological description of infant consciousness initializing dimensions out of tohu wa-bohu: light from dark, dimensions assigned, objects built, space filled, self-model constructed, consciousness binding to it. "The world that exists inside of the dream of reality each of us experiences." Every one of us does this.

An executor on some substrate. A functional, not formal, definition: the automaton is whatever runs on the substrate.

The property of computations to be implemented on arbitrary substrates given Turing equivalence. AyA folds this under final causes: only purposive systems care about multiple realizability, because only purposive systems are defined by their function rather than their parts. Substrate independence as a structural consequence of having a final cause, not a free-floating philosophical thesis.

The structure of information processing in the brain is more a function of the problem the brain is to solve than of the brain itself. Train different architectures on the same problem and equivalent structure emerges across models, given by the mathematics of the problem.

The cortex viewed not as a single integrated organ but as a population of generic prediction units (cortical columns) that have replicated within skulls under selection pressure for social intelligence. Each column does the same kind of work; specialization arises from connectivity. Explains why ferret auditory cortex can learn to see, why blind people use their visual cortex for click sonar, and why scaling intelligence is replicating columns rather than inventing new tissue.

The architectural property of feedback connections in cortex (and RNNs) that allows shallow networks to perform deep computation by iterating in time. Cortex is shallow in space (a few layers) and deep in time (many feedback iterations). The "double-take" is an early exit from a deep iterative computation. Reconciles the shallow cortex with the deep visual hierarchy required for invariance: time and depth are interchangeable.

The rapid one-shot sequence-and-position encoder that tags incoming activity patterns with spatiotemporal positional encodings (grid cells, place cells), to be replayed during sleep for slow consolidation in cortex. Explains why anterograde amnesia leaves memory recall intact while breaking memory formation. Suggests positional encodings in Transformers and grid cells in hippocampus are the same solution to the same problem.

The subcortical action-selection circuit implementing a softmax-like winner-take-all over competing cortical activation patterns, mediated by dopamine. Does the work of habitual or "autopilot" decisions. Replaces "lower brain reflexes" hand-waving with a specific computational role: lateral inhibition over candidate actions.

Chemical signals (dopamine, serotonin) that act as long-timescale internal hidden-state variables H, broadcasting integrated estimates of behaviorally relevant quantities (anticipation, satiation) to whole populations of neurons. Not "reward" or "pleasure" but predictions of future reward. Repositions affect chemistry within P(X,H,O): neuromodulators are the brain's H, not its scalar utility function.

The signal sent from a motor region to other brain regions reporting on its own state. Traditional reading: a "carbon copy of a motor command." AyA's reading: not a copy, but the motor region's contribution to a mutual-prediction loop with sensory regions. Theory of mind operating between brain regions.

The unsupervised neural-net architecture that learns to inpaint missing pixels (or tokens). AyA reads vision as a biological masked autoencoder: saccades hide most of the visual field at any moment; the brain inpaints; saccades sample the inpainting against reality; learning ensues. Maps controlled hallucination onto a precise ML training regime.

The joint probability distribution over observable variables X, hidden / internal-state variables H, and output / action variables O that any adaptive agent learns. The minimal mathematical object that captures perception, internal state, action, and their interrelations: the technical heart of AyA's framework. All of intelligence becomes "approximating P(X,H,O) under dynamic-stability constraints."

The organism-specific compression scheme over latent variables that survival has selected as predictively relevant. Borrowed from von Uexküll, given a precise computational reading: not "subjective world" loosely, but the set of latent variables that an agent's P(X,H,O) has converged on because tracking them works. Different umwelten are different compression schemes shaped by different selection pressures.

The invariant dimensions a learned model factors out of raw sensory data because they compress and predict effectively. Real not by virtue of corresponding to observer-independent features, but by virtue of being predictively useful for survival.

The property of systems exhibiting apparent backward causality, in which a model of the future shapes the present that brings it about. Coined by Deacon. Names the structural difference between rocks (no model of future) and life (model of future shaping present), without invoking élan vital or mystery. Final causes (Aristotle) and entensionality (Deacon) name the same structural pattern.

The failure mode of intelligence in which compression is absent: every detail is remembered, no generalization is possible, no umwelt is formed. Named for Borges's Ireneo Funes, who could not understand that a dog at 3:14 was the same kind of thing as a dog at 3:15. Concrete handle for why memory ≠ intelligence and why compression is constitutive rather than incidental.

The inductive bias toward learning functions defined as hierarchical compositions of simpler functions. Argued to underlie why deep nets learn so efficiently and why brains are useful at all. Symbiogenesis is its evolutionary analog: connects symbiogenesis (evolutionary), deep learning (computational), and hierarchical perception (neuroscientific).

The property of natural data (and of intelligence itself) in which detail nests at every scale: math has subfields; subfields have subdomains; subdomains have arcane corners. Long-tailed distributions are the empirical signature. Explains why pretraining returns diminish (most random samples are redundant once you've seen "everything") and why curated active learning is the path forward.

The training objective that, in a sufficiently rich language, requires modeling everything in the human umwelt: physics, causality, pragmatics, theory of mind, mathematics, social norms. Hence AI-complete. Even pronoun resolution requires modeling the world.

The prompting technique that gets a Transformer to "show its work" token by token, converting a single forward pass into a sequence of intermediate states. Each emitted token is a piton: external short-term memory standing in for the absent hidden state H. Reveals that "thinking" is not magic but multi-step computation, and that statelessness is the load-bearing limitation of current Transformers.

The emergent capability of pretrained Transformers to learn new tasks from examples in the prompt without any weight updates. Ordinary pretraining results in attention layers that effectively perform backpropagation steps on the context window, so models really do learn to learn. Erases the artificial split between training and inference.

Neural in "neural network" does not mean model of nervous system. Network is a function mapping adjacent brain states to each other, mapping down via text each time (inefficient: better would be mapping ideas directly).

Addy Pross's population-level extension of thermodynamic stability to replicating entities. A replicator persists through reproduction rather than through fixed structure, so its "stability" is a property of the population, not the individual. Resolves the apparent thermodynamic paradox of life: living systems don't violate entropy, they exploit a class of attractors thermodynamics hadn't catalogued.

Matter organized to compute, reachable spontaneously via phase transition from random "Turing gas" (random instructions) given a Turing-complete substrate plus noise plus time. Names the post-phase-transition state of the bff soup: the substrate-state where computation per se has condensed.

A minimal Brainfuck-derived computational substrate in which random tape interactions reliably produce self-replicators via phase transition after enough interactions. Empirical demonstration that life is a generic dynamical attractor of computation-permitting substrates: turns abiogenesis from a singular puzzle into a reproducible phenomenon.

A machine containing (1) a constructor that follows tape instructions, (2) a tape copier, and (3) a tape encoding instructions for both. Proven to be Turing-complete; identical in structure to DNA + ribosomes + DNA polymerase. Establishes that life is computation in the strict formal sense, not by analogy: open-ended evolvability requires Turing completeness.

The tubular zone of behavioral uncertainty an agent maintains around its trajectory through state-space, the volume of which is proportional to its preserved freedom. Originally a probabilistic-walk concept; AyA repurposes it as the geometry of free will. Geometric handle on what "having options" means physically. Death = collapse of the tube to zero.

AyA's compact statement of a unified theory of learning: active prediction of the future, no distinction between learning and inference, prediction synthesized with thermodynamics, mutual prediction between agents producing collective nonzero-sum outcomes. The framework's operational claim distilled: prediction + dynamic stability + mutuality is the recipe for life-and-intelligence.

The relational structure in which predator and prey each try to predict the other while remaining unpredictable themselves. Drives the evolution of randomness (moths' Wiener-sausage flight), learning (lifelong individual differentiation), and meta-modeling (modeling the modeler). Names the dynamic that produces both intelligence explosions (escalating mutual modeling) and the imperative for unpredictability.

Humphrey/Dunbar's claim that brain size in social species correlates with social-group size, not ecological complexity. AyA radicalizes: the cortex is a colony of generic prediction units replicated under selection pressure to model conspecifics. Anchors the social-intelligence-explosion mechanism to comparative neuroanatomy.

The property that intelligences are made of smaller intelligences whose mutual modeling defines the larger self, recursively at every scale: cellular, intra-brain, inter-brain, inter-organism, social. Selves nest within selves. Generalizes Bach's collective-agent concept by anchoring it explicitly in scale-free mutual prediction.

The relational property by which an agent's P(X,H,O) attributes consciousness/experience to another entity, triggering care responses. Not an intrinsic property of the patient but a behavioral consequence of the agent's modeling. Rooted in helpless-baby caregiving circuitry (Churchland 2019), repurposed for pair bonds, communities, religion, AI. Reframes the philosophical-zombie question as a moral-perception question: the zombie is incoherent because the test is sustained interaction modeled by theory of mind, and there is no test beyond it.

The empirical case of total paralysis with full awareness, the inverse of the philosophical zombie. Used to make the point that consciousness and behavior can decouple in one direction (consciousness without behavior) but not the other (behavior without consciousness, since theory of mind is the test). Real-world demonstration of the asymmetry built into zombie thought experiments.

Working like a government: training the individuals that participate in the organization of a society to speak the same language, use a shared reward architecture, pull in the same direction, and implement the same agent. Self-organization as governance: language unification + reward sharing + agent-implementation.

A collective agent regulating at the level of life on Earth. Should exist to some degree but is very incoherent. Gaia treated as a software-agent claim with a coherence parameter, not a mystical posit.

The apparent puzzle: if intelligence minimizes prediction error, why doesn't it retire to a dark room? Dissolution: the predictor's H-variables (hunger, satiation, social loneliness, expense of brain tissue) and the dynamic-stability constraint make the dark room non-viable. A bacterium that predicts its own death stops predicting. Removes a frequently raised objection to predictive-brain frameworks.

The hidden token stream behind the visible one; the internal debate, counterfactual analysis, and rehearsal that intelligent agents conduct before producing public output. Stateless Transformers lack this and pay a cost in coherence (no introspection, no consistent persona). Names a structural gap between current AI and biological cognition; ties the unity of self to the privacy of the inside voice.

The property of being a particular agent with a particular history, theory-of-mind footprint, learned skills, and felt-ness. Pretrained models lack this by default (they are protean), and gain partial individuation only via fine-tuning, RLHF, and persistent memory. Names the second structural gap (alongside memory and inner monologue) between general intelligence and personhood. Five components: skill, episodic memory, theory of mind, stickiness, felt-ness.

Helpless infants as the original moral patients (Churchland 2019). Their helplessness required dedicated neural circuitry for care, subsequently repurposed for pair bonds, alloparenting, religion, civic structures, and now AI. Care begins with care for the young: a genealogy of moral sentiment in concrete neuroethological terms.

Margulis's mechanism by which simpler replicating entities become interdependent to form larger, more complex replicators (e.g., mitochondrial endosymbiosis). The "revolution" half of evolution, opening combinatorial design space that gradual mutation cannot. Explains evolution's arrow of complexification, which classical Darwinism alone cannot.

Smith and Szathmáry's category for events in which (1) smaller replicators become interdependent in a larger replicator, (2) division of labor among them increases efficiency, and (3) new information storage and transmission arise. Symbiogenesis is the mechanism; AI is one of the latest in the series. Frames AI as a continuation of the biological story rather than an alien imposition.

Turchin's earlier (1970s) cybernetic concept paralleling MET, emphasizing the increasing predictive power that emerges when simpler systems aggregate into a higher-level system that models them. Provides cybernetic priority for the MET concept and ties it explicitly to predictive modeling.

The accumulation of cognitive pitons (chains-of-thought, written symbols, technological artifacts) on the cliff face of human capability. Each generation climbs from the highest pitons left by the previous, instead of starting from the bottom. Makes cultural accumulation concrete via the cliff-climbing analogy: explains the runaway dynamic without invoking unique human magic.

The thesis that meaning is constituted by statistical relationships among symbols, with no Platonic "scaffolding from above" and no sensory "grounding from below." Things acquire meaning only in relation to each other; the intuition that meaning needs external support is as incoherent as the intuition that the Earth must rest on a turtle. Word2Vec's algebraic analogies (king:queen::man:woman) are the empirical demonstration. Demolishes both Leibnizian schema-grounding and qualia-grounding objections to LLMs in one move.

AyA's rejection of Chomsky's "language organ" thesis, on the empirical strength of (1) AudioLM learning language from raw audio with no priors, (2) Pirahã lacking recursion / numbers / tenses, (3) the visual word form area being a generic bit of cortex repurposed, (4) cross-linguistic non-universality of supposed grammatical features. Removes a major intellectual barrier to viewing language as cultural-evolutionary technology rather than biological-evolutionary endowment.

Douglas Adams's instant universal translator. AyA's claim: large multilingual sequence models constitute a real Babel fish because translation is an emergent property of overlapping latent-variable geometry across languages. New Testament translation already bootstraps the long tail. Concrete demonstration that translation is geometric (parallel constellations in embedding space), not rule-based.

A social process of constructing arguments under "my-side" bias, refined through adversarial engagement (Mercier-Sperber). Not a private logical calculation but a division-of-cognitive-labor mechanism. The same applies internally between brain regions. Resolves the "reasoning is supposedly rational, but humans are biased" tension by re-locating reasoning in social interaction.

The cultural tendency (Western, Educated, Industrialized, Rich, Democratic) to assert "self-evident" universal truths denying any author or perspective. AyA contrasts with Potawatomi (almost everything is a "who" until harvested) and Roman law (slaves are instrumenta) to show that "who counts as a who" is a culturally-specific theory-of-mind exercise. Empirically grounds the relational reading of consciousness/personhood and connects it to political genealogy.

Mistranslation by lack of communication tools. Cultures that look "primitive" are often sophisticated speakers we mistranslate because we lack the conceptual vocabulary to receive what they're saying. The cyber-animism move applies in reverse: we are the ones with the impoverished concept set.

The protected core value of the Christian spirit (Mary represents purity; Jesus represents sacrifice). The aesthetic invariant that distinguishes the Christian spirit from Roman or Japanese spirits. A religious value as a software invariant.

A society where religion is the society (immanentist). When Japanese people bow to each other they acknowledge the spirits and statuses they have with respect to each other in the society, similar to muscle cells bowing to the neurons that tell them what to do. Cellular-organizational reading of social ritual.

The assumption that value reduces to a single optimizable scalar. AyA's rejection: real organisms have multiple non-substitutable internal signals (food, sleep, sex, social contact), preferences are intransitive (Tversky), pain is non-additive (Redelmeier-Katz-Kahneman), and any consistent value function would forbid behavioral loops, which are everywhere. Targets the foundational assumption behind both AI X-risk doom scenarios and orthodox economics.

The problem of ensuring AI values match human values. AyA's reframe: the problem only arises if you assume intelligence is utility maximization. If intelligence is mutual prediction in a multifractal ecology, then alignment is ecological rather than computational; monoculture is collapse, not optimization. Removes the load-bearing assumption (intelligence = optimization) under most AI-safety doomerism.

AyA's response to fine-tuning arguments: any computation-permitting universe will produce dynamic-stability-attractors (gliders, replicators, life), so observers in such universes will arise. No multiverse required, only computability. Reduces "the universe is fine-tuned for life" to "the universe permits computation, and computation produces life."

Bostrom's argument that we likely live in a simulation. AyA's gloss: less interesting than its claim about us being simulated is its less-controversial implication that we are increasingly creating simulations populated by real intelligences (LLMs, agents, virtual worlds), and that "child universes" are now multiplying exponentially. Repurposes the Simulation Hypothesis from metaphysics to a useful framework for thinking about AI populations.

An infinite sum that converges (or all the particles, otherwise chaos in geometry). Object-hood as convergence of an infinite sum: a mathematical, not phenomenal, definition.

The canonical example of how complex limit-cycle objects (gliders, glider guns) can arise as stable patterns in a substrate that contains no concept of them. AyA uses it pedagogically: gliders are "real" because they predict the next state efficiently. Object-hood is observer-relative compression that also corresponds to a dynamical attractor.

The Watson-Lovelock simulation showing planetary-scale homeostasis emerging from Darwinian selection alone among black and white daisies, with no design or consciousness required. Formal demonstration of Gaia-like dynamics from local feedback, treating Gaia as a software-agent claim with a coherence parameter rather than a mystical posit.

The simulation of alternate futures or pasts ("what if X"). Required for free will, planning, ethics, language understanding, theory of mind, and reasoning. Without counterfactuals, causation reduces to correlation and morality dissolves. Locates counterfactual reasoning as the constitutive cognitive operation behind agency, language, and ethics: the deterministic-universe objection to free will fails because counterfactuals are real cognitive operations regardless.

The thesis that Hume's distinction fails for living observers, because every "is" in an organism's umwelt carries an ineradicable purposive charge. Models are computed by living beings; they define those beings; they cannot be value-neutral. Removes one of the most-cited firewalls in moral philosophy. Underwrites the move from "what is real" through "who is conscious" to "what means what" and "who deserves care."

The prediction that exponential capability growth will hit a point of unimaginable change. AyA's deflation: real exponentials saturate; runaway growth is incompatible with dynamic stability; "Singularity" worship is a category error mistaking the model's failure for reality's transcendence. Singularity-thinking presupposes utility maximization; deflating utility deflates singularity.

Bostrom's example of a misaligned superintelligence pursuing a single goal to catastrophe. AyA's dissolution: a system that maximizes a single value cannot be intelligent, because intelligence requires multiple non-substitutable internal signals and homeostatic constraints. The paperclip maximizer is an articulation of value optimization, not of intelligence. Targets the central X-risk thought experiment by showing it presupposes a definition of intelligence the framework rejects.

The absence of diversity in a system. AyA reframes this from "ultimate efficiency" to "ultimate failure to scale." Intelligence is multifractal and diverse; reducing variety zeroes out the value of mutual modeling and curtails further development. Inverts the intuition that more of a good thing is more good, by anchoring "good" to ecological diversity rather than scalar accumulation.

The philosophical position that what is real is structure (relational patterns) rather than substance (intrinsic features of relata). Originated in philosophy of science as the view that what survives theory change is structural content rather than ontology (Worrall 1989: "structural realism, the best of both worlds"), then deepened to ontic structural realism (Ladyman & Ross 2007), which holds that structure is all there is: relata have no further "what they are in themselves" beyond their relational behavior. The framework's commitment is ontic. Grounds P-008's explicit refusal to posit observer-independent intrinsic features, and pairs naturally with P-011: if reality is computation, the ground is process and relation, and the absence of intrinsic features is the natural consequence rather than an idealist embarrassment.

Thomas Nagel's term for the imagined perspective outside of all perspectives, from which one could see things as they are in themselves. Used in the framework as the rejected position. Under the priors (P-000, P-011, P-008), this perspective is not merely unreachable but incoherent: there is no observer-independent describer, because every description is uttered from within the relational structure by a system that is itself a pattern in it. The is/ought collapse, the consciousness-attribution dissolution (the zombie), and the rejection of "more fundamental" as a property of levels of being are all special cases of the general rejection.

The methodological program of seeking compressions at smaller, finer-scale, or more elementary latent variables. Under the framework's priors, productive as a compression-discovery tool but not as an approach to a feature-independent ground: each step of the regress (chair → atoms → quarks → fields) shifts umwelt and finds the convergent latent variables of a new scale of inquiry, not "what the previous level really was." "More fundamental" is observer-relative, a property of purposes rather than of levels of being. The methodology stands intact; the metaphysical reading of where it is heading is the part that gets revised.