State-Space Semantics

In the [YOU] on AI Field Guide

The cycle’s confrontation with large language models is, among other things, a confrontation with the discovery that meaning can be geometric. When a language model answers a question about history or writes a poem or explains a scientific concept, it does so by transforming vectors through a learned space in which the relationships among ideas are encoded as distances and directions. This is precisely Churchland’s state-space semantics running at planetary scale, and understanding it reframes the question the cycle asks: not whether the machine “really understands” in the folk-psychological sense, but whether geometric relationships in a high-dimensional activation space can constitute understanding—and if so, what kind.

State-space semantics also provides the cycle with its most precise account of why modern AI learns the way it does. The machine is not programmed with the facts it comes to know; it is exposed to vast quantities of data, and a training procedure nudges its billions of weights, gradually shaping its activation spaces until they capture the structure of the domain. No propositions are inserted. The knowledge that emerges lives in the geometry of the trained network, precisely where Churchland said human knowledge lives. The cycle’s central observation—that AI is a fundamentally different kind of intelligence from the rule-following, checklist-executing image that most people carry—is, at its deepest level, an observation about what state-space semantics means for how these systems work.

Origin

Churchland developed state-space semantics most fully in A Neurocomputational Perspective (1989) and The Engine of Reason, the Seat of the Soul (1995), building on the color-perception case he used repeatedly: the human visual system has three types of cone, each sensitive to a different band of wavelengths, so any color corresponds to a triple of activation levels across those channels—a point in a three-dimensional activation space. The familiar color solid is not a metaphor for color but, on Churchland’s account, the literal shape of the activation space of the visual system given geometric form. He treated this as a template for all cognition: faces, words, sounds, moral situations—everything is a point in some learned activation space, and knowledge is the sculpted structure of that space.

The thesis was developed in sharp opposition to the symbolic AI tradition and to the language-of-thought hypothesis associated with Jerry Fodor—both of which held that cognition requires sentence-like representations manipulated by rules. Churchland argued from neuroscience that brains simply are not built this way, and from philosophy that the sentence-based picture generates the infinite regress problem for learning: if learning is adding sentences to a store, the capacity to handle sentence-like representations must already be in place, which cannot itself have been learned by the same process. State-space learning—the continuous adjustment of connection weights—faces no such regress.

Key Ideas

Meaning as position, not predication. In folk psychology and symbolic AI, meaning is propositional: to mean something is to predicate a property of a subject, to stand in a logical relation to a sentence-like content. In state-space semantics, meaning is positional: to mean something is to occupy a location in a structured space in which similar meanings are nearby. Two representations are similar to the degree that their corresponding points are close, and the brain computes by moving from one location to another along learned trajectories through the space.

Prototype concepts and fuzzy categories. If concepts are regions of activation space rather than definitions, then category membership is a matter of proximity to a prototype rather than satisfaction of necessary and sufficient conditions. This explains why some cats are more cat-like than others, why category boundaries are fuzzy, why we handle novel instances by analogy rather than by deduction, and why expertise produces the perception of distinctions invisible to novices—the expert has sculpted a space with finer-grained structure in the relevant region. Modern neural networks develop exactly this prototype structure spontaneously, without being given definitions.

Learning as space-sculpting. On the state-space account, learning is not the accumulation of propositions but the continuous reshaping of a network’s activation space through weight adjustment. Experience changes the geometry of the space until similar things are represented similarly and the space supports good generalization. A newborn’s neural spaces are largely undifferentiated; a radiologist’s activation space for medical images has been sculpted by thousands of cases into a structure where subtle pathologies stand out as distinct regions. This is the model of learning on which all modern machine learning operates, without the folk-psychological furniture that competing theories required.

The convergence with AI embeddings. The word embeddings used by large language models—high-dimensional vectors in which words with similar meanings have nearby representations and semantic relationships correspond to geometric directions—are state-space semantics implemented in silicon. That the same representational strategy is used by the brain (on Churchland’s account) and by the most capable AI systems (by engineering necessity) is not a coincidence but a deep fact about what any system must do to represent the rich relational structure of the world efficiently.