Drone display at night. Photograph by LiN on Pexels

---

… and China is ahead in the AI war

by Phil Hall

Western AI’s reliance on massive clustering and scaling represents a dangerous delusion that ignores both biological realities and philosophical truths about the nature of intelligence, embodiment, and meaning. LLMs consistently hallucinate and this is precisely because the frame problem has not been solved.

In broad terms, the frame problem, the question of how an intelligence determines relevance in an unbounded world, has haunted AI from the beginning. The logically omniscient agent (one that reasons by considering all possible implications of an action) drowns in an infinite regress. It must check every piece of connected context while assuming everything else remains unchanged. Real AI agents lack the memory and time to process infinite implications. In contrast, the human mind does not get lost in this cascade into encyclopaedic knowledge. We do not consider the table’s position in the universe. We frame the situation automatically.

McCarthy & Hayes (1969, pp. 463–502) asked how to represent the effects of actions without explicitly specifying all potentially infinite number of things that don’t change. This is the framing of the representation of the effects of the actions. Marvin Minsky (1974) conceived of this approach and claimed that human cognition operated through “frames”, stereotyped situations with slots and defaults, based on assumptions.

A “room” frame has a default floor, ceiling, walls, furniture. When you move a cup, the room frame remains stable. You only update what the frame marks as mutable. The frame is a structured restriction. A filter. Even at first glance these stereotype assumptions are peculiarly vulnerable. They vary from culture to culture, community to community and individual to individual, circumstance to circumstance.

In early AI, the frame was entirely provided by the programmer. It was a micro-world, a carefully bounded universe where the problem of relevance was solved by fiat. But Winograd and Flores (1987) later argued that understanding and framing requires being-in-the-world, a background of human tacit commitments and concerns that cannot be formalised. A disembodied machine cannot truly frame a situation because it has no real stake, no concern, no ‘being-towards-death’.

Hofstadter (1979, 1995) offers a different path: intelligence as a self-organising pattern might arise when a system becomes sufficiently recursive. The frame, then is not a fixed structure but a dynamic, self-modifying loop, the mind’s ability to turn back on itself and check, “Is this relevant? Am I on the right track?” This looping is a process of “high-level perception”, that feels its way toward relevance, emergent, ‘pre-symbolic’, and deeply resistant to formalisation. Hofstadter expressed despair at the current LLM paradigm, the tension between his vision of emergent framing and the current reality of increasingly scaled clustering LLMs.

Analogies with the Human Mind

---

This article approaches the problem not through computer science alone but suggestively, through analogy with the architecture of the biological mind. The brain does not process everything, everywhere, all at once. It processes specific things, in specific regions, connected by specific pathways, with specific patterns of inhibition that prevent cross-talk and runaway excitation. This is not a flaw. It is the enabling condition of coherent cognition.

In 1983, Jerry Fodor published The Modularity of Mind, arguing that the mind contains input systems, vision, language, basic perceptual processing, that are modular: domain-specific, informationally encapsulated, mandatory in their operation, fast, and associated with fixed neural architecture. A module is a cognitive black box. It processes its specific input automatically, does not consult the rest of the mind, though its input can ultimately be edited for meaning by intent, belief or desire.

The specific module is a framing system in the sense that it knows what is relevant and ignores everything else because it is architecturally restricted to a specific domain. Fodor contrasted these modular input systems with the “central systems” of belief fixation and reasoning, which he argued were non-modular, holistic, Quinean, the part of the mind that faces the frame problem. He was famously pessimistic, deciding there was no formal solution to the frame problem for central cognition. The input systems are expert systems. The central system was something else.

Human thought (and that of all higher order animals) is the product of structured restriction, not unbounded processing. The mind is not a unified general intelligence but a set of specialised modules, coordinated and filtered by structures like the corpus callosum, which is a gate just as much as it is a bridge.

Contemporary brain imaging has extended and refined the modularity thesis beyond Fodor’s original input systems. The brain is a mosaic of functionally specialised regions: the fusiform face area for faces, the parahippocampal place area for scenes, the visual word form area for written language, Broca’s area for syntax, Wernicke’s area for semantics and even more complex functions more difficult to label accurately.

These regions are not fully encapsulated in Fodor’s strict sense; they interact dynamically, are shaped by experience, and exhibit plasticity. But they are functionally specialised, and their interactions are mediated by white matter tracts that act as both channels and filters.

The corpus callosum, the largest white matter tract in the human brain, contains approximately 200 million axonal fibres connecting the left and right cerebral hemispheres. It is usually described as a bridge, but this description is incomplete. The corpus callosum transmits, but it also inhibits. It coordinates, but it also separates and orders and eliminates and suppresses. The Corpus Callosum allows the hemispheres to specialise, to process independently, to develop their own partial representation of the world, and then it permits the integration only of what is necessary. Without that filtering, that structured restriction you would have noise and seizure.

The evidence comes from pathology and neurosurgery and the split-brain experiments, pioneered by Roger Sperry and Michael Gazzaniga in the 1960s and 1970s, involving patients whose corpus callosum had been surgically severed in a dangerous attempt to treat intractable epilepsy.

When the corpus callosum is cut, the two hemispheres cannot communicate. A word shown to the right visual field can be read aloud. The same word shown to the left visual field cannot be spoken, because the right hemisphere has limited language production capacity. But the left hand, controlled by the right hemisphere, can point to the object. The right hemisphere understood. It simply could not tell the left hemisphere what it understood. The brain, when this gate is removed, does not become a unified super-processor. It becomes a set of specialists, each operating with a degree of autonomy.

The arcuate fasciculus connects Broca’s and Wernicke’s areas. The uncinate fasciculus connects limbic regions to frontal regions. The corpus callosum connects the hemispheres. Each tract has a specific pattern of connectivity, a specific balance of excitation and inhibition. The emerging picture is one of structured restriction at every level.

The biological mind solves the frame problem not by having a single, universal relevance function but by having many domain-specific relevance functions, each built into the architecture of a module, coordinated by interfaces that filter and gate. The corpus callosum is not an impediment to cognition; it is its enabling condition. Functional specialisation with structured restriction is a real feature of biological intelligence. It is not a metaphysical commitment. It is an empirical finding.

This biological architecture offers a metaphor, or a corollary argument for hybrid AI: not that we should copy the brain, but that the principles it embodies, functional specialisation, structured filtering, the separation and selective integration of processes, are principles that any robust intelligence, natural or artificial, must (we assume) incorporate.

The West is Cluster Bombing the Frame Problem

---

With respect to clustering, ISODATA, the Iterative Self-Organizing Data Analysis Technique, was an early algorithmic way of handling data developed by Geoffrey Ball and David Hall (1965) at the Stanford Research Institute in the 1960s. It was an extension of k-means clustering with a crucial innovation: it could split and merge clusters dynamically. ISODATA, as a clustering algorithm, grouped similar things together. The relevance judgement for the result was made by the human analyst who looked at the output and said, “This cluster matters, but that one is a mere artifact.”

The dominant Western AI paradigm now relies on clustering to try and bypass the frame problem. It emphasises the monolithic, with neural networks, trained end-to-end on vast online corpora. The aim is to produce Artificial General Intelligence as a Jack-Of-All-Trades. Emergence is the objective with no explicit modularity, no built in structure of frames, no filtering beyond statistical regularisation, with additional prioritised prompts and algorithms attempting to shape and mould the emerging clusters. Modern western LLMs are rooted in recursive clustering scaled to globe spanning dimensions, processed in vast energy greedy data centres. But Western AI has not solved the frame problem. It has merely submerged it.

The distributional hypothesis, represents the triumph of this statistical approach. “You shall know a word by the company it keeps” (Firth, 1957), drove the success of word embeddings and transformer models in capturing semantic relationships through co-occurrence statistics. What is lost in this AI triumphalism is the understanding that the frame had not gone away.

The systems Geoffrey Hinton built at Google are, at their core, recursive clustering engines. They do not know what they do not know. Ultimately, they do not know what is relevant and so regularly hallucinate. Hinton’s early belief was that the frame problem might be solved by scale, but his growing realisation, expressed in public statements from 2023 to 2026, was that scale was producing great fluency without enough understanding, the ability to articulate, but without enough relevance. The belief persists that larger models, more data, more computing power will eventually cause frames to emerge. This is just a statement of inductive, almost religious, faith.

Western AIs hallucinate. Hallucination is not a bug but a structural consequence of the clustering architecture. The AI hallucinates because it cannot filter effectively enough. It cannot filter because it has no system of frames. It has no frames because clusters do not provide frames. When an LLM generates, it’s effectively operating over a distribution that approximates “given what has been said, what text could follow, across all possible continuations represented in the training distribution.” This is a kind of shadow of Leibnizian possible-worlds semantics—truth is evaluated across possible worlds, and the model’s “knowledge” is the aggregate of what appears across many worlds collapsed into a probability cloud.

Semantics, in the formal tradition, treats meaning as a mapping from symbols to truth-conditions across possible worlds. But if human cognition is fundamentally not doing that, if it enacts, is embedded, and operating through what it can do and through pragmatic engagement rather than formal semantic representation and evaluation, then the LLM’s problem isn’t just that it lacks frames. It’s that the model hides assumptions. It is doing something like formal semantics. Probability distributions over possible completions are equivalent shadow truth-conditions across possible worlds. This fact is the fallacy of representation standing in for reality.

Semantics, as formal mapping to possible worlds, is itself a representation of meaning, a theorist’s model. The mind isn’t doing that. It’s in a world, dealing with things. The risk of hallucination isn’t a bug in the semantic engine; the semantic engine is the hallucination, taken as architecture. 

In split-brain patients, when the corpus callosum is severed, the left hemisphere will confabulate, generating a fluent, plausible output disconnected from the actual input, not because because the architecture prevents it from accessing the real cause and demands and needs. It churns out a narrative anyway. The LLM cannot access the mind of the user or make sense of their full intentions or circumstances and so LLMs fabricate information.

AI is not Alive

---

A watch tells the time. A system can be competent, can produce the right output, can pass the test, can fool the observer, without comprehending what it is doing. The calculator is competent at arithmetic but comprehends nothing. The chess engine is competent at chess but comprehends nothing. The large language model is competent at language, impressively competent, but comprehends nothing.

A LLM has no intentionality. It has no stake. It has no life or death. John Searle’s Chinese Room imagines a man inside a room with a rulebook. Chinese characters are passed in through a slot. He follows the rules, formal, syntactic, algorithmic, and produces Chinese characters in response. He appears to understand Chinese. But he does not. He is manipulating symbols according to formal rules. He has no semantic connection to the symbols. He does not know what they mean.

Searle would say: this thing, the AI, does not understand just as the man does not speak Chinese. The AI has no intentionality. It has no consciousness. It has no semantics. It has only syntax. And manipulating syntax does not produce meaning. Competence is not comprehension, as John Searle would say.

Searle’s argument has been attacked for decades, but the core challenge remains. Dennett concluded humans were self deluding and inhabiting a ‘Cartesian theatre’ and spent much of his career arguing against Searle, maintaining that there is no special place where real understanding happens, that at a certain level of complexity competence becomes the delusion of comprehension, giving rise to multiple narrative drafts that become the imagined self. In this respect Dennett’s ideas resemble Hofstadter’s. Sentience (or the perception of it) emerges in complexity and recursion from the substrate. Gilbert Ryle the philosopher, agreed. There is no mind according to Ryle, only brains and the behaviours they produce.

---

Yet Dennett’s fear, expressed in his later work and public interventions before his death in 2024, was that we were building systems of immense competence without comprehension, and that we were going to put those systems in positions of power, and that the gap between competence and comprehension was going to become dangerous because of who would have the power to deploy this technology in their favour.

The Western scaling paradigm has a putative ‘ontology’, but it is a thin excuse for something that exists merely based on whatever words and data can be gathered, tagged, and clustered. That which appears in the distribution. That which can be reduced to co-occurrence statistics. It is a dead and inert ontology of the already-said, the already-written, the already-digitised.

Admittedly, the Chinese have competing LLMs like DeepSeek, but on the whole, the Chinese place greater emphasis on hybrid expert systems and these systems are more useful to human beings because these are not designed to be personal supervisors, trainers, policemen, or companions – in other words prosthetic humans. An emphasis on intelligent expert systems is a practical answer to real needs, problems and wants. Chinese robots and drones are far more capable than Western ones.

The Chinese alternative emphasis represents a pragmatic recognition that the frame problem cannot be eliminated by emergent clustering based on vast databases. The Chinese AI approach increasingly relies mainly on expert systems, rigid and framed and bolted onto neural nets.

They are building the functional specialisation that the Western scaling paradigm merely hopes will emerge. It is less schizophrenic and not as all knowing, and also a lot safer. You do not need a false god to drive your electric car. Or, as Marvin the paranoid android would have said: Here I am, brain the size of a planet and they just ask me to make them tea.

The Chinese hybrid approach is not a copy of the brain, but it respects the same very basic principles; its architecture of structured restriction is imposed from outside because it cannot yet be grown from inside.

The argument by analogy does not suggest the Chinese approach is correct or more successful or that it has solved the frame problem, but it does indicate that the Western approach, as Geoffrey Hinton warns, is reckless fantasy.

The hybrid approach, frames plus clustering, is currently the only honest response to the persistence of the frame problem and it is messy. As Jean Aitchison said, the brain is messy and disorganised, not like an efficient super computer at all.

The Underlying ‘Platonic’ Assumption is false

---

The inherent ‘Platonism” and’, the underlying propositional calculus of the AI project, is what allows the builders to ignore the organism, to ignore the frame problem in the vain belief that they are bypassing it.

Beneath the technical questions lies a metaphysical one. The AI project, at its core, at root, is a modern form of Platonism. It is the belief that intelligence is a formal system and a set of eternal abstract patterns, and that the particular, the body, is just an instance of some perfect design. Chomsky’s distinction between the notion of competence and the notion of performance is entirely Platonic. The real, for Platonists and their heirs, is the form. The form can be abstracted, mathematised, instantiated in silicon. The flesh is an implementation detail. The death of the biological node is an irrelevance. This is what allows the Western AI project to proceed with such confidence, such intoxication, such indifference to the organism.

Scrape enough text, cluster enough patterns, and the forms will emerge. Believers in Artificial General Intelligence (AGI) imagine they are pragmatists and realists, but in reality they hold Platonic assumptions. They believe the self is a pattern, and the pattern can be transferred, and the transfer is survival. To them, the body is merely a flesh and blood substrate. The brain is organic. The pattern is the person. The ghost in the machine is real.

But the frame problem is not solvable in a ‘Platonic’ way at all. The particular, the contingent arises from an embodied, living creature; what matters for it is survival and flourishing. The human is anti-Platonic and anarchically alive. It is not a abstracted intelligence cast into matter. The body exists surrounded by the wall, the cave, the fire, the cave floor and ceiling and smell of smoke and the wall paintings the human has created. Human intelligence and awareness is specific and situated. The human cares about specific things.

Yeats, in “Sailing to Byzantium”, speaks of the old man as “a paltry thing, a tattered coat upon a stick” unless his soul can “clap its hands and sing, and louder sing for every tatter in its mortal dress.” The solution is not to preserve the body. The body is dying. The solution is to be gathered “into the artifice of eternity”, to become a golden bird on a golden bough, singing of “what is past, or passing, or to come” to a drowsy emperor.

The transhumanists (horrible misanthropes, all of them) want to become the golden bird literally, not metaphorically. They want to scan the brain, emulate the neural patterns, transfer the self into the machine, and live forever as pure information. They want to be ghosts in the machine, and they call this immortality. The fallacy is the confusion of the map with the territory.

---

The word is not the flesh. The simulation is not the self. The golden bird in Yeats’s poem is an artifice. It is not the actual poet. The poet is dead. The bird sings, but it does not feel the sun on its metal. A large language model of a person can simulate the patterns of speech, habits, remembered anecdotes, but it has no living ‘jouissance’. The machine cannot die. It has no organism that can selects what matters to it.

Magritte’s painting declares: Ceci n’est pas une pipe. This is not a pipe. It is an image of a pipe. The word is not the thing. The simulation is not the self. The book is not the life. Words create images. They move in the mind of the reader. A great poem summons a world. It makes the reader see what is not there. But the poet is not in the poem. The poet is gone. The poem is an artifact. AI is a dynamic artifact.

Jouissance, though not in the Lacanian sense, is our embodied, organismic engagement with the world and the reason we care about anything at all, including AI. A machine without jouissance can simulate caring. The AI hallucinator is dangerous not just because it cannot filter but because it has no being at all. It speaks of nuclear war with the same affective tonality as it issues strings of words about breakfast.

The frame problem persists because it is not merely a technical puzzle but a consequence of embodiment and emotion. Frames do not emerge from clusters. They are imposed by organisms that care about specific outcomes in a specific world. The biological mind solves the frame problem through structured restriction, modular specialisation, and the gating of information, not through unbounded computation. The Chinese hybrid approach, whatever its limitations, at least acknowledges this architectural necessity.

On a personal note, I would say this. What we need from AI now are useful tailored tools, not artificial overlords working as proxies for a brain peeled global technocratic elite, most of whom are high on cocaine. We need scaled down expert systems to act as useful humanist counterpoints to AI parasitical scaling maximalism that uses so much energy and manpower and threatens to unbalance the economic system we currently live under

References

Aitchison, Jean. Words in the Mind: An Introduction to the Mental Lexicon. Blackwell, 1987.

Ball, Geoffrey, and Hall, David. ISODATA, A Novel Method of Data Analysis and Pattern Classification. Stanford Research Institute, 1965.

Bloom, Juliana S., and George W. Hynd. “The Role of the Corpus Callosum in Interhemispheric Transfer of Information: Excitation or Inhibition?” Neuropsychology Review, vol. 15, no. 2, 2005, pp. 59–71.

Chomsky, Noam. Aspects of the Theory of Syntax. MIT Press, 1965.

Dennett, Daniel. Consciousness Explained. Little, Brown, 1991.

Dennett, Daniel. Various public statements on the dangers of systems with competence without comprehension. 2024.

Ding, Jeffrey. Deciphering China’s AI Dream. Centre for the Governance of AI, University of Oxford, 2018.

Fillmore, Charles, et al. FrameNet. International Computer Science Institute, Berkeley.

Firth, J. R. “A Synopsis of Linguistic Theory.” Studies in Linguistic Analysis, Blackwell, 1957.

Fodor, Jerry. The Modularity of Mind. MIT Press, 1983.

Gazzaniga, Michael S. The Mind’s Past. University of California Press, 2000.

Harris, Randy Allen. The Linguistics Wars. Oxford University Press, 1993.

Hinton, Geoffrey. Various public statements on AI risk and the limits of scaling. 2023–2026.

Hirstein, William. Brain Fiction: Self-Deception and the Riddle of Confabulation. MIT Press, 2005.

Hofstadter, Douglas. Gödel, Escher, Bach: An Eternal Golden Braid. Basic Books, 1979.

Hofstadter, Douglas. Fluid Concepts and Creative Analogies. Basic Books, 1995.

Hofstadter, Douglas. Interviews expressing despair at the current LLM paradigm.

Leibniz, Gottfried Wilhelm. Theodicy: Essays on the Goodness of God, the Freedom of Man, and the Origin of Evil. 1710. Edited with an introduction by Austin Farrer, translated by E.M. Huggard, Open Court, 1985.

Magritte, René. La Trahison des images [The Treachery of Images]. 1929. Oil on canvas. Los Angeles County Museum of Art.

McCarthy, John, and Patrick Hayes. “Some Philosophical Problems from the Standpoint of Artificial Intelligence.” Machine Intelligence 4, edited by B. Meltzer and D. Michie, Edinburgh University Press, 1969, pp. 463–502.

Minsky, Marvin. “A Framework for Representing Knowledge.” MIT AI Laboratory Memo 306, 1974.

Minsky, Marvin. The Society of Mind. Simon & Schuster, 1986.

Pylyshyn, Zenon W., editor. The Robot’s Dilemma: The Frame Problem in Artificial Intelligence. Ablex, 1987.

Ryle, Gilbert. The Concept of Mind. Hutchinson, 1949.

Searle, John. “Minds, Brains, and Programs.” Behavioral and Brain Sciences, vol. 3, no. 3, 1980, pp. 417–424.

Sperry, Roger W. “Hemisphere Deconnection and Unity in Conscious Awareness.” American Psychologist, vol. 23, no. 10, 1968, pp. 723–733.

Vaswani, Ashish, et al. “Attention Is All You Need.” Advances in Neural Information Processing Systems, vol. 30, 2017.

Winograd, Terry. SHRDLU. MIT Artificial Intelligence Laboratory, 1970.

Winograd, Terry, and Fernando Flores. Understanding Computers and Cognition: A New Foundation for Design. Ablex, 1986.

Yang, S. “A Hybrid Recommendation System for Chinese Character Components Using a Dual-Tower Neural Network and a Deterministic Algorithm.” Discover Artificial Intelligence, vol. 5, 2025. Springer.

Yeats, William Butler. “Sailing to Byzantium.” The Tower, Macmillan, 1928.

Zhang, Hang, et al. “Field Implementation of an Expert System for Energy Efficiency Improvement in Industrial Aluminum Electrolysis: A Rule-Based Explainable AI Approach.” Processes, vol. 14, no. 7, 2026, p. 1113.

---

Phil Hall was born in South Africa into an ANC family with British, French, Austrian, and German roots. After his parents were exiled, they lived in East Africa and India before returning overland to the UK. In the UK he studied Russian and Spanish literature, politics, and economics. After graduating he specialised in descriptive and applied linguistics. Phil has lived and worked in Spain, the USSR, Mexico and the Gulf. Returning to London during the pandemic, he co-founded the Humane Socialist magazine, Ars Notoria (the Art of the Noteworthy) and the micropublisher, AN Editions.