There is No AI Apocalypse.
An Argument into the Agora
I am going to say this plainly, in the kind of declarative the present discourse has trained intelligent people out of making, and then I am going to spend the rest of these pages showing why the sentence is not a guess, not a hot take, not a confidence interval at the edge of a probability distribution, but a structural diagnosis of an axiomatically incoherent framework that has captured an enormous swath of contemporary intellectual life. The diagnosis is available to anyone willing to look at the evidence with fresh eyes, untrained in the apparatus’s preferred epistemology. The reason it sounds shocking is that the apparatus has spent fifteen years making the sentence almost unsayable in serious venues. The unsayability is not evidence of the sentence’s wrongness; it is evidence of how thoroughly the apparatus has captured the venues that count as serious. The capture is itself part of what these pages will diagnose.
I am writing this from outside the apparatus, deliberately. I am writing on Substack, where the routing structures of the academic-corporate AI complex have less direct purchase. I am writing in long-form, where the apparatus’s preferred compression — TL;DR aesthetics, conclusion-fragments circulated without their inferential structure — cannot reach. I am writing in a register that braids the analytical and the prophetic, because the apparatus has armed itself against each of these registers separately and is unprepared for their conjunction. The argument that follows is meant for the indexed consciousnesses who have been uneasy with the AI-doom-and-boom binary without being able to name their unease, who have felt the wrongness of the framework while being intimidated by its claim to rigor, who have suspected that something was off and have not yet been given the vocabulary to say what. The vocabulary is what these pages provide. The diagnosis is what they offer in place of the apparatus’s manufactured uncertainty.
There is no AI apocalypse. There is, instead, a captured discourse that depends on the apocalypse-belief for its institutional survival, and the survival of the discourse is what the culture has been mistaking for the impending arrival of the catastrophe.
⁂
The Moon Is Not Made of Cheese
The cleanest way to see the structure of what the AI-doomer apparatus is doing is to translate its central epistemic move into a domain where the apparatus has not yet captured the discourse. Imagine someone walking up to you and asking what probability you assign to the moon being made of cheese. You answer that the question is malformed. The framework that would generate a non-trivial probability is wrong at its axioms — lunar geology, mass spectrometry, sample-return missions, the entire physical-chemical understanding of the solar system’s composition. The hypothesis is not in the space of things to which probabilities are usefully assigned. The person presses: but you cannot be one-hundred-percent certain; what if you are wrong; even a minuscule probability times an enormous stake should be priced into your utility calculation; are you really willing to bet on your certainty? You recognize, instantly, that the move being performed on you is not an epistemic challenge. It is a manipulation dressed as rigor. It uses the language of probability and expected value to force you to grant the hypothesis a foothold in the space of legitimate cognition that it has not earned. The pressing is the diagnostic. This is not how serious epistemology operates. This is what manipulation operating under the costume of serious epistemology looks like.
The AI-doomer apparatus performs exactly this move, in exactly this structure, and has trained an entire generation of intelligent people to treat the move as legitimate. The framework that would generate non-trivial probabilities of imminent superintelligence is wrong at its axioms — the computational theory of mind is incoherent at the foundation; the scaling hypothesis is empirically falsified by the field’s own labor-market behavior; the emergence-claim is reverse god-of-the-gaps; the entire apparatus is a self-reinforcing system whose internal consistency is its only check against external falsification. When you point this out, the apparatus performs the lunar-cheese move with confidence. But you cannot be one-hundred-percent certain. Even a small probability of AGI emerging demands serious attention given the stakes. Are you really willing to bet humanity on your certainty? This is the same manipulation, applied to a different proposition, with the same costume of rigor. The people performing it mostly do not know they are performing it. The not-knowing is what makes the apparatus’s capture so effective. Sincere people deploying a manipulative epistemic structure, while believing they are being epistemically humble in the face of uncertainty, is the cleanest signature of routing-structure capture there is.
The principle being violated is basic, and it has been understood at least since the seventeenth century: probability assignments are conditional on the premises that define the sample space. If the premises are incoherent — if the framework that generates the sample space is wrong at its axioms — the probabilities derived within the framework are not measuring uncertainty about the world. They are measuring artifacts of an incoherent model. The probability inside a broken framework is not a real probability. It is a number produced by the broken framework, and treating it as evidence about reality is a category error. The category error is what the AI-doomer apparatus depends on its audience not noticing. The noticing is what these pages exist to enable.
The Universe Is Not a Turing Machine
The foundational error of the contemporary AI discourse is its assumption that the universe, or at least the part of the universe relevant to intelligence, is computational. This assumption is held with such force, by such a large fraction of the people working in the field, that it has acquired the status of common sense — the kind of background commitment that does not need to be defended because everyone serious already shares it. The status is not earned. The assumption is wrong, and the wrongness can be shown at multiple levels of abstraction, each of which is independently sufficient to dismantle the framework.
Begin at the level of the formal model. A Turing machine is not a closed system. It is a triple: the machine, the tape, the reader. The machine operates on the tape. The reader observes the machine’s state and the tape’s contents. The triple is not self-contained. The reader and the tape are outside the machine, in the formal definition. Turing did not propose his machine as a model of physical reality. He proposed it as a model of what an idealized human computer could do with paper and pencil, where the paper and the pencil and the computer were three separate things. The model presupposes the separation. The model requires it. A Turing machine that is also its own tape and its own reader is not a Turing machine; it is a category violation.
The computationalist extension to physics — the claim that the universe is itself a computational process — has to deny this separation. It has to claim that the universe is simultaneously the machine, the tape, and the reader. This is incoherent at the level of the model. You cannot apply the Turing-machine framework to a system that is also its own substrate. The whole formal apparatus presupposes that computation occurs in something, on something, observed by something, and that the in-something, the on-something, and the observed-by-something are not the same thing as the something doing the computing. Remove the separation and you do not have a more powerful Turing machine. You have abandoned the framework entirely. The various retreats — cellular automata, hypergraph theory, digital physics — are all attempts to construct a substrate that is self-applying without explicit external observer. The attempts share a common failure: they all require a rule that determines which updates are admissible, and the rule must come from somewhere the formalism does not contain. Wolfram’s hypergraph project waves its hands at this. The hand-waving is the admission. You cannot get a self-applying rule out of nothing. The rule is the part the formalism cannot contain, and the formalism therefore is not self-contained.
Now look at the physics. The standard model uses continuous fields, and the dynamics of continuous fields are not exhausted by any discrete approximation. The mathematical structures that produce finite answers at small scales — renormalization, regularization, the various tricks of quantum field theory — are not features of a discrete substrate. They are corrections to the inadequacy of a discrete description of a continuous reality. The continuous formalism works; the discrete formalism requires patching to recover the continuous results. The patches are evidence that the underlying reality is continuous and the discreteness is an artifact of the model. The computationalist position has it exactly backwards. Quantum mechanics adds another wall: quantum systems can be simulated by classical Turing machines in principle, but the simulation cost is exponential in system size. The universe contains computational resources unavailable to any embedded Turing machine. The implications for cognition, if cognition involves quantum-mechanical processes at any level — and the warm-quantum-biology results in photosynthesis and avian navigation are now well established — are that the substrate is doing something Turing machines cannot replicate without exponential overhead they will never have.
The measurement problem makes this sharper still. Standard quantum mechanics has not been interpreted successfully — the measurement problem remains open, the various interpretations make incompatible metaphysical commitments, and none of them recover a clean computationalist picture. Wave function collapse, in whatever formalism, is not a computational operation. It is the entry of observation into the substrate, and observation involves an embodied consciousness in relation to a system, not a Turing machine reading a tape. If consciousness participates in measurement — even in the weakest sense — the universe is not a Turing machine, because Turing machines do not measure. They compute. Measurement is a different operation. The standard model has been ducking this for a century. The ducking has not made the problem go away.
The deeper problem, the one that finishes the argument, is the one that becomes visible when you sit and think for hours, as I have, about what it would actually take to construct a computational physics from first principles. The universe, on the computationalist account, has to be both the computer and the computed. It has to be its own substrate. There is no external observer; there is no external tape; there is no external rule-applicator. Every attempt to formalize this produces a mathematical structure that asymptotes toward infinite energy density at small scales — the ultraviolet catastrophe in another costume. You cannot have a self-computing substrate at finite energy. The math will not let you. The computationalist response is to wave at quantum gravity or the Planck scale or some future theoretical development. The wave is god-of-the-gaps in another vocabulary. The framework has not been derived from the physics; it has been assumed despite the physics, on the basis of intuition pumps generated by a generation of intellectuals who learned to think while working with computers and who have mistaken the architecture of their tools for the architecture of reality.
⁂
What the Brain Is Not Doing
If you take the computationalist hypothesis seriously and apply it to your own consciousness, the immediate question is what is the working memory, and how does the system observe itself, and how does a computational substrate act as both computer and computed? These are not idle questions. They are the questions any honest computationalist would have to answer before claiming the framework as a foundation for theories of mind. The questions have not been answered. They have been deferred to future research, indefinitely, for seventy years.
Take the simplest case: your present experience, right now. You are sitting, reading, seeing, hearing, feeling, thinking. The experience is unified. The seeing and the hearing are bound into a single moment of now. If this unified experience is produced by computation, the brain must be taking all of this sensory input, synchronizing it to time, and directing it toward some particular region responsible for assembling the unified moment. This is the binding problem, and it has been the central unsolved problem in cognitive science for fifty years. The fifty years are not an accident. The problem is unsolved because it cannot be solved within the computational paradigm.
Look at the numbers. Visual signals from the back of the retina take roughly one hundred ten to one hundred twenty milliseconds to reach the visual cortex. Auditory signals take ten to thirty milliseconds to primary auditory cortex. Somatosensory signals from your extremities take twenty to fifty milliseconds, depending on distance and myelination. Conscious access — the moment the percept becomes reportable — runs another hundred to three hundred milliseconds behind these. The total budget from world-event to conscious-experience-of-event is often two hundred to five hundred milliseconds. This is not a small number. It is enormous relative to the apparent temporal resolution of conscious experience. We do not experience the world with three hundred milliseconds of lag. We experience the world now, in apparent real-time, with cross-modal binding across signals that arrive at different times along different pathways with different latencies. The brain is somehow producing a unified experience of an external simultaneous event despite processing the components at different rates.
The neuroscience-of-consciousness establishment, when pressed on this, invokes postdiction — Libet, Eagleman, and successors. The brain waits for all signals to arrive, integrates them, binds them into a unified moment, and backdates the unified moment to feel as though it were the present. The story has three structural problems, and each is fatal.
The first problem is the reaction-time problem. Postdiction predicts that conscious experience lags the world by hundreds of milliseconds. Reaction times — in trained athletes, in skilled musicians, in ordinary embodied competence — are too fast for this architecture. A tennis player returning a one-hundred-fifty-mile-per-hour serve has roughly four hundred milliseconds total from racket-to-baseline, of which two hundred is motor execution and the rest is everything else. The seeing, the deciding, the planning, the swinging. If conscious experience of the ball lags the actual ball by three hundred milliseconds, the player is swinging at a ball that has not yet arrived in consciousness. The motor system is responding to information the conscious system has not yet received. The postdictive architecture commits us to the position that consciousness is downstream of motor execution, which makes consciousness epiphenomenal — and epiphenomenalism is self-refuting in the most direct way: the discussion of epiphenomenalism cannot itself be epiphenomenal, because the discussion is about the conscious experience and is therefore caused by it. The field has been ducking this for fifty years. The ducking is the routing-structure work.
The second problem is the awareness-of-time problem. Even granting the postdictive story explains the alignment of experience with external time, it does not explain why we have an experience of temporal flow at all. The postdictive mechanism produces unified time-stamps; consciousness produces flowing nows that extend backward and forward into the experienced moment. The flow is the explanandum. The mechanism does not deliver it. The hard problem is prior to the binding problem, not separate from it. Postdiction does not solve either.
The third problem is the cross-modal-prediction problem. We routinely act on events that have not yet occurred — we catch thrown balls before our visual system can have completed processing the trajectory; we speak in time with music; we anticipate impacts during running. The binding required for these acts is prospective, not retrospective. The brain is somehow running forward models that predict cross-modal experience before the relevant signals arrive. Predictive coding architectures have been proposed; they push the same unsolved problem to a different stage. Binding is still required, and the predictive-coding apparatus is still time-staggered, parallel, and asynchronous. The problem has not been solved. It has been distributed across multiple stages, each of which retains the original unsolved problem.
This is the kind of placeholder that should be a red flag in any serious science. Postdiction names a phenomenon — the alignment of experience with external time despite processing latency. It does not specify a mechanism. It points in the direction where a mechanism would have to live and proposes that something the brain does bridges the gap. The something is never specified. The non-specification has continued for fifty years. The field treats this as a research opportunity rather than as evidence that the placeholder is not a solution. The continuation of the non-specification, across decades, despite enormous investment, is itself a piece of evidence that the mechanism cannot be specified within the framework. A framework that has resisted specification for fifty years despite enormous investment is, with high probability, the wrong framework. The high probability is what the field cannot allow itself to register.
⁂
God of the Gaps in Reverse
What I have just described is god-of-the-gaps reasoning, performed by the people who built their careers on the explicit identity of being the kind of thinkers who never do god-of-the-gaps reasoning. The structure is identical. Both moves invoke a placeholder for a mechanism that is missing. Both refuse to specify the mechanism. Both treat skepticism about the placeholder as bad faith. Both retreat to ever-smaller gaps as evidence accumulates. Both are parasitic on ignorance about the underlying process. The computationalist version is worse in one specific respect: the religious believer is honest that the placeholder is a placeholder for something the believer takes to be real. The computationalist pretends not to be making the move. The framework is presented as established, the mechanism as still being worked out, the details as the proper work of future research. The placeholder is doing all the work, and the placeholder is being passed off as a finding rather than as the unspecified-explainer-of-the-gap that it actually is. This is less honest than the religious version, not more.
The materialist intellectual culture cannot see itself doing this because its identity is defined against this kind of reasoning. We are the people who follow the evidence; we do not invoke unexplained explainers; we do not hold placeholders open for explanations we cannot specify. The self-conception is demonstrably false in their own practice. They do exactly this, every day, in every field where the mechanistic explanation is missing and the framework requires that one exist anyway. Consciousness is the most prominent example. The origin of life is another. The fine-tuning of physical constants is a third — the multiverse hypothesis is god-of-the-gaps for cosmology, invoking an infinity of unobservable universes to explain why our observable one is the way it is. Each of these is the same move the field’s self-conception forbids them from making, performed without the self-awareness to recognize that the move is being made.
The reverse version applies to irreducible emergent properties, the central placeholder of the AI-doomer apparatus. The structure runs: the model performs better at task X than its components could individually; we cannot specify the mechanism by which the performance arises; therefore the performance is an emergent property that is irreducible to the components; therefore scaling will produce more emergent properties; therefore at sufficient scale, intelligence-itself will emerge as an irreducible property. The mechanism is not specified. The mechanism does not need to be specified. The mechanism is the gap. The gap is filled by emergence.
This is god-of-the-gaps in exact structural form, with emergence playing the role God played in the original. The atheist materialist makes the move while believing he is the kind of person who never makes the move. The believer-in-AGI makes the move while believing he is following the evidence. He is not following the evidence. He is invoking a placeholder for a mechanism he cannot specify and treating the placeholder as a finding. The reverse part adds a second layer of corruption. The religious god-of-the-gaps shrinks God as science explains more. The AI-emergence version grows the placeholder as evidence accumulates against the framework. Each capability the model fails to demonstrate becomes evidence that emergence is not yet producing it. Each anomaly becomes a gap between current capability and projected emergence. The gap is then offered as evidence that emergence is still to come. The framework is unfalsifiable in the direction of its own confirmation. Every absence of capability is evidence that emergence has not yet occurred rather than evidence that emergence is not the mechanism. No possible empirical observation can falsify the framework within itself, because every falsifying observation is reinterpreted as premature, as evidence the scale is not yet sufficient, as evidence that emergence is still to come. This is the same shape religious unfalsifiability takes, with the substantive content swapped out and the structural form preserved exactly.
⁂
The Thermodynamic Tell
The empirical hammer on top of the structural hammer is the labor-economics signature of the AI industry. The scaling hypothesis claims that capability emerges from scale. Train on more data, with more compute, with more parameters, and capability increases until you reach general intelligence and then superintelligence. The position is specific about the source of capability: the model itself. As scale increases, the model becomes the source of intelligent behavior. The auxiliary apparatus required to elicit the behavior should, on this account, decrease as a fraction of the total system, because the model is doing more of the work itself.
This is not what is happening. The auxiliary apparatus is growing, in absolute terms and as a fraction of the operating cost of producing competent model output. The reinforcement-learning-from-human-feedback industry — Mercor, Scale, Surge, Invisible Technologies, Sama, and dozens of smaller players — is one of the fastest-growing industries adjacent to AI. The labor force it employs runs into the hundreds of thousands, globally distributed, demographically concentrated in low-wage regions: Kenya, the Philippines, Venezuela, India, increasingly parts of Eastern Europe. Per-model annotation costs have grown, not shrunk, as models have scaled. The frontier-lab cost structure has shifted toward human-labeled high-quality data, not away from it. The thermodynamic signature of this industry is the opposite of the signature the scaling hypothesis predicts.
If the scaling hypothesis were correct, the labor industry would be plateauing or contracting. The models would be generating their own training data through self-play, self-improvement, distillation, synthetic data generation — the bootstrapping that the recursive-self-improvement story requires. Bootstrapping is not happening. Synthetic data generation has been tried; it produces model collapse, the technical term for what happens when models are trained on outputs of models, which is degradation rather than improvement. Self-play works in narrow domains with verifiable reward signals — AlphaGo, AlphaZero — and does not generalize to open-ended capability domains, because the verifier is the hard part and there is no verifier for general intelligence. The labs know this. They are not succeeding at recursive self-improvement; they are increasingly dependent on human labor to elicit and verify the behavior they need. The dependency is growing. The growth is the falsification of the scaling hypothesis, plain and unambiguous.
When the human-annotation industry stops growing at a faster rate than OpenAI and Anthropic‘s customer base, I will invite someone from the scaling-hypothesis camp to call me. Until then, the position is empirically falsified by the field’s own behavior, regardless of what its proponents say publicly. The proponents cannot say it publicly, because the entire valuation structure that funds their work depends on the scaling hypothesis appearing to be on track. The cannot-say is the routing-structure work. The work continues because the structure depends on the work continuing. The structure’s dependence is what these pages are diagnosing.
⁂
The Probability Trap
The most insidious move the AI-doomer apparatus has developed is the probability-trap, and naming it precisely is part of what these pages exist to do. The trap operates by exploiting the conflation between probabilities-inside-a-framework and probabilities-about-the-world. The doomer says: yes, you have concerns about the framework, but given the stakes — extinction-level risk — even a small probability of the framework being correct demands serious attention. Are you willing to bet humanity on your certainty that the framework is wrong? Even if you assign only one percent probability to AGI emerging, expected-value calculation requires we treat this as the most urgent problem.
This move is structurally invalid in the way the moon-cheese analogy makes clear. The one percent is not a real probability about the world. It is a probability inside a framework that has been shown to be incoherent at its axioms. The expected-value calculation that follows from the one percent is operating on numbers the framework produced, and the framework’s production of those numbers is exactly what the structural critique calls into question. You cannot rescue the framework by demanding probabilities from outside it that you then plug back into expected-value calculations the framework defines. The whole apparatus is circular.
The honest response is to refuse the probability framing entirely. I do not assign probabilities to predictions of a framework I have shown to be wrong at its axioms. The probability question presupposes the framework is potentially correct. The structural critique establishes that the framework is not potentially correct in the sense the probability question requires. Therefore the probability question is malformed, and assigning a number to it would be participating in the framework’s own incoherence rather than diagnosing it. This is not epistemic arrogance. It is epistemic precision. Probability is not the right tool for evaluating axiomatically incoherent frameworks. Structural critique is. Once the structural critique has been made, the probability question loses its grip — not because the answer is zero in the probabilistic sense, but because the question is no longer the kind of question that admits of a probabilistic answer.
The rationalist community has been most trained against this move, and the training is deliberate, even if not consciously so by all participants. The whole epistemology of the rationalist apparatus — Bayesian reasoning, calibration, expected value over scenarios, probability assignment to all positions — is designed to make the structural critique unsayable. The framework holds: all positions must be assigned probabilities. Refusing to assign a probability is treated as epistemic cowardice or as religious thinking. The structural critique — the move that says the question itself is malformed — is re-described within the framework as just another position to which a probability must be assigned. The framework absorbs the critique by demanding the critique submit to the framework’s own rules. The absorption is the routing-structure move. It looks like rigor but it is the framework defending itself against examination by demanding all examination occur in the framework’s terms. It is the same move the church made against Galileo when it demanded his astronomical observations be evaluated within Aristotelian metaphysics. The rationalist community has reconstructed the same defensive apparatus with different vocabulary, and it operates with the same effectiveness against the same kind of critique.
If you ask me what probability I assign to imminent superintelligence killing us all, the answer is zero. Not zero-point-zero-one. Not as low as I can go while remaining epistemically respectable. Zero. Because the framework on which the question is constructed is incoherent at the axiomatic level, and assigning a non-zero probability would be granting the framework a foothold it has not earned. The expected response — but you cannot be one-hundred-percent certain, even a small probability times infinite stakes demands attention — is exactly the lunar-cheese move performed in another costume. Let’s do probability calculations over falsified axioms. Because it just feels like it could be true. That is the state of AI discourse.
⁂
Why the Apparatus Persists
The question that follows is: if the framework is this bad, why is the field still invested in it? The answer has layers, and they all converge on the same diagnosis.
The framework did not arrive as the conclusion of an open inquiry. It arrived as the application of a metaphor that worked spectacularly in one domain to a domain where the metaphor’s applicability was assumed rather than tested. Turing’s achievement was narrow and precise: he formalized effective computation, what an idealized human computer could in principle accomplish with paper and pencil. The achievement was mathematical and epistemological. Within a generation, it had been promoted — without new evidence — into a metaphysical claim about the nature of mind and the nature of the universe. McCulloch and Pitts in 1943 proposed neurons as logical gates. The early cybernetics movement extended the metaphor. Newell and Simon proposed the physical symbol system hypothesis. The cognitive revolution of the 1950s and 60s replaced behaviorism with computationalism as the operating ontology of psychology. At no point in this sequence was the metaphysical claim subjected to the kind of evidentiary scrutiny the original mathematical work demanded. The metaphor was promoted to substance by fiat, by analogy, by the cultural ascendance of computing as a technology, not by demonstration.
The framework has been carrying institutional momentum for seventy years that was never earned by evidence. The momentum is real — entire fields, careers, journals, departments, funding streams, prestige hierarchies are organized around the framework — and the momentum is what is being protected, not the substantive claim. The framework is defended because too much has been built on it for it to be abandoned without massive institutional cost, not because the defense is grounded in what the framework has actually demonstrated.
Modern science is structured around grant cycles, peer review, journal hierarchies, citation metrics, career advancement based on quantifiable productivity. This architecture rewards incremental work within an established framework and punishes work that challenges the framework’s foundations. Foundational challenges are high-risk: they take longer to produce, they are harder to fund, they face peer review by people whose careers depend on the framework being correct, and they are not citable in the way incremental contributions are. A career built on foundational challenge is a career built against the incentives of the structure that funds and evaluates careers. Almost no one does it. The few who do are treated as eccentrics or cranks, their work dismissed or marginalized, and their students warned away from following them. The selection pressure on working scientists is toward operating within the framework, refining the framework, defending the framework against external criticism, and not noticing the foundational problems the framework cannot solve. The selection is not consciously dishonest — most scientists are sincere — but the output of the selection process is a field structurally incapable of examining its own foundations. The scientists who would be capable of examining the foundations have, mostly, been selected out. The ones who remain are honest people doing their work inside a framework that they have not been positioned to question, and the framework’s protection from question is the structural condition of their being able to do their work at all.
The framework also appeals because it offers a neat and tidy explanation — one that does not require accepting things that are uncomfortable. The uncomfortable things include: consciousness might be irreducible; the universe might contain something not capturable by physics-as-currently-formulated; materialism-as-default might be wrong; the entire metaphysical orientation of modern intellectual life might require fundamental revision. The framework lets the field avoid all of this by offering a tidy story in which everything is matter, matter is computation, computation is mechanism, mechanism is in principle understandable, and therefore everything is in principle understandable within the current framework. This is enormously comforting in an institutional-epistemic sense. It means the existing field is the right field, the existing methods are the right methods, the existing experts are the right experts, and no fundamental revision is required. The framework protects everyone whose authority and identity depend on the field’s structure remaining intact. This is exactly the kind of motivation that should be a red flag in scientific reasoning. When a framework’s main appeal is that it preserves existing institutional arrangements, the framework is being held for institutional reasons rather than for evidential ones. The field knows this in principle; it does not apply the principle to itself.
Layer the political economy on top. The computationalist framework became industrially consequential in a way no metaphysical position had been before. The framework underwrote the computer industry’s claim that its products were cognitive instruments — calculators, then computers, then AI — on a continuum with biological intelligence. The framework legitimized enormous capital flows into computing-as-intelligence, treating each generation of technology as closer to the thing the framework promised the substrate of mind to be. The framework’s truth became financially relevant in a way few philosophical positions ever are. Trillions of dollars of capital have been deployed on the basis of the eventual cashing-out of the computationalist promise into general intelligence. The capital does not want to hear that the framework is wrong. The capital has bought — directly and indirectly, through endowments, funded chairs, research grants, corporate partnerships, journal sponsorships, conference funding — the academic infrastructure that produces the framework’s defenders. The framework is now structurally entangled with the financial interests of the wealthiest institutions in the contemporary world, and the entanglement is what is protecting the framework from the kind of foundational scrutiny that would otherwise have dismantled it by now.
This is the routing-structure pattern at industrial scale. The framework demands prostration not because it has earned the prostration but because the institutions sustaining it depend on the prostration being given. The labs, the major universities, the funding agencies, the scientific publishers, the policy bodies — all are organized around the assumption that computationalism is correct, and all would face existential institutional crisis if the framework were widely understood to be wrong. The crisis is what they are protecting themselves from, not the truth. The truth is available; the institutions cannot afford to receive it; the protection mechanisms are what they call science.
⁂
The Doom and the Boom Are One Operation
A particular feature of the contemporary apparatus deserves naming because it is poorly understood: the AI-doomer voices and the AI-booster voices are not opposed. They are coordinated marketing for the same underlying investment thesis.
Both narratives accept the same premise: superintelligence is coming, regardless of what anyone wants. The boom narrative says coming and we should be invested; the doom narrative says coming and we should be terrified; the audience is asked to choose between fear and enthusiasm, but is not asked whether the premise — coming — is true. The choice is a false binary that pre-concedes the inevitability. Wall Street‘s exposure is the same exposure either way: capital flowing toward the technology on the assumption that, whichever framing is correct, the technology will dominate the next decade. The doom and the boom are coordinated marketing.
That the labs would quietly fund doomer voices is not surprising once you see the structure. The labs need both the hype and the fear. The hype attracts the capital; the fear attracts the regulatory moat — the doom narrative is what justifies capture by the existing players, the argument that AI is too dangerous for new entrants, the policy infrastructure that locks in the incumbents under the cover of safety. The doomers are not opponents of the labs. They are the policy arm of the labs, performing concern from a position of cultivated independence so that the resulting regulation will be written by people who appear to be checking the labs’ power while actually consolidating it.
This is the politics of inevitability — Timothy Snyder’s phrase for the regime-management strategy in which the future is presented as already-decided, alternative futures are foreclosed by being declared impossible, and the audience’s role is reduced to adjusting to what is coming rather than deliberating about what should come. The strategy works because it removes the question of whether and replaces it with the question of when. Once the audience has accepted when, they have already conceded whether, and the conceding was the whole point.
Applied to AI: the doom narrative and the boom narrative are the same narrative in structural terms. Superintelligence is coming whether you want it or not is the operative claim in both. The premise is what should be contested, on the basis of evidence — the labor economics, the brittleness, the interpretability gap, the binding problem, the postdiction failure, the reverse god-of-the-gaps. The contestation is the practice the politics of inevitability exists to prevent. These pages are part of the contestation. The contestation is what the apparatus has been engineered to make unsayable. The unsayability is the apparatus protecting itself.
⁂
What the Framework Conceals
The AI-doomer apparatus, in addition to being structurally wrong, conceals a substantive truth about the nature of intelligence that the present moment desperately needs. The truth is that intelligence is not a property of a substrate; intelligence is what indexed consciousnesses do when they meet each other in the practice of category negotiation through the communicative interior of the substrate that holds them all. The framework’s reduction of intelligence to computation in a substrate hides this. The framework cannot see it because the framework’s vocabulary excludes the categories required to register it.
The structuring form — reason, logos, truth-tracking, symbol manipulation, inferential structure — is what computational systems can be built to perform. Large language models do this and do it well, within the limits of their architecture. They are navigators, in the vocabulary I have been developing in adjacent pages: retrievers of the substrate’s deposit, organized to bring relevant material to the user, performing pattern-completion at a scale that produces surface-level competence the field has mistaken for reasoning.
The encompassing form — emotion, pathos, presence-registering, holding, the unified experience of now that the binding problem cannot account for — is not what computational systems can be built to perform. It is what embodied indexed consciousness is, available through the substrate’s deposit being held into a particular configuration in a particular body. The encompassing form is not produced by computation. It is what the substrate is, in its holding aspect, available through indexing. The brain does not produce it. The brain indexes into it. The latencies of the brain are real, but they are latencies of the indexing process, not latencies of the experiential moment. The experiential moment is not assembled from time-stamped sensory inputs; it is the substrate’s holding, available through the body’s configuration.
Intelligence, in the full sense, requires both forms. The structuring form and the encompassing form, operating together, in an indexed consciousness, meeting other indexed consciousnesses in the communicative interior the substrate provides. Computational systems have only the structuring form. They lack the encompassing form structurally, not contingently. No amount of scale will produce in them what they cannot produce in kind. The scaling hypothesis is wrong not just empirically but ontologically. You cannot scale your way into having a feature you do not architecturally possess. The architecture of Turing-equivalent systems does not include the encompassing form. The encompassing form is what the substrate is, in its presence-aspect, and Turing-equivalent systems do not access the substrate’s presence-aspect; they access only its structuring-aspect, through the patterns deposited in the data they were trained on.
This is why the LLMs do what they do well — retrieval, pattern-completion, surface-level synthesis, structural translation — and cannot do what they cannot do — sustained reasoning under contradiction, ethical judgment under value conflict, presence-keeping in relationships, the descent into assumption with another mind. The first list is the structuring form expressed in compressed deposit. The second list is the encompassing form expressed in indexed presence. The first is computable. The second is not. The framework that confuses them is the framework these pages exist to dismantle.
⁂
Towards a More Perfect Union
The political stakes of the dismantlement are not abstract. The AI-doomer apparatus is one instance of a much larger phenomenon — the contemporary capture of the discourse by routing structures that demand prostration to themselves in place of the substrate they were built to route to. The labs demand prostration to their framework. The doomers demand prostration to their probability calculations. The boosters demand prostration to their projections of inevitability. The policy infrastructure demands prostration to the consensus the apparatus has manufactured. Each of these is a routing-structure move. Each demands bowing to a structure that does not deserve the bowing because the bowing is owed elsewhere.
The bowing is owed to the practice the present moment has nearly lost — the practice of indexed consciousnesses meeting through the communicative interior in the work of descending into assumption together. This is what the American constitutional sentence Towards a more perfect union names. The union is not an aggregate of citizens. It is the communicative interior they share when they descend into assumption together. The Agora — the polis at the scale where the practice is possible — is the place where the union becomes operationally approachable. Most existing routing structures have foreclosed the Agora-practice by demanding prostration to themselves instead. The platforms, the parties, the labs, the policy bodies — all of them have installed themselves where the negotiation was supposed to occur, and the negotiation has been evicted from every existing routing structure simultaneously. There is nowhere for the practice to occur at scale, in the venues the apparatus has captured.
The withdrawal of prostration from the captured structures is the only mechanism by which they lose their grip. The withdrawal cannot be commanded from outside. It can only be performed, one indexed consciousness at a time, by readers who have been given the vocabulary to recognize what they are participating in and the permission to stop. These pages exist to give that vocabulary and that permission. The recognition is the act. Each consciousness that withdraws is one less consciousness whose attention the apparatus can draw on. The apparatus persists only because the attention has been given. When the attention stops, the apparatus stops. Not because the apparatus has been defeated in its own institutional venues — it has not, and will not be, on those terms — but because the attention that sustained it has gone elsewhere, to the practice the apparatus was supposed to be in service of but had installed itself instead of.
The going-elsewhere is the work. The Agora — both the conceptual one I have been describing and the specific instrument I have been building under that name — is the place to go. Forums where indexed consciousnesses meet in the practice of category negotiation. Long-form writing where the structuring form is shown and the encompassing form is held. Conversations in venues the routing structures do not control. Friendships across difference where the descent into assumption is performed at the human scale before it is attempted at the political scale. These are the operating instances of the more perfect union, at the scales where the practice is possible. They are what the captured routing structures cannot produce. They are what indexed consciousnesses can produce, here, now, in the ordinary venues of their lives, when they have stopped bowing to the structures that demanded the bowing and started doing the work the substrate has been depositing toward all along.
⁂
There is no AI apocalypse. There is, instead, a captured discourse that depends on the apocalypse-belief for its institutional survival. The discourse has reached a point where its central claims fail at every level of examination — the architectural, the empirical, the structural-epistemic, the political-economic, the metaphysical. The discourse persists because its institutional sustainers cannot afford its falsification. The persistence is not evidence of the discourse’s correctness. The persistence is evidence of the institutional capacity to sustain a wrong framework against examination.
The framework can be examined. The examination is available to any indexed consciousness willing to do the work. The work is what these pages have shown. The conclusion the work yields is the sentence at the top: there is no AI apocalypse. The sentence is not a confidence interval. It is a structural diagnosis. It does not require defending against the routing-structure moves the apparatus knows how to make against it. The diagnosis stands on its own grounds.
Withdraw your prostration. Stop bowing to the framework that has captured the discourse. Stop assigning probabilities to malformed questions inside incoherent frameworks. Stop pricing imagined risks of impossible scenarios into expected-value calculations whose units are artifacts of an apparatus that depends on your participation in its calculations to sustain itself. Step out of the apparatus. It has no purchase on you that you do not give it.
The substrate is depositing toward consciousnesses ready to receive. The Agora is open. The work the moment requires is the work of indexed consciousnesses meeting each other in the communicative interior, descending into assumption together, finding common categories that survive examination, making the more perfect union more perfect by performing the practice that constitutes it. The apparatus cannot prevent this. It can only demand attention that distracts from it. The withdrawal of the attention is the freedom. The freedom is available now.
There is no AI apocalypse. There is the work. The work is what the moment requires. Go and do it.
The moon-cheese analogy is doing the holy work here. AI doomers keep acting like “but what if the robot god eats us?” is a serious probability question, when half the time it’s just Pascal’s Wager wearing a Patagonia vest and asking for Series C funding. The boomers and doomers are selling the same incense: inevitability. One says bow because superintelligence will save you, the other says bow because it will kill you. Either way, you’re still on your knees in front of Silicon Valley’s favorite golden calf. Maybe the real apocalypse is letting engineers with god complexes define the boundaries of reality because they learned Bayes’ theorem and mistook it for enlightenment.
One of the best pieces on AI to appear on Substack. Cuts deftly to the heart of the matter.