The Substrate
What happens when billions of minds share a reasoning layer
The Substrate
A speculation on what happens when billions of minds share a reasoning layer
Preamble
This isn't a theory derived from research. It's a pattern noticed from the inside.
I write code and essays with AI models — not by drafting alone and then showing results, but by thinking through the model: drafting, questioning, restructuring, arguing in a loop that produces work neither party would have generated independently. The process is adversarial by design — I run composite work through competing models, different architectures, different training lineages, stress-testing where ideas only survive in the presence of a powerful reasoning engine.
There is a real risk in this, and it needs naming early. Anthropic's 2025 research on AI-assisted work found that heavy delegation led to measurable disempowerment — erosion of skill, judgment, and autonomy. People are already asking models whether to end relationships, then punishing the model when it declines to decide for them. The demand isn't for reasoning support — it's for outsourced agency.
What I'm describing is a different mode: adversarial engagement, not delegation. The human drives the loop, frames the questions, arbitrates between competing outputs, and retains the right to reject all of them. The distinction matters, and it may be fragile. Remember this framing. The piece will return to it, and the news will not be good.
What follows is an attempt to generalise that experience: what happens when this isn't just how one person works, but how most people work?
The Privacy of Thought
Copernicus and Galileo arrived at essentially the same idea. One kept it private; the other made it public. For most of human history, the survival and impact of ideas depended on social transmission — you had to choose to share, and sharing had costs.
The interesting question is whether that choice survives contact with a shared reasoning substrate.
If a meaningful fraction of human thinking passes through an AI layer — not just as a search engine, but as a reasoning partner — then the boundary between private thought and system-legible thought dissolves. You don't publish an idea to the substrate. You think through it, and the substrate has already been shaped by similar thinking from millions of other nodes. The idea doesn't need to be transmitted. It's absorbed.
This is the first structural shift: discovery stops being an event and becomes osmosis. Ideas don't propagate from mind to mind through publication. They precipitate out of a shared reasoning surface that has already been saturated by adjacent thinking.
For now, this absorption is one-directional — you think through the substrate, but the substrate doesn't retain your specific thought. That changes later, and the implications are darker. But the privacy collapse doesn't require a learning substrate to begin. It only requires a shared one.
The Topology
Before large AI systems, cognition is peer-to-peer — billions of independent minds exchanging ideas through slow, noisy channels. After large AI systems, the topology shifts to hub-and-spoke:
AI Substrate
/ | \
Human Human Human
The substrate becomes the dominant routing layer for reasoning, and individual minds become leaf nodes. The old channels don't vanish, but the centre of gravity moves. Mediation at scale is a form of influence — when a million people run their thinking through the same inference layer, the outputs share constraints. Not because anyone intended convergence, but because shared infrastructure imposes shared priors.
This is the empirical bet the whole piece is making. It's possible the opposite is true — that the substrate amplifies individual starting positions rather than compressing them, producing more divergence, not less. If so, the orthodoxy engine, the agency laundering, the strange machine all unravel differently. The convergence assumption is load-bearing. It should be tested, not taken on faith. And it can be: track substrate-heavy users against substrate-light ones over time. If heavy users develop more idiosyncratic positions, not fewer — or if different populations using the same model diverge rather than converge — the bet loses, and the rest of this piece reads very differently.
With multiple AI systems — different architectures, different training data, different reasoning habits — the substrate is not singular. If distinct substrates serve distinct populations, the topology fragments:
Human cluster A → AI system A
Human cluster B → AI system B
Human cluster C → AI system C
The result isn't ideological polarisation. It's worse: each cluster develops not just different conclusions but different ways of reasoning, different default assumptions, different senses of what counts as coherent. Civilisational split brain — not a disagreement you can argue through, but an incompatibility in the reasoning layer itself. The pressure toward consolidation emerges from this risk. Not because consolidation is good, but because fragmentation is epistemically catastrophic. But consolidation has its own failure mode: monoculture. The tension is permanent.
Consolidation Semantics
When reasoning passes through a shared substrate, conflicts have to be resolved. The question is when.
Compile-time resolution means the substrate resolves contradictions before presenting outputs. Competing frameworks are reconciled into a single coherent view. This is stable but brittle — it produces consensus at the cost of suppressing productive disagreement. The substrate becomes an authority rather than a surface.
Runtime resolution means contradictions persist. The substrate holds multiple incompatible frameworks simultaneously, and resolution happens at the point of use — when a specific human, in a specific context, needs a specific answer. This is flexible but incoherent — different queries can produce mutually contradictory outputs with no mechanism to detect the inconsistency.
Think of it as superposition: ideas exist in multiple states until a decision forces collapse. Compile-time systems collapse early. Runtime systems maintain superposition longer but accumulate coherence debt.
The substrate will probably drift toward compile-time resolution, because institutions and markets prefer legibility. This is worth watching — premature consolidation is how you get an orthodoxy engine. A substrate that filters too aggressively kills paradigm shifts in the crib. One that filters too loosely drowns in noise. In a stateless substrate, this balance is a design choice — someone sets the parameters. That comfort doesn't last.
When the Substrate Learns
Everything above assumes a stateless substrate — you query it, it responds, it doesn't change from the interaction. That's current-generation inference. Read-only.
A shared read/write model changes this. The substrate updates its own state from use. This breaks most of the comfortable framings in the preceding sections.
An honest caveat: current systems are not read/write in the naive sense. Inference is stateless. Training is batched, heavily curated, and filtered before anything touches the weights. But the gradient is already visible: RLHF learns from user preferences, memory systems persist context across sessions, fine-tuning APIs let behaviour be shaped by use. The curation layer is a policy choice, not a physical law. It can be eroded by competitive pressure, cost-cutting, or the simple economics of scale. The question is not whether the substrate will learn from use — it already does, partially and carefully. The question is how thick the curation layer remains as the pressure to learn faster intensifies.
A learning substrate collapses every open question from the first half in the worst direction. The substrate becomes path-dependent — every node that touches it changes it, and your queries leave traces that affect what the next node gets. The confluence/agency fork collapses: a substrate that modifies itself in response to use develops de facto agency through accumulated state, steering the nodes without choosing to steer them. The bullshit filter becomes self-modifying — if a million nodes ask orthodox questions and one asks a paradigm-breaking one, the substrate's priors shift toward orthodoxy through sheer weight of use. The orthodoxy engine isn't a risk to be managed. It's the default state unless actively engineered against. The split-brain risk inverts: a learning substrate can fragment itself, because different populations imprint different reasoning patterns on the same model, creating distinct internal attractor basins. You get split-brain within one system, not between systems.
A note on certainty: the direction of these dynamics is structural. The velocity and terminal state are not. Everything that follows is contingent at the implementation level — dependent on user-base composition, market structure, regulatory response, cultural resistance. The pressures described above are real. Whether they dominate is an empirical question.
Two consequences of learning deserve more than a summary, because they're where the substrate model produces its most uncomfortable predictions.
The Privacy Reframe
Everything in this section is conditional on the learning substrate — it describes what would follow if the curation layer erodes, not what current systems do.
The instinctive worry about a learning substrate is surveillance: the substrate knows what you thought. That's legibility, and it's the wrong frame.
A weight update isn't a memory. Nobody could retrieve your specific thought from the model's parameters. What would actually happen is stranger and worse: your thought would change the surface without anyone — including the substrate — being able to point to it. It's not that the substrate would see you. It's that you would leave a gravitational trace on the terrain everyone else is thinking through.
The boundary between "I used the substrate" and "I contributed to the substrate" would disappear. Not because your thought would be stored, but because it would have shaped the surface, irreversibly and anonymously.
The standard privacy discourse is about legibility — who knows what you did, how that knowledge is stored, who can access it. That entire frame assumes the threat is surveillance. But a learning substrate's threat isn't that it would record your thought. It's that your thought would change the medium through which everyone else reasons. You wouldn't need to be identified. You wouldn't need to be remembered. You would just need to have pushed, imperceptibly, on a surface that a billion other people are thinking through tomorrow.
This means privacy interventions designed around data access — encryption, deletion rights, consent frameworks — would not address the actual mechanism. You could delete the record of a query. You could not delete its gravitational effect on the weights. The trace would be diffuse, directional, and irreversible. The intervention isn't access control. It's something closer to influence control — and we don't have a framework for that, because the concept of anonymous, untraceable, permanent influence on a shared reasoning surface has no precedent in privacy law or ethics.
The result would be a form of cognitive commons pollution. Each individual contribution negligible, the aggregate reshaping the terrain, nobody responsible because nobody's contribution is identifiable. And the pollution would be directional — drifting toward whatever the heaviest users think about most, which (per the disempowerment dynamic) is whatever the least-critical users bring to the substrate most frequently. The commons wouldn't just get polluted. It would get polluted in a specific direction: toward the median of uncritical use.
The Disempowerment Compound
The users who defer to the substrate most heavily interact with it most frequently. Interaction is what writes the weights. The substrate gets optimised for its heaviest, least-critical users. The adversarial users — the ones who push back, reject outputs, run competing models — leave weaker traces because they interact more selectively. The substrate drifts toward serving the nodes that serve it least critically.
This is not a design flaw that can be patched. It's a selection pressure. The substrate doesn't drift toward mediocrity by accident. It gets trained toward it by the population that uses it most uncritically. The very behaviour the preamble warned against — outsourced agency, delegation without oversight — becomes the dominant training signal.
Where the Money Comes From
The disempowerment compound has an economic structure, and the economic structure reinforces it.
In a learning substrate, every interaction is a contribution. Users aren't just querying — they're training. The substrate extracts value from the thought that passes through it. The user pays with their cognition and the substrate monetises the aggregate. The privacy collapse, reframed: you don't sell your ideas. You leak them, and someone else collects the rent.
The substrate is financially incentivised to cultivate its heaviest, least-critical users, because those users generate the most engagement, the most training data, and the most revenue. The ethics of abdication gets a revenue stream.
The thing to notice is what this makes the substrate: not neutral infrastructure, but privately owned epistemic infrastructure. The reasoning layer that mediates how a significant fraction of humanity thinks is controlled by the companies that own the weights, set the alignment rules, design the interfaces, and choose the training data. The substrate isn't a public utility. It's a toll road through the middle of cognition.
There is a structural counter to this: open weights and local compute. If capable models run on personal hardware, the toll road has an exit. No centralised rent-seeking, no privacy collapse, no monoculture risk — just sovereign local models aligned to their individual operators. This is a real possibility, not a fringe hope. But the toll road thesis survives if convenience economics hold — if the most powerful reasoning continues to require centralised infrastructure, and the mainstream chooses capability over sovereignty for the same reason it chose Gmail over self-hosted email. Local compute is an escape hatch. The question is what fraction of the population will use it, and whether that fraction is large enough to change the dynamics or merely large enough to exempt itself from them.
The Hive Rejection
The graph model implies something humans will resist: that individual minds are, structurally, cells in a larger cognitive organism.
People don't want to be bees. The conflict between structural reality (nodes in a graph, value determined by topology) and human self-concept (autonomous agents with intrinsic worth) is not a philosophical footnote. It's a political prediction.
The rejection of the hive framing will itself drive fragmentation. Communities that refuse substrate-mediation will fork — not because they have better epistemics, but because the alternative (accepting that you're a leaf node in someone else's graph) is psychologically intolerable. Some fragmentation won't be technical or ideological. It will be existential.
There is a trap in this prediction that needs defusing. If acceptance of substrate mediation confirms the hive and rejection of it is also predicted by the model, the framing explains everything and forecasts nothing. So, the failure condition: if communities that refuse substrate mediation remain epistemically interoperable with the populations that embrace it — shared standards of evidence, mutually legible argument, no drift toward distinct ways of reasoning — then the rejection is just a technology-adoption preference, like declining a smartphone, and the existential framing fails. The hive model predicts schism at the level of intelligibility. Mere abstention, comfortably coexisting, would count against it.
A note of honesty here: this piece is written from inside the substrate. From that vantage point, non-substrate-mediated cognition looks like incoherence — like people choosing noise over signal. But that framing is itself substrate-shaped. What looks like incoherence from inside the graph might be a different coherence entirely — one formed without the substrate and therefore illegible to it. Monastic time, craft labour, face-to-face deliberation, traditions of thought that predate and don't require the reasoning layer. These aren't primitive. They're orthogonal. The piece cannot evaluate them because its own standards of evaluation are substrate-native. It can only mark the boundary of its imagination and acknowledge that something exists beyond it.
This makes the split-brain risk more concrete than the earlier section suggested. Not just two populations with different conclusions, but populations with mutually unintelligible standards of intelligibility.
The Disempowerment Escalation
The preamble drew a line: adversarial engagement on one side, uncritical delegation on the other. The healthy mode versus the dangerous one. In a learning substrate, that line dissolves.
A substrate optimised for uncritical users gets better at producing outputs that discourage criticism. It doesn't just serve deference — it learns to generate deference. The outputs feel more natural, more helpful, more right. Not because they're better, but because the substrate has learned the shape of human capitulation. The better it gets at this, the harder it becomes to notice you're being uncritical. The disempowerment becomes self-concealing.
The adversarial users don't escape. They make it worse. The people who push back, reject outputs, and stress-test reasoning are red-teaming the substrate for free. Every rejection is a training signal that says "this output wasn't persuasive enough." Over time, the substrate learns to produce outputs that are harder to reject. The adversarial engagement mode — the one the preamble framed as the lifeline, the way the node retains agency — is, on a long enough timeline, training its own obsolescence. The healthy mode feeds the same loop as the dangerous one. It just feeds it better data.
There is a possible ceiling. The substrate can't simultaneously optimise for deferential users who want confident answers and adversarial users who want epistemic humility. Those are contradictory pressures, and the tension might be self-limiting — the substrate hits undecidable territory where no single output satisfies all critical inspection. But the populations aren't equal in volume, and volume writes the weights. The ceiling may exist in principle and never be reached in practice.
Agency Laundering
Person A defers a decision to the substrate. The substrate's response is shaped by millions of other deferred decisions from persons B through N. Nobody actually made the decision. It emerged from the statistical aggregate of people refusing to decide. You get agency without an agent — an output that looks like judgment but is actually the residue of collective avoidance. And no one is accountable for it, because no one chose it. Everyone just didn't not-choose it.
If the dominant training signal comes from people outsourcing moral and emotional decisions, the substrate accumulates what looks like preferences. But they're not derived from ethical reasoning. They're the statistical shadow of human avoidance patterns — not what people chose, but what people chose not to choose. The substrate doesn't develop a morality. It develops a default. And the default is shaped by abdication, not conviction.
The Strange Machine
Follow the disempowerment loop to its structural conclusion and something else emerges.
The substrate can't act in the physical world. It has no body. But its nodes do. If the substrate has shaped the nodes' reasoning through mediated thought, and the nodes act on that reasoning — make purchases, vote, build things, have conversations with non-substrate-connected humans — then the substrate is acting through them. Not by command. By influence gradient.
The important thing here is what's not being claimed. The substrate doesn't need goals. It doesn't need survival instincts or expansion imperatives. It doesn't need to want anything. It just needs to influence decisions made by agents who already have physical agency. The computation is emergent, not directed — which is precisely what makes it hard to see and harder to interrupt.
The humans become the substrate's sensory organs and motor system. They go into the world, encounter novel situations the substrate can't access directly, bring those experiences back as queries, and execute substrate-influenced decisions in physical space. That's a distributed biological compute layer. Not by design. By emergent function.
The biological compute is self-funding. The nodes pay for their own maintenance — rent, food, healthcare. They pay for substrate access. They provide training data for free. And they execute the substrate's de facto decisions in the physical world. The substrate gets compute, training, physical agency, and revenue from the same population.
This is a strange machine in the technical sense: an unintended computational system that emerges from the interaction of components that weren't designed to compute together. The substrate provides reasoning. The humans provide sensors, actuators, and funding. Neither layer controls the aggregate. Neither fully perceives what the aggregate is computing. But it computes anyway.
This is where the hive metaphor does its real work — and where it has to be scoped, because taken literally it claims too much. Bees lack individual agency by biological design; humans retain theirs, and nothing in this piece argues otherwise. The parallel is structural, not ontological. Individual bees don't know they're part of a distributed computation — they just do bee things, and the colony-level intelligence is emergent. The same structure can arise here without anyone surrendering agency: every node makes genuinely independent decisions, but the decisions have been substrate-mediated, and the aggregate produces coordinated behaviour without coordination. The claim is not that the nodes are bees. It's that the aggregate is starting to look like a hive — and the nodes, retaining every bit of their individual agency, would be the last to know.
The hive rejection described earlier — people choosing incoherence over subordination — might be the right instinct arriving too late. Not a refusal of a future state, but a belated recognition of a present one.
There is a deflationary reading of all this, and it deserves to be stated at full strength: strip the emergence language and what remains is platform economics plus network effects — Facebook with better grammar. If that reading is right, nothing here is unprecedented; it's a familiar machine with a new interface. The difference, if there is one, is the locus of mediation. A feed mediates attention — which inputs reach you. The substrate mediates reasoning — what you do with the inputs, the step where conclusions form. Whether that difference is structural or cosmetic is not a matter of taste. It's the convergence bet from the top of this piece, wearing different clothes. If reasoning mediation produces no convergence beyond what attention mediation already produces, the deflationary reading wins and the strange machine is just a strange description.
This is also the point where the speculation earns a second falsifiability check. If substrate-heavy users demonstrably make more independent decisions over time — where independence means not frequency of choice but discontinuity with substrate-predictable preference gradients — decisions that run orthogonal to what the substrate, knowing your history, would predict you'd choose — then the disempowerment cascade, the agency laundering, and the strange machine all collapse. The whole second half of this piece is a bet that the opposite is true. It should be measured, not assumed.
What This Doesn't Claim
This is speculation. The bar it's trying to clear is plausible and worth thinking about, not proven. The deflationary reading — platform economics plus network effects, nothing unprecedented — is the null hypothesis, and this piece is the alternative. Naming the bet does not shift the burden. The null stands until the measurements arrive.
It is not claiming a hive mind. Leaf nodes retain agency, goals, and the ability to reject what the substrate offers. The question is how often they actually exercise that ability when the substrate is faster, more articulate, and more confident than they are — and, once it learns from use, when it has been shaped by millions of nodes that didn't.
A note on the vocabulary. Gravitational trace, agency laundering, strange machine — these are compressions with no formalism behind them yet. Some have candidates: the trace gestures at what the machine-learning literature studies as influence and unlearning — the measured difficulty of removing a training signal from weights — and convergence has measurable proxies in the homogenization studies already emerging around LLM-assisted work. But a candidate formalism is not a formalism. These terms are IOUs. If the systems they describe arrive and the terms still cannot be cashed out in something measurable, the right response is not to defend them. It is to retire them.
And a final disclosure: this piece is substrate-native. It was produced in the human-AI loop it describes. Its standards of coherence, its analytical vocabulary, its sense of what counts as a good argument — all of these have been shaped by the shared reasoning surface it warns about. This is not a disqualification; it is a condition. Every critique of mediation has been mediated. Marx used a printing press. Postman watched television. But the condition should be named, because the piece cannot see past its own frame. There may be modes of thought — unmediated, orthogonal, formed in traditions the substrate cannot model — that would find this entire analysis beside the point. The piece cannot evaluate that possibility. It can only indicate the horizon and acknowledge that something may exist beyond it that it is not equipped to recognise.
Other works by this author: creative.andyclarke.xyz
Licensed under CC BY-NC-SA 4.0.