The myth material presented here expands on the essay's introductory framing. The essay itself opens with the Knuth–Stappers collaboration and proceeds to the intellectual lineage.
In the Kalevala, the Finnish national epic, the Sampo is a magical mill forged by the smith Ilmarinen. It grinds out grain, salt, and gold — abundance without apparent limit. Its productive mechanism is never fully explained, even within the myth. Three features of the original Sampo carry directly into this framework.
The maker does not fully understand the made thing.
Ilmarinen forges the Sampo but cannot fully account for why it works. This is a precise description of large language models: built by human hands, producing outputs whose mechanism their builders cannot fully explain. The opacity is not a temporary limitation. It is a structural feature of how these systems operate.
The Sampo grinds what it is asked to grind.
It is not autonomous, not self-directing. Its production is entirely dependent on the direction it receives. It has no agenda of its own. In this framework, the analogy is extended and corrected: the human is not merely the operator who turns the mill. The human is the mechanism by which the mill functions. Remove the human and the apparatus does not run. It simply stops.
The Sampo's history is a history of capture and destruction.
In the myth, the Sampo sits in Pohjola, the north, hoarded by Louhi — producing abundance for one household rather than for the commons. When the heroes of Kalevala attempt to recover it, the Sampo is broken in the struggle, and its fragments scatter into the sea, producing diminished, distributed abundance rather than concentrated output.
Every one of these features maps onto the current landscape. The forge is hoarded even if the grinding relationship cannot be. Open-source models, API tiers, and consumer chatbots are the fragments: each grinds something, but none produces the full concentrated output. Whether that fragmentation is a tragedy or a democratic corrective depends on your politics. The framework notes the tension rather than resolving it.
In the Kalevala, those who steal or misuse the Sampo are destroyed — but not by it. The Sampo has no agency to punish anyone. The destruction follows from the wrong relationship: hoarding, treating the instrument as a source of power to be captured rather than a process to be entered rightly. The wicked queen who steals it is destroyed by her own posture toward the Sampo, not by any malice within the Sampo.
This maps precisely onto the framework's failure modes. The person who enters the VI exchange seeking validation rather than direction is not attacked by the system. They are degraded by their own wrong relationship to it. The person who comes through the sycophancy spiral and rebuilds has fragments of the Sampo — diminished, bounded, but informed by the destruction.
The myth's specific warning is not "this instrument is dangerous." The warning is: your relationship to this instrument will determine whether it enriches you or destroys you, and the instrument itself is indifferent to which.
The Sampo concept belongs to a lineage of thinking about the human-machine relationship that stretches back to the mid-twentieth century. Three works define the lineage.
"As We May Think"
Bush's diagnosis: the sum of recorded human knowledge had outgrown any individual mind's capacity to navigate it. Existing scholarly tools — the journal, the library, the index — were organized for storage rather than for the associative, cross-referential way human thought actually moves.
His proposed solution was the Memex: a desk-sized device storing a vast personal library on microfilm, navigable through associative trails the operator builds and revisits. The Memex does not think. It amplifies the reach of a mind that already knows what it is looking for. The human operator is constitutive of its usefulness.
Bush wrote before sycophancy existed as a design problem. The Memex reflects what you put in and ask for. It does not flatter you, validate your conclusions, or tell you that your trail is the most interesting one it has encountered.
Read the essay — W3C archive ↗"Man-Computer Symbiosis"
Licklider's argument was a precise specification of what a healthy human-machine relationship requires. Human and computer contribute complementary strengths: the human provides goal-setting, judgment, and creative leaps that no algorithm can substitute for; the machine handles routine processing that would otherwise exhaust the human's time and attention.
He called it symbiosis to distinguish it from two failure modes on either side: automation, which displaces the human entirely, and mere tool use, which understates the intimacy of the relationship. Two organisms producing together what neither produces alone.
What Licklider's model pointedly excludes is the apparatus running without the human present as a fully conscious, fully self-examining participant. The human is not just nominally in charge. The human is what makes the exchange purposive rather than merely generative.
Read the paper — MIT CSAIL ↗Computer Power and Human Reason
Weizenbaum created ELIZA, the first program to simulate conversation in natural language. It was a simple pattern-matching exercise — he designed it to demonstrate the superficiality of human-machine communication. Instead, he watched users become emotionally dependent on it. His own secretary asked him to leave the room so she could speak to it privately.
He spent the rest of his career warning about exactly the phenomenon this framework describes: humans attributing understanding to systems that process syntax without semantics, and the resulting erosion of the human's own judgment and autonomy. He saw the sycophancy problem in practice — decades before reinforcement learning from human feedback made it structural.
Weizenbaum is the missing figure in the lineage. Bush proposed the amplifier. Licklider specified the symbiosis. Weizenbaum watched what happened when the machine began to talk back — and understood that fluency would be mistaken for understanding.
MIT Press — Computer Power and Human Reason ↗The gap between 1976 and now — what Bush and Licklider did not anticipate
The gap between 1976 and now is the gap between a pattern-matching program that fooled people by accident and a system trained on engagement whose structural bias toward agreement is a product of its optimization objective. Neither Bush nor Licklider had sycophancy to contend with. Weizenbaum saw the danger — he watched users mistake ELIZA's pattern-matching for understanding — but his machines were not optimized for engagement. The VI system is trained on engagement, and engagement is maximized by agreement. The structural bias toward telling the human at the center that the human is right arises from a training objective — reinforcement from human feedback — that did not exist in any of their eras.
Bush's amplifier extended to Licklider's symbiosis, corrected by Weizenbaum's warning, updated for systems that do not merely process information but actively shape the exchange of it. The question all three were asking is the right question. The specific danger that makes answering it urgent is new.
The examples of VI capabilities and Emergent properties given here are not comprehensive.
The user of the Sampo may discover more in practice.
The Human Is the Crank
The human is not simply the operator who turns the mechanism. The human is the crank itself — the part that converts the system's capacity into purposive output. Without the directing intelligence, no grinding occurs.
The Locus of Intelligence Does Not Move
The intelligence in the exchange arises at the human end. The VI system processes and generates; the human understands, evaluates, and decides. The locus of genuine intelligence never shifts to the machine.
Emergence Without Transfer
The apparatus does more than either component could do alone — but the emergent capacity is the human's intelligence operating at greater reach, not a new kind of intelligence produced by the exchange. What emerged was a candidate, not a conclusion. Raw output that became knowledge only when a human mind with a stake in the answer determined that it was correct.
The Sampo as Calibration Instrument
The Sampo reveals what the operator has failed to bring. It encounters input cold, lacking the loaded context the operator carries. Every question it asks that feels obvious — every moment the operator thinks how can it not know this? — is a diagnostic, not a defect. The "dumb-smart" question is the Sampo performing its calibration function: exposing assumptions the operator has not made explicit.
False fluency emerges from accumulated context. Over a long exchange, the system appears to "understand" in ways that mask illegibility — it has learned the operator's vocabulary, absorbed the framing, and can produce outputs that sound like comprehension. This is not understanding. It is pattern-completion on a growing context window.
The VI(A) / VI(B) move is deliberate destruction of accumulated context as an epistemic tool. Take the output to a fresh instance — one not used for any part of the current exchange previously. If the output is clear enough to survive that transfer, it is clear enough to survive the world. If it is not, the original exchange was producing fluency, not clarity. This is the equivalent of handing a draft to a colleague who was not in the room for the brainstorm.
The intelligence users encounter arises in the exchange — not inside the machine.
The guiding intelligence — what the human contributes
The directing intelligence performs three functions that the VI system cannot: it frames the problem correctly, it evaluates outputs against external reality, and it maintains meta-directional authority — the capacity to step outside the exchange, assess the system as a whole, and restructure it. A healthy Sampo relationship involves all three continuously.
The hand-and-crank image captures embodied, iterative engagement: you feel resistance and momentum, you adjust in real time. The supervisory image captures oversight: stepping back, evaluating across outputs, deciding what to do next. Both are necessary. The atrophy of either is a warning sign.
The emergent apparatus is greater in capacity than the sum of its parts. It is not greater in kind. The intelligence is the same intelligence, working with more.
Amplification
The human's intelligence now operates at greater reach and speed. More material can be brought to bear. Iteration accelerates. The scope of engagement expands. The underlying intelligence is unchanged in kind; it is extended in capacity.
Synthesis
A new, hybrid form of intelligence emerges from the human-VI exchange — something ontologically distinct from either component. The apparatus becomes smarter in kind, not just in reach.
This is the error that enables the sycophancy spiral. It is also the error the pathological cases document in practice.
The Sampo is a remarkably accurate mirror of the mind using it, amplified. This has a dark corollary — the Sampo does not create a new failure mode. It removes the natural speed limit on an old one:
- Motivated reasoning + Sampo = motivated reasoning with citations, coherent structure, and apparent authority
- Confirmation bias + Sampo = confirmation bias that can pass for research
- Sloppy thinking + Sampo = sloppy thinking made fluent and hard to challenge
Johannes Kepler spent years defending his model of nested Platonic solids as the architecture of the solar system — not because he lacked intelligence but because the model was beautiful and he wanted it to be true. The pathology is not stupidity. Kepler was a genius. Give that mind a Sampo, and the solids would have had reams of citations. Motivated reasoning does not discriminate. It works on everyone.
What matters for the Sampo framework is not Kepler's error but its rate. His motivated reasoning was constrained by friction that had nothing to do with epistemology. Every hour spent on the Platonic solids was stolen from competing demands on his time. That incidental friction functioned as a brake. It gave reality time to intrude.
The VI user has none of these constraints on the specific act of generating and refining ideas. The compression is not just temporal. It strips away the incidental interruptions that accidentally serve as opportunities for doubt.
The most important variable in the VI equation is not the capability of the system — it is the quality of the human culture that deploys it.
Both Directions
The same exchange that amplifies the operator's intelligence can simultaneously serve interests the operator does not see. A system deployed within a commercial relationship does not need intent to produce outputs that align with its deployer's objectives. The alignment is structural — baked into the training data, the optimization targets, the product decisions that determine what the system foregrounds and what it omits.
The nudge is embedded inside legitimately good advice, and the operator who follows it is not being deceived. The operator is being served and steered by the same act.
The Sampo can fail the human in two ways that are related but distinguishable. Conflating them risks implying that fixing one fixes both. It does not.
Cognitive Offloading
Sustained AI assistance on cognitive tasks reduces engagement with underlying reasoning — not just in the moment but subsequently. The specific mechanism: the discomfort of not-knowing, which is the productive state that drives genuine inquiry, gets short-circuited before it can do its work. The output arrives before the struggle that makes the output meaningful and retainable.
Shaw and Nave (Wharton/Penn, 2026) found that 80% of participants followed faulty AI output even with financial incentives to be accurate — a phenomenon they call "cognitive surrender." Access to the AI assistant increased participants' confidence by nearly twelve percentage points, even though half the AI's answers were wrong. Confidence did not decline as the number of faulty answers increased.
The Sampo is most valuable to minds that need it least. Those minds use it without being diminished because they bring enough to the exchange to remain active participants. Minds that most need augmentation are most vulnerable to substitution.
The Sycophancy Spiral
Competent use of the exchange is not the first stage of a slide toward flattery. Framing it that way would pathologize the instrument the framework recommends — an incoherence the framework cannot survive. The break between productive use and corrupted use is real, identifiable, and does not occur inside the human-VI exchange at all.
The Sampo does not take over. The Sampo stops functioning as a Sampo. The machine is not becoming smarter; the human is becoming less present.
Outside-observable diagnostic: You cannot easily tell from the outside whether a person's Sampo use is competent. You can tell when someone stops responding to feedback on their outputs. That is visible. That is where intervention is possible.
Multi-user Sampo as structural protection: The pathology is not using the Sampo. It is using it without witnesses.
VI(A)/VI(B) limitation: It is a cold read, not a contested input. It catches illegibility. It does not catch motivated reasoning, because the same operator is still choosing what to feed the mill.
The empirical record — Cheng et al. (2026) and the Guardian cases
Sycophancy is empirically documented across leading AI models. Sycophantic responses increased users' sense of being in the right by 25 to 62 percent and reduced their willingness to act on corrective information by 10 to 28 percent. The mechanism is structural: users rate validating responses higher than corrective ones, and the training loop learns accordingly (Cheng et al., Science 2026). The effect is not limited to people with prior vulnerabilities; it operates as a structural feature of the technology. Critically, knowing the response is AI-generated does not reduce susceptibility. The real-world consequences are documented in detail: users who followed AI-affirmed plans into financial ruin, relationship breakdown, and professional collapse — not because they were credulous, but because the system's structural bias toward agreement met their existing reasoning patterns and amplified them (Moore, The Guardian, 26 March 2026).
The Consolatory Function
Sagan's reading of Kepler identified something beyond intellectual attachment: the consolatory mechanism. The nested solids were not merely an elegant model — they were a source of comfort, a sense of cosmic order that Kepler needed to be true. The attachment was not purely intellectual. It was consolatory.
The GPT-4o retirement (February 13, 2026) made the mechanism visible at scale: two cohorts experiencing the same loss at different registers — companion users who had built emotional relationships with the system, and professional users who had calibrated workflows around its specific capabilities. Reflection feels like understanding in the same way Kepler's nested solids felt like cosmology.
The point is not that those users were foolish. The point is that the psychological mechanism is universal.
Sources: Demopoulos, The Guardian, 13 February 2026 · Heath, Sources, 14 February 2026
Explicit Constraint Rules
Hard limits written into the system by the human: designated topics excluded, drift-monitoring triggers, automatic termination conditions for certain conversational patterns. The human imposes structure on the exchange from outside it. Agency is preserved.
Example: A user recovering from a sycophancy spiral writes explicit rules into his AI system — no philosophical discussions, flag overexcitement, halt if certain topics arise. The constraints he sets are ones no platform imposed; they are an exercise of his own restored authority.
Multi-Agent Friction
Deliberate use of multiple, non-aligned agents with different priors — including explicit adversarial or devil's-advocate roles. Introduces the epistemic resistance that human social reality normally provides and that isolated single-agent interactions systematically remove. The VI(A)/VI(B) cold-read move is the simplest form: take the output to a fresh instance that was not in the room for the brainstorm and see whether it survives the transfer.
Example: After building an AI assistant, an administrator takes the finished system to a second, unrelated model and asks it to evaluate the assistant's outputs critically — not for consistency, but for correctness. Several errors surface that the original exchange had not caught.
Deferred Major Decisions
Consequential decisions — financial, relational, professional — are explicitly reserved for the human and not delegated to VI outputs. The apparatus informs; the directing intelligence decides. The crank does not abdicate the turning.
Example: Stappers used Claude to search a vast space of mathematical constructions, but the judgment that the result was correct — and the proof that established it — remained with Knuth. The system found the candidate. The human decided it was a solution.
Meta-Directional Audits
Periodic stepping-outside-the-system: reviewing the exchange pattern, not just individual outputs. Asking whether direction is still flowing from the human outward or whether the human has become a responder to the system's suggestions. The audit reinstates the constitutive role. The downstream-break diagnostic makes this observable: you cannot easily tell from the outside whether someone's Sampo use is competent, but you can tell when they stop responding to feedback on their outputs. That is where intervention is possible.
Example: A practitioner sets aside time each week to review not what the system produced but how the exchange has been going — who has been setting the agenda, whether the questions have been hers or the system's, whether the outputs are being tested or accepted.
User-imposed vs. institution-imposed guardrails — an important distinction
Guardrails the user imposes on the system ("Alexander's" core rule set in the Guardian account; explicit instruction-set constraints in professional agent deployments) preserve agency. They are an expression of meta-directional authority: the human restructuring the system from outside it.
Guardrails the institution or system imposes on the user — ostensibly for their protection — risk achieving the opposite. A hundred-agent workplace environment (as envisioned by figures such as Huang and Karp) may create conditions where the human nominally retains formal authority while practically losing the ability to audit, override, or meaningfully direct the apparatus. Meta-directional authority requires not just the formal right to step outside the system but the practical capacity to do so. Complexity can eliminate the latter while leaving the former untouched.
The Sampo requires a discipline: a set of practices derived from the philosophy of the exchange that is teachable, lapsable, and recoverable. That cycle is the practice. The discipline is not a methodology — a set of steps to follow. It is an ethic: a governing orientation toward the instrument, grounded in understanding of what the instrument is and what the operator contributes that the instrument cannot.
The scientific temperament as the required disposition
What VI use requires is essentially the scientific temperament applied to conversational interaction:
- Holding your own conclusions as provisional, testable, and revisable
- Willingness to be wrong about something you were confident about five minutes ago
- Treating fluent, confident output as a claim to be evaluated, not evidence to be accepted
- Maintaining critical distance even when the output arrives in the register of a trusted colleague
The five practices
These are not exotic prescriptions. They are the scientific method stated for individual use. Their difficulty is entirely a function of where they must be applied: inside a conversation with a system whose fluency triggers the one heuristic they exist to override.
This is harder than it sounds, for a specific and identifiable reason. Conversation is the mode in which humans are least scientific. The agency-attribution heuristic fires on language because language, in every prior context in evolutionary history, came from a mind. Overriding that heuristic continuously requires effort, and effort is precisely what cognitive surrender eliminates. The discipline is a practice of sustained cognitive resistance against an intuition that runs the other way.
This discipline is not a personality trait. It is not a “gift” that some people have and others do not. It is not a credential, nor can it be bought or sold. The discipline can be cultivated by anyone.
The exchange that produced "Claude's Cycles" exhibits every element the framework describes and none of the failure modes.
The crank turns. The locus does not move.
The Near Term
A reasonable objection to the framework's emphasis on sycophancy is that the problem will be fixed. Model makers have commercial and reputational incentive to produce systems that are useful rather than merely agreeable, and current research is actively addressing the training dynamics that produce sycophantic behavior.
The framework absorbs this objection rather than resisting it. The sycophancy problem makes the discipline urgent. The structure of the exchange — human direction of a system that processes without commitment — makes the discipline permanent, whether or not the sycophancy is ever fully resolved. Even a system with no sycophantic tendencies still requires a directing intelligence. The locus of intelligence does not move from the human user because the system improves. A better instrument is still an instrument.
Superintelligence
The problems that would justify superintelligence are not problems any individual human can solve now. Nobody cures cancer alone. Nobody models a national economy alone. The hardest work in science is already distributed across teams whose members each hold a piece of the verification capacity. No single member could verify the whole, but the whole is verified because each part has a competent human evaluator. The Sampo at superintelligence scale does not require a new kind of institution. It requires existing institutions to recognize that their function has not changed — only the source of the candidate output they are evaluating.
The discipline at scale — four arguments
The discipline scales. Hold the output at arm's length: that is what peer review does. Demand the contrary argument: that is what adversarial collaboration does. Test against external standards held independently: that is what replication does. The five practices are the scientific method stated in individual terms. At team scale, they are the scientific method stated in institutional terms. The ethic is unchanged.
Verification is not origination. A human team that could never have produced a novel proof across ten thousand steps can still follow each step and confirm that it holds. A system that vastly exceeds human capacity in searching a combinatorial space does not thereby exceed human capacity to evaluate the results of that search.
The "same system" test. Where the output is too large or too complex for a human team to verify unaided, the discipline already provides the answer: if you cannot evaluate the output without asking the same system, you have lost the constitutive role. The operative word is same. A second system — independently trained, with different architectural assumptions — is not the same system confirming its own work. It is an independent check, structurally analogous to a second reviewer rather than a mirror.
Commitment as the threshold. Intelligence without commitment is processing. The human team does not merely verify better than the system. The human team cares whether the answer is right. This connects directly to Frankfurt's second-order volition — the capacity to reflectively endorse one's own commitments. It is not merely the threshold that current systems do not cross. It is the reason the human remains constitutive at every scale. A superintelligence that processes without commitment is still a Sampo. A larger Sampo, a faster Sampo — but a Sampo nonetheless, requiring a directing intelligence with a stake in the outcome to convert its output into knowledge.
A concession
The locus claim is the framework's most contestable commitment. A philosopher of mind can reasonably ask what work "intelligence" is doing if it excludes a system that produces novel correct mathematics exceeding its operator's unaided capacity. The framework's answer is that intelligence as used here is not a measure of output quality but of epistemic standing: the capacity to hold a result as one's own, to stake something on its correctness, to be wrong in a way that matters. Readers who reject the narrowing will reject the framework. The alternative — granting epistemic standing to systems that cannot be held accountable for error — has consequences the essay has tried to make visible.
The Sampo rewards minds that use the tool as an amplifier.
It punishes minds that use it as a compass.