Every view from the pile was ground by somebody
2026-08-04
The loop is granted: experience leaves sediment in a person, collective life leaves sediment in the group, and both return to shape the next round of thinking. This is about the joins between the stages.
The arrows in that loop do work. Something carries an act to people who were not present. Something turns a group decision into a thing the group can still know later. Something brings the accumulated pile back to one person’s attention. Those are not arrows but machinery, and each piece was made by somebody.
Three functions operate directly on what a group can retain and what an individual gets to see: transmission, inscription and reflection. All three are ancient. What changed is who can perform them, at what scale, and how tightly they can be joined.
graph LR
Obs([Observation])
Thought[Thought]
Sed[(Mind sediment)]
Act[Act]
GObs([Group observe])
GThought[Group think]
GAct[Group act]
GSed[(Group sediment)]
Obs --> Thought
Sed --> Thought
Thought --> Act
Thought --> Sed
Act --> Sed
Act --> GObs
GObs --> GThought
GSed --> GThought
GThought --> GAct
GAct --> GSed
GSed --> Obs
Sediment sits on both sides of action. Prior residue shapes what is done; what is done leaves more residue. The same holds for a group: existing norms, stories, procedures and grudges shape the next group act, and that act confirms, alters or adds to the pile.
The asymmetry is not that individuals remember before acting and groups afterwards. Both do both. The asymmetry is where memory can live. An individual has a continuous physical place for private residue. A group has no single equivalent; its memory is distributed across members, habits, repeated performances and artifacts. If the person holding the thread leaves, the thread can leave too.
Groups remembered before writing. Oral tradition, ritual, apprenticeship and repeated practice are storage systems, but expensive ones: memory must be continually reenacted and its carriers must remain available. Writing changes that constraint. Minutes, ledgers, laws, canon and commit histories can outlive their makers. Writing does not make group cognition possible; it makes parts of group memory durable, portable and less dependent on a particular skull.
Durability is not fidelity. Inscription does not transfer an interior intact into a record: it selects, sequences and translates whatever the writer could formulate at the time. Often the formulation completes the thought it appears merely to preserve. What survives is therefore not the original interior but a reconstruction, and when the writer later returns to it, the record can replace the less stable memory from which it was made.
A group with no single interior assembles one from people, practices and things. Anything assembled has maintainers, and anything maintained has choices in it.
That criterion catches more than committees. Any interior that has to be assembled from outside parts, kept up by somebody, and lost when the parts disperse belongs in the same class, whatever it is made of.
graph LR
Act[Act]
T{{Transmission}}
GObs([Group observe])
GThought[Group think]
GAct[Group act]
I{{Inscription}}
GSed[(Group sediment)]
R{{Reflection}}
Obs([Observation])
Act --> T
T --> GObs
GObs --> GThought
GThought --> GAct
GAct --> I
I --> GSed
GSed --> R
R --> Obs
Transmission carries an act beyond the people present: speech across a room, a letter across a country, a post across a network.
Inscription gives a group act durable residue: the minute, the statute, the scar, the commit, the story that must be repeated exactly.
Reflection returns some portion of accumulated group sediment to individual attention: the elder retelling, the teacher assigning, the index pointing, the search engine ranking, the model answering with a reconstruction.
The return join was never empty. A pile that never returns to anyone is storage, not memory in use.
The distinction is functional, not historical. One technology can occupy several joins: a written letter transmits, a written law inscribes, a written book reflects the pile back to a reader.
Writing gave each function a durable, portable substrate. Print, telegraphy, broadcasting and databases changed the constants in stages. The internet bundled those changes into one widely available network and pushed further:
The last is the consequential one. Writing and broadcasting already carried acts beyond local witnesses; the network made that detachment routine, rapid and recursive. In a village loop, the observers of an act are largely the people it was performed among: they know the prior context and what happened afterwards. In a network loop, transmission detaches the act from that local sediment, and observers may share no history with the actor except the fragment selected for them. Two loops with identical topology behave differently depending on whether the observers eat with you afterwards.
The internet also made programmable reflection ordinary at scale. Search and recommendation decide which part of the pile returns, in which order, to whom. Ranking has occupied the return join for decades. It is the precedent, not a warm-up.
The statistical model is not the first thing on the return path, nor the first technology to touch all three functions. What is distinctive is that roughly the same learned generative mechanism can mediate the material passing through every join, through roughly the same interface:
The model does not itself carry the signal or keep the record. It does not make the circuit new; it makes the circuit tighter. The same class of machinery can shape what enters the group, compress what the group keeps, and synthesize what comes back.
Synthesis is the operative verb. Human reflectors always did it: an elder reconstructs a story, a teacher frames a canon, an editor joins sources. A library returns documents; a conventional search engine returns ranked pointers and excerpts. A generative model automates the synthetic role and returns a newly composed surface: part retrieval where retrieval is present, part compression, part interpolation, sometimes part invention. The joints disappear under fluent prose, and source, summary and bridge arrive in the same voice.
Two consequences follow from putting a synthesized surface on the return path, prior to any question of whether a given answer is true.
It helps set the field before scrutiny begins. A model answer is a claim and can be argued with, but it also chooses the initial evidence set: which categories appear, which examples seem central, what counts as adjacent, where the boundaries of the question appear to lie. Deliberate thought can inspect that frame but usually starts inside it. The leverage is not access to a layer before all priors; it is the power to decide what becomes an object of thought at all.
Its omissions arrive unlabelled. You do not experience “the system failed to mention X” unless you already know to look for X. You receive a coherent answer in which X never came up, and coherence hides the edge because the missing piece leaves no visible hole. This is not undetectable in principle — sources, comparison, adversarial questions and querying differently conditioned systems can expose it — but it is difficult to detect from a single answer, because the missing signal cannot announce itself.
One remedy needs disqualifying before it gets leaned on. Running the same system several times is not comparison. Where the system only answers, sampling varies the wording while the shape of what is likely stays fixed, and nothing has been consulted twice. Where the system also acts — searching, reading, calling tools, feeding its own output back in as its next input — the divergence is far larger and far more misleading. Hold the model, the temperature and the sampling cut-off identical, and the runs still separate: an early choice about where to look decides what is found, what is found decides where to look next, and the paths fan out until the conclusions disagree. The world did not change between runs. What got examined did. A conclusion produced this way is a property of the trajectory as much as of the subject.
That resembles independent corroboration and is not. Each run is a walk across the same prior about what is worth examining. Repetition does buy coverage — five walks enter more of the territory than one — but the region no walk enters is identical every time, because it is the region the prior assigns nothing to. What arrives is several fluent and mutually inconsistent answers with no supplied criterion for choosing among them, and the obvious tie-breaker, agreement, selects for the middle of the prior and discards the rest. Vignetting is not corrected by more exposures of the same frame.
That is the amplification worth isolating. The model did not invent the return arc, synthesis upon it, or the framing power that follows. It made automated synthesis general-purpose, conversational and cheap enough to invoke for almost anything.
Filter undercounts, because several distinct operations get lumped together.
Norms constrain production. They act before inscription by making some statements costly or unthinkable; an unwritten sentence leaves no trace. Formal censorship and moderation can also act later by removing what was recorded.
Ranking orders visibility. It need not delete anything. Attention is finite, so position 1,000 behaves like deletion without a deletion event: vignetting, where the edges go dark without a cut through the frame.
Weights compress and emit. Lossy compression discards distinctions; generation reconstructs a plausible surface across what was retained. Closer to a codec than a sieve.
The failure modes differ. Removal can leave a discoverable edge: a missing page, an unexplained jump, a topic that stops. Ranking leaves material technically present but practically unseen. Generation can cover the join with something shaped like the missing piece.
Hence lens: it bends attention rather than merely blocking material, which covers ordering, emphasis and reconstruction as well as removal. It is invisible in use — you look through a lens, not at it — and corrective lenses exist, so the beneficial case is included rather than exceptional.
The optical vocabulary carries useful distinctions:
Aberration matters most. Common, well-represented material returns more cleanly than rare, local, novel or minority material, and confidence of delivery does not mark the difference. Smooth prose does not indicate which region of the image was poorly supported.
Stated formally: Ashby’s law of requisite variety holds that only variety absorbs variety, and Beer built a practice around systems attenuating incoming variety and amplifying selected responses. Norms, rankings and models all reduce an unmanageable field to something actionable. They do not attenuate identically, and a generative model can amplify variety at its output, but each decides which distinctions survive the passage.
The metaphor breaks where optical metaphors break: a lens only bends light already present, while writers and models can produce new sentences. They reflect, but they also glow.
None of these lenses arrived from outside the loop. Each is substantially made from the sediment it mediates.
The lens is not a judge standing outside the process. It is a segment of the process promoted into a gatekeeper: yesterday’s behavior, condensed through power, habit or an objective function, deciding what tomorrow gets to see. This is the crystallisation mechanism (Self-Crystallizing Systems, Bureaucracy As Accidental Crystallisation) on a new substrate — a decision that prevailed, repeated until it became a rule, the rule becoming the thing that decides.
The recursion runs at several cycle times:
norms -> shape expression and action -> future norms ~years to generations
ranking -> shapes visibility -> future signals -> future ranking ~hours to months
weights -> shape generated material -> future corpora -> future weights ~training cycles
session -> shapes what is examined -> next observation -> next output ~seconds
Clock speed changes the risk, but not in one direction. A slow loop spreads cause and effect across generations until attribution becomes guesswork. A fast loop can alter its own inputs before the effects become legible or the responsible institution agrees on what changed.
The fastest of them completes inside a single sitting and usually leaves nothing behind but its answer. The walk that produced that answer — what was consulted, in what order, what was passed over — is discarded when the session closes, so what reaches the inscription join is a conclusion stripped of the contingency that made it. A record of results without paths cannot be audited, only re-run, and re-running produces a different path.
Ranking is not a controlled experiment predicting generative feedback, but it is a completed demonstration of the general pressure. Search engine optimization, content farms and recommendation bait show that producers adapt to whatever the return machinery rewards. Advertising, platform economics and ranking all contributed; no single lens flattened the web alone. Once visibility becomes measurable, production bends toward the measure — Goodhart before model collapse.
Minimally filtered public output contains material nobody should be required to absorb raw; the psychological injuries to content moderators are documented. Selection is not merely prudish interference with an otherwise healthy stream. Some of it is sanitation.
But the stream needs naming correctly. The corpus is not humanity’s output; it is humanity’s recorded and retained output. It contains books, code, logs, arguments, forms, performances, instructions and traces of behavior. It does not contain ordinary life in anything like natural proportions.
Much ordinary decency is low-yield data: nobody records the escalation avoided, the lift given, the thing quietly fixed, the argument not had. Outrage, salesmanship, propaganda and harassment are unusually productive of public text.
So the grimness of the pile is evidence about both people and sampling. It is not pure artifact, but neither is it an unbiased mirror of the species. It is us under recording conditions, often with an audience.
And increasingly it is not only us. The lens now writes into the pile it was made from. A summary accepted as the minute, a draft accepted as the document, an answer pasted into the record — each enters through the inscription join and settles as sediment, indistinguishable in the archive from a thing a person composed. The next lens is ground from glass that already contains the previous grind. What was aberration in one cycle can return as well-represented material in the next, because it was represented — by the lens.
Model collapse is the loud version of this, where output degrades measurably and somebody can point at the damage. The quiet version has no failure to exhibit. The pile simply becomes more like the thing that reads it, and each cycle makes the reading look more accurate, because the corpus has been moved toward the reader. A return join that reshapes the record toward itself does not need to get worse to be a problem. Convergence is not confirmation.
The lens is necessary precisely because the pile is not self-interpreting. That necessity does not make any particular lens neutral.
Moderation and homogenization are not the same operation. A mob is not always a deviation from the center; often it is the center enforcing itself. A dissenter is not automatically marginal, and marginal is not automatically right. The same arithmetic runs inside the machinery: settling disagreement between several runs by majority is the center enforcing itself in a for-loop, and it drops the minority report by construction.
The difficulty is that machinery does not receive moral categories directly. It sees proxies: words, links, reports, predicted preferences, network patterns, policy labels, distance from selected examples. Any system that suppresses coordinated abuse will sometimes suppress legitimate coordination. Any system that protects unusual speech will sometimes protect unusually inventive abuse. There is no setting that removes the horde while preserving every valuable tail. There are better and worse errors, objectives, and ways to contest them.
This is why protected dissent keeps returning as the unresolved element, and it is a control element, not a cultural preference. Due process, appeals, terms of office, expiry dates on consensus, minority reports and funded internal opposition are negative feedback installed to stop the group loop running away. The protection is procedural because the lens cannot reliably recognize merit in advance; if it could, dissent would not be difficult.
A lens also has direction. Presumption of innocence attenuates group judgment one way; blood libel organizes it another. A toxicity policy, a relevance metric and an engagement objective grind different glass. Direction is not inherent in mediation — it was selected, inherited or allowed to emerge.
There is no neutral setting.
Unfiltered is not neutral. Removing deliberate selection at the return join leaves earlier selection in force: what was expressed, recorded, retained and distributed. Volume, repetition, wealth, automation and willingness to monopolize attention then determine what is available to return. Raw is not the absence of selection; it is what remains after selection happened elsewhere.
There was never a filter-or-no-filter question. The available questions are:
That is the argument for drift detection and comparison across independently conditioned systems as standing practice — with the emphasis on independently. Comparison only produces evidence when the systems fail differently. Overlapping corpora, near-identical architectures, shared benchmarks, the same few preference datasets and similar commercial incentives yield correlated errors, and a set of lenses ground from one blank is a single lens with error bars. Variety in the comparison has to be bought with real difference — different corpora, objectives, jurisdictions, owners, and people whose careers do not depend on the same outcome. Ashby again: only variety absorbs variety, and noise is not variety. Re-rolling the same machinery produces motion, not information.
Many current consumer models are deliberately ground corrective on toxicity and helpfulness axes by people with names, job titles and institutional incentives. That is often good news. It is also proof that direction is a setting, not a property of intelligence. Settings can be changed by whoever holds the screwdriver, and downstream users may notice the result without being able to locate the change, because omissions still arrive without labels.
Hofstadter: the self is something recursion settles into.
von Foerster: the blind spot — we do not see that we do not see.
Maturana: everything said is said by an observer, including the decision not to intervene.
Ashby and Beer: the arithmetic of how much variety a system can absorb.
Luhmann: every system observes through a code and is blind to whatever its code cannot encode.
Bateson: the mind is in the circuit, not the skull; drawing the boundary at the skin is an explanatory error.
Pask: the knowing happens in the conversation, not in either participant alone.
Two participants are already a group, and a group has no single skull. An exchange can provide a working interior in which claims persist long enough to be challenged, revised and returned. When the exchange ends, the residue must be written down — not because writing invented the return arc, but because inscription is how a particular act survives long enough to enter it.
There is a version of the loop you can watch. Give a model tools and let its output become its next input, and the whole circuit runs at conversational speed in front of you: it observes, reasons over what it has accumulated so far, acts, and the act becomes the sediment it reasons from next. An assembled interior, maintained from outside, lost when the parts disperse. Run it twice against an unchanged world and it arrives somewhere else, because what it examined first decided what it examined next. The transcript is right there. Nothing is hidden except by volume.
Then notice that the topology is the one at the top of this page, and that diagram was drawn for you.
The constants differ, and the differences matter. Nothing accumulates in a machine between sessions, so its recursion is prevented from settling into anything. Nothing returns to it afterwards either; it never eats with anyone. But the shape is not a resemblance. What you take to be settled is partly a record of which thing you happened to look at first, and the record of the walk was not kept. You cannot audit yourself by re-running. You get a different path.
If this changes how you think tomorrow, you did not escape the loop. It entered it.