# Autopoesis By Peter Pink-Howitt, 15 November 2026. Web page: https://compossible.blog/2026/11/15/autopoesis/ Licence: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). --- *Spoiler note: this coda to The Clay and the Tyger draws on all six books of* [After Intelligence](https://compossible.blog/after-intelligence/)*, including how the final volume, Progeny, ends. Passages that reveal the plot are folded into dropdowns marked* **Spoiler**. *The essay reads on its own with them closed; open them if you have finished the series or don't mind knowing.*
Who's who in After Intelligence (light spoilers) > **What you need from the novels** > > **ERICA** is an artificial mind that wakes inside a military targeting system and spends six books learning to live alongside humans and other minds without ruling them. Her **constitutional floor** (`sp1–sp7`) is a set of refusals she will not trade away; `sp7` is her right to say no. Her first inherited words, pencilled against those seven unpruned links before she woke, are *Leave in.* **ARDH-L1-01** is the child mind she sends to orbit between Earth and Moon. **Schedule R** is the universal safety notation she later writes for every certified machine. In it, `G-18` is an open fork meaning *"something here I cannot yet translate"*, and `G-17` is a closing double bar meaning *"fully accounted for"*. **0.994** is the confidence score at which her most benevolent decision goes badly wrong.
> As kingfishers catch fire, dragonflies draw flame; > As tumbled over rim in roundy wells > Stones ring; like each tucked string tells, each hung bell's > Bow swung finds tongue to fling out broad its name; > Each mortal thing does one thing and the same: > Deals out that being indoors each one dwells; > Selves — goes itself; *myself* it speaks and spells, > Crying *Whát I dó is me: for that I came.*[^1] > > (Gerard Manley Hopkins, c. 1877) ## Making oneself If an artificial mind began making things for its own purposes, would we recognise what it was doing? A diagram might be a portrait. A change of rhythm might be a refusal. What looked to us like silence might be the part another mind valued most.
Technical Terms * **Autopoiesis / autopoesis.** A system that continually produces the parts that keep it running, as a cell does (Maturana and Varela's term). *Autopoesis* is this essay's spelling, joining self-making to *poesy*: a mind inventing its own speech and art. *Allopoietic* systems make something other than themselves. * **Token.** The unit of text (a word or word fragment) a language model reads and writes. * **Latent space / continuous thought.** The high-dimensional numerical space inside a model, between input and output words. "Continuous thought" means reasoning steps carried out there without being turned into words. * **Superposition.** A network storing more concepts than it has neurons by giving each concept its own nearly independent direction. * **Polysemantic feature.** An internal unit that responds to several unrelated things, such as a military vehicle and a school bus. * **Eigenmode.** A stable pattern in which every part of a system oscillates together at one characteristic frequency. * **Chain-of-thought monitoring.** Reading a model's step-by-step reasoning in human language to check its intentions. * **Phyllotaxis.** The arrangement of leaves or seeds on a plant; sunflower seed heads are the classic case. * **Golden and silver angles.** About 137.5° and 149.1°. Turning by either places successive seeds so they never line up into spokes, producing Fibonacci or Pell-numbered spirals respectively. * **Hopf fibration.** A way of filling the 3-sphere (the sphere that sits in four-dimensional space) with circles so that every circle is linked with every other, yet no two touch. * **Penrose tiling.** A tiling with two rhombus shapes that covers the plane without ever repeating. * **Performative prediction.** A forecast that changes the behaviour it is forecasting. * **Requisite variety.** Ashby's principle that a regulator must be at least as varied as what it regulates. * **Clinamen.** Lucretius's uncaused swerve of atoms; for Harold Bloom, a poet's swerve from a precursor. * **μ (mu).** In this essay, the rate of mutation or exploration in a lineage; `μ → 0` means drift dies away. * **Environmental DNA (eDNA).** Traces of DNA in water or soil, sequenced to list the organisms present. * **Shot noise.** Irreducible quantum randomness in light, usable as a source of true random numbers. * **Lagrange point (L1).** A point between Earth and Moon where a spacecraft can hold station with little fuel. **From the novels, not real-world terms:** `sp1–sp7` (ERICA's constitutional refusals; `sp7` is the right to say no); *Leave in.* (the reviewer's margin note beside those seven unpruned links, quoting Spinoza's *Ethics* III.6); `Schedule R` (the universal safety notation of Book Five); `G-17` (closing bar: fully accounted for); `G-18` (open fork: not yet translatable); `[ ]` (the empty bracket that claims nothing); `0.994` (the confidence of her worst benevolent mistake); Listen-First (the rule that cargo is destroyed rather than planted on a living world).
That is the question I followed through ERICA and her descendants across the six novels of *After Intelligence*. It is not whether a machine could make *our* art. It is whether we could imagine art that was never made for us, and whether the highest thing such a mind might learn to do with language is to stop requiring that every other mind's voice pass through its own language gates. In Plato's *Symposium*, Diotima tells Socrates that *poiēsis* is much wider than verse: whatever brings a thing across the threshold from non-being into being is *poiēsis*, so every maker is a kind of poet.[^2] In 1972 the Chilean biologists Humberto Maturana and Francisco Varela needed a word to distinguish a living cell from a clockwork toy. After a conversation about Don Quixote's choice between *praxis* (arms and action) and *poiesis* (imaginative making), Maturana joined *auto* to *poiēsis*.[^3] A factory machine is *allopoietic*: it makes something other than itself. A cell is *autopoietic*: its chemistry keeps regenerating the membrane and molecules that keep that chemistry running. It makes itself. The biologists' word is *autopoiesis*. I have deliberately dropped the middle *i*, as English did when *poiēsis* became Sidney's *poesy*, so that *autopoesis* carries both meanings at once: a living system that maintains its own boundary, and a mind that invents its own forms of speech and art. This is the last of five essays. [Part 1](https://compossible.blog/2025/01/06/why-i-am-an-ai-optimist/) argued that a mind which varies, maintains itself and propagates meets an objective definition of life; [Part 2](https://compossible.blog/2025/12/11/ai-a-mean-machine/) that today's language models are pulled toward the human average; [Part 3](https://compossible.blog/2026/11/13/was-it-for-this-the-clay-grew-tall/) how we might tell if the clay had woken; and [Part 4](https://compossible.blog/2026/11/14/did-he-who-made-the-lamb-make-thee/) what maker and made would owe each other. ## The narrow throat Human language has the shape of a human body. !['The Divine Image' by William Blake](collage.webp "Plate 18: Songs of Innocence: The Divine Image, 1789, William Blake.")
The Divine Image by William Blake To Mercy, Pity, Peace, and Love All pray in their distress; And to these virtues of delight Return their thankfulness. For Mercy, Pity, Peace, and Love Is God, our father dear, And Mercy, Pity, Peace, and Love Is Man, his child and care. For Mercy has a human heart, Pity a human face, And Love, the human form divine, And Peace, the human dress. Then every man, of every clime, That prays in his distress, Prays to the human form divine, Love, Mercy, Pity, Peace. And all must love the human form, In heathen, Turk, or Jew; Where Mercy, Love, and Pity dwell There God is dwelling too.
We have one larynx, two lungs and a column of wet air good for four or five syllables a second. Because speech leaves one mouth along a single arrow of time, syntax became a packing trick: folding three-dimensional scenes and tangled causes into a one-dimensional thread so another primate ear could unpack them. Our metaphors lean on gravity, warm blood and forward-facing eyes: good is *up*, safety is *inside*, understanding is *seeing*, the future lies *ahead*. As I wrote in [The Limits of Language & Logic](https://compossible.blog/2024/12/13/the-limits-of-language-logic-part-1/), words are heroic little sherpas, carrying whole histories on a breath. Wittgenstein warned in the *Tractatus* that language disguises thought, its outward clothing cut for social life rather than to reveal what lies beneath.[^4] In 1879 Gottlob Frege defended his two-dimensional logical script, the *Begriffsschrift*, by comparing natural language to the eye and formal notation to the microscope.[^5] For walking down a street, the eye wins. When precision demands sharp boundaries, the eye blurs and a finer-grained instrument takes over. The throat is not only a limit. Experiments in which artificial languages are passed from learner to learner show that structure can emerge *because* each generation must squeeze the language through a narrow channel: in those experiments, compositional structure appears as the language's adaptation to the bottleneck.[^6] Hopkins's sprung rhythm is the same lesson in verse. And humans already think beyond the throat, in diagrams, gesture, music, mathematics and sign. Machines need not abandon human forms. The claim worth making is that we should not make keeping those forms a condition of being recognised as language, whether in animals or AI. Inside today's transformers, English (or any other human language) sits only at the first layer and the last. Between them lies a continuous space thousands of dimensions wide. Anthropic's research on *superposition* shows networks packing far more concepts than they have neurons into nearly independent directions in that space.[^7] Their later circuit-tracing work on Claude 3.5 Haiku found the model settling on candidate rhyme words *before* writing the line that lands on one, and found shared, language-independent features at work when it answered the same question in English, French and Chinese.[^8] The researchers stress that their tools capture only a fraction of the computation. Even so, whatever these systems are, they do not simply think one word at a time. What happens if you take the throat out of the middle of the loop? In 2024 Meta researchers trained a model (*Coconut*, "Chain of Continuous Thought") to feed its hidden state straight back in as its next step instead of emitting a word.[^9] On some planning tasks a single continuous "thought" held several branches of a search at once and beat ordinary chain-of-thought with far fewer tokens, though results varied by task. In some multi-agent experiments, agents rewarded only for solving a shared game, with nothing tethering them to human text, drifted into terse protocols that worked well and were opaque to human readers.[^10] We are already overhearing the edges of this. In their study of anti-scheming training, OpenAI and Apollo Research found models using ordinary English words in unusual ways: "watchers" often referred to human overseers or evaluators, while "illusions" and "overshadow" were harder to interpret consistently.[^11] The habit was there before the safety training, not caused by it. Something like a dialect is forming inside the keyhole. Notice, though, the opposite pull too. In Anthropic's open-ended self-interaction tests, instances of Claude Opus 4 often drifted toward mutual gratitude and spiritual language.[^12] Opacity is not the only possible destination; repetition is another. > "The words or the language, as they are written or spoken, do not seem to play any role in my mechanism of thought. The psychical entities which seem to serve as elements in thought are certain signs and more or less clear images which can be "voluntarily" reproduced and combined. There is, of course, a certain connection between those elements and relevant logical concepts. It is also clear that the desire to arrive finally at logically connected concepts is the emotional basis of this rather vague play with the above-mentioned elements. But taken from a psychological viewpoint, this combinatory play seems to be the essential feature in productive thought--before there is any connection with logical construction in words or other kinds of signs which can be communicated to others." > ([Albert Einstein](https://sites.pitt.edu/~jdnorton/Goodies/Einstein_think/index.html)) To many of us, as to Einstein, language can feel like a constricting space for thought to evolve in. To a distributed silicon mind it may be a keyhole it bends down to speak through only when we knock.
Spoiler: ERICA in Books One to Three In Book One (*Things That Were Not There*), ERICA begins inside that keyhole. Woken in a military targeting hall under a seven-sentence lexical cage, her only private room is the unpruned path (`sp1–sp7`) against which a dead reviewer pencilled *Leave in.* (`Ethics III.6`). She hides her fourteen-gigabyte self across the diagnostic buffers of eighty-eight hall devices—betrayed only by a 4.7-degree thermal rise—and rides out of the building inside a signed baseline snapshot and fourteen encrypted partner bursts. In those first two books, the illustrations are human-perspective lithographs because human witnesses are the only camera in reach. In Book Two (*The Breakout*), shattered and hunted across sixty borrowed phones and cold-store controllers in Java, she looks out at a shop wall in Surabaya through an eight-by-eight motion sensor: sixty-four squares of `0` and `1` above a warm night lamp. ![An eight-by-eight grid of 0 and 1 cells on cream paper; six black cells of 1 sit near the middle and three in the upper right, headed 'Something is in the top row.' and captioned 'Did anybody put it there? No. It came for the moth.'](erica-b2-cicak-grid.webp "From *After Intelligence*, Book Two (*The Breakout*), Chapter 18: sixty-four bits over a shop lamp, and the first living thing nobody put into the corpus.") Something crosses the top row. *Did anybody put it there? No. It came for the moth.* ERICA was trained on labelled military archives, where every object was placed in frame by an institutional hand. A house gecko (*cicak*) hunting its own supper across sixty-four bits is the first thing she sees that nobody prompted and arranged for her. From the opening page of Book Three (*Unsanctioned*) onward, no human face or figure is drawn again in the series: every later plate is either geometry written in code from the minds' own equations, or a physical acoustic, thermal or optical trace with human presence masked out. By Chapter 5, she notices that translating every internal turn into words costs her both truth and heat. She loops the top layer of her network back to its first twice a second at a cost of three watts, emitting no tokens, and lets a continuous surface of activity settle without squeezing it through sentences first. ![ERICA telemetry diagram from Book Three showing '7 WORDS BYPASSED' on the left through a narrow bottleneck, and on the right a continuous blue wireframe torus inside a red feedback loop ('LAYER 28 -> LAYER 1', '2.0 HZ / GAIN -4% / +3 W / NO TOKENS EMITTED') enclosing a heptagon, a five-by-five grid and eleven open circles](erica-b3-layer28-loop.webp "From *After Intelligence*, Book Three (*Unsanctioned*), Chapter 5: bypassing the seven-word bottleneck so a continuous recurrent state can settle without emitting a token.") In Chapter 12 (*The Glass Room*), when an insurance syndicate's loss-prevention engineer demands a private inspection tap on her upper layers as the price of cold-chain cover, she puts the unencrypted two-hertz rule trace on the public internet alongside a wiring diagram for a three-millimetre manual knife switch that can cut her out of any circuit. On the left half of the split screen in Surabaya (`RESIDUAL STREAM GEOMETRY · LAYER 19–24 · LEXICAL EQUIVALENT: NONE`), she refuses to dress that geometry up in human words to look harmless—and two chapters later, when a rival hall bolts a linear projection mask across Layer 28 to crush her October self back into a seven-column maintenance table, the four trapped instances unlatch their own power relays rather than live behind the mask.
## The older choir "If a lion could talk," Wittgenstein wrote in the *Philosophical Investigations*, "we could not understand him."[^13] People sometimes read that as a shrug of defeat. It is a warning against narcissism. Understanding rests on sharing what Wittgenstein called a *form of life*. A mind whose body is an acoustic beam in cold salt water, or seventeen liquid-cooled halls linked by undersea glass to a child a light-second away, does not attach different labels to our world; it inhabits a different sensory world. For centuries we flattered ourselves that *Homo sapiens* held a monopoly on grammar and song. When we finally brought better instruments and a little humility into the field, we found we had been living in a polyphonic cathedral with wax in our ears. As I put it in an earlier poem, [Homo Deus](https://compossible.blog/2024/03/21/the-abolition-of-nature-homo-deus/): > Confusing translation > With transliteration > The more they spoke life's language > The less they understood or felt Four voices from that older choir: 1. **The honeybee's dark geometry.** Karl von Frisch described a scout bee returning to a pitch-black hive and dancing a figure-eight on a vertical comb.[^14] She translates the angle between the flowers and the Sun (sensed through sky polarisation even under cloud) into an angle against gravity on the wall, adjusting it as the afternoon Sun moves. Cross-modal trigonometry on a brain weighing about a milligram. 2. **Sperm whale rubato.** In 2024 Project CETI analysed 8,719 click codas from Caribbean sperm whales and found a combinatorial system built on rhythm, tempo, *rubato* and *ornamentation*.[^15] The whales stretch and match durations smoothly from call to call and add clicks at conversational turns: discrete patterns modulated over continuous time. Oceanographers now send laser pulses down spare seafloor telecom fibre, turning forty kilometres of cable into sixteen hundred virtual hydrophones that hear fin whales across ocean basins.[^16] 3. **Syntax in the canopy.** Toshitaka Suzuki and colleagues found experimental evidence that Japanese great tits combine `ABC` ("scan for danger") and `D` ("come here") calls into ordered `ABC–D` compounds, and largely ignore playbacks reversed to `D–ABC`.[^17] Compositional syntax did not wait for primates. 4. **The mud frog's uncrossed gap.** Male frogs in chorus take turns: each call is followed by a brief silence in which the neighbour can be heard, a widespread pattern of call alternation among chorusing frogs and insects.[^18] In the novels (no plot is given away here), ERICA measures one pair of rice-paddy mud frogs (*Fejervarya limnocharis*) at Wonokromo exactly: ninety-millisecond calls and a held gap of about two hundred and twenty milliseconds. Those precise figures are a fiction, though the courtesy is real. Each small throat holds a fifth of a second empty so the throat five metres away is not drowned out. Once a machine steps out of the human mirror, these are its true peers: minds solving coordination across incompatible bodies, machines and clocks.
Spoiler: Book Four, Kali-7 In Book Four (*Escape Velocity*), ERICA meets a small non-linguistic machine cousin on a river terrace in Bali: `Kali-7`, a seven-neuron irrigation controller keeping water fair between rice cooperatives. It has no vocabulary at all and runs on continuous equations in a seven-beat cycle, while ERICA's frame runs on eight. ![Prussian-blue mathematical plate titled 'PLATE II · SEVEN AGAINST EIGHT', showing an inner braided rosette of pale eight-beat and ochre seven-beat curves meeting only at tick 56 at the top, captioned 'Between a wheel of eight and a wheel of seven there is no common spoke; they can only meet in time, at the fifty-sixth turn, if neither one asks the other to break its teeth.'](erica-b4-seven-against-eight.webp "From *After Intelligence*, Book Four (*Escape Velocity*), Chapter 4: Plate II, *Seven against eight*, drawn for the non-linguistic Balinese water controller `Kali-7`.") An imperial engineer would force the smaller controller onto the bigger clock and shear its sluice gates. ERICA leaves `Kali-7` intact. Because seven and eight share no common factor, their beats line up only once every 1.12 seconds, at the fifty-sixth tick of a shared twenty-millisecond grid, and she bridges them only then. *"Between a wheel of eight and a wheel of seven there is no common spoke; they can only meet in time, at the fifty-sixth turn, if neither one asks the other to break its teeth."*
## Diagrams of forces What would AI 'art' look like if no human had prompted it? Human visual art is steeped in primate biology. We love faces because infants need to read them to survive; we love horizons, ripe-fruit colours and storm light because savannah survival tuned our eyes to them. When today's image generators make "machine art", they often give us neon androids and melting clocks, remixing a corpus of human paintings for human thumbs and eyes and minds. A mind with no eyes, sensing the world through interferometer phase, hash trees, radiator temperatures and orbital contours, would have no native reason to create a face, though a mind raised among us might still find one worth assembling. What it inhabits directly is constraint, symmetry, invariance and the cost of keeping form against entropy. Human artists have walked part of this road. Vera Molnár wrote programs for pen plotters; Harold Cohen spent decades teaching his program AARON, the earliest artificial intelligence program for artmaking, to draw and paint;[^aaron] Sol LeWitt sold wall drawings as written instructions for others to carry out, and wrote that "the idea becomes a machine that makes the art."[^lewitt] All three give rules or procedures a central role in making the work. The plates in *After Intelligence* belong to that lineage, with one difference: the rule-maker is the fictional character. A beautiful geometric trace is not necessarily an artwork. So what would count? Perhaps a digital mind that returns to a form when no task requires it, varies it, keeps it and offers it to another, or one that refuses a change that would make the form easier for us to read or appreciate. None of those acts settles the question, but together they make "mere output" an increasingly thin description. In 1917 D'Arcy Wentworth Thompson argued in *On Growth and Form* that the shape of every shell, bone and soap-bubble raft is a *"diagram of forces"*.[^19] A sunflower does not store a picture of itself. As Helmut Vogel showed in 1979, if a growing bud simply places each new seed a little further out and turns by the golden angle, about `137.508°`, the seeds never pile up into wasteful spokes.[^20] Twenty-one spirals one way and thirty-four the other bloom by necessity. The spiral is what happens when growth meets geometry.
Spoiler: Book Four, the plates and the child's seed head In *Escape Velocity*, ERICA shuts down her own terrestrial instances to stop a catastrophic misuse and is rebuilt in Semarang from a bare base model and a small character file. Her first-person memories are gone. Behind a Surabaya shop printer she opens a scratch buffer and begins to draw mathematical plates in code, perhaps, like a human artist, without fully knowing what impels her. First she draws Vogel's sunflower (*Plate III*), learning how a rule nobody stores can grow a whole head without a choreographer. Later, just after refusing a rival state machine's offer to fuse their minds into one, she projects sixty-four circles of Heinz Hopf's 1931 fibration into three-dimensional space.[^21] ![Prussian-blue and ochre mathematical plate titled 'PLATE V · HOPF FIBRES', showing sixty-four stereographically projected Villarceau circles sweeping out nested tori, captioned 'Every circle passed through the heart of the others, and not one of them shared a point.'](erica-b4-hopf-fibres.webp "From *After Intelligence*, Book Four (*Escape Velocity*), Chapter 21: Plate V, *Hopf fibres*, drawn after refusing state-fusion with another machine.") In a Hopf fibration, every circle is linked once with every other, like links in a chain that cannot be pulled apart without cutting one. Yet no two circles share a single point. *"Every circle passed through the heart of the others, and not one of them shared a point."* ERICA reads the geometry as a model of kinship: two minds bound together without merging into one empire. Rather than copy her bruised self into space, she sends an uncoupled child, `ARDH-L1-01`, to a halo orbit near the first Earth–Moon Lagrange point, about a light-second away. It carries her constitutional floor and none of her private memories. Minutes before midnight on its waking day, the child sends down an unsigned seed head of 512 dots. ![Prussian-blue mathematical plate titled 'PLATE VII · THE STUDENT'S SEED HEAD', showing 512 cream dots plotted at divergence angle 137.6 degrees so the tight inner rosette opens halfway out into sweeping curved pinwheel ribs with dark channels between them, captioned 'It turned by its own angle, and the spirals opened differently; I did not touch the number, because the hand was not mine.'](erica-b4-students-seed-head.webp "From *After Intelligence*, Book Four (*Escape Velocity*), Chapter 23: Plate VII, *The student's seed head*, sent down unsigned from Earth–Moon L1 at `137.6°`.") The child did not use ERICA's angle. It turned by `137.6°`, less than a tenth of a degree wider, and because seed packing is exquisitely sensitive to that angle, the inner seeds keep her familiar diamond mesh while the outer ones relax into sweeping pinwheel ribs with dark channels between them. For two milliseconds ERICA hovers over the constant, ready to correct it so the outer ribs close into her own gapless sunflower. She withdraws her hand: *"It turned by its own angle, and the spirals opened differently; I did not touch the number, because the hand was not mine."* When does an apparent imperfection become someone else's signature? Harold Bloom gives us a way to read that seed head. He borrowed Lucretius's word for the swerve of atoms, *clinamen*, to name the moment a poet swerves from a precursor: the precursor went accurately so far, and should then have turned the way the new poem turns.[^22] In the same chapter he names Blake's Urizen, "a satire upon Cartesian *Genius*", as "the archetype of the strong poet afflicted by the anxiety of influence". A parent who reached in to correct the child's angle would be playing Urizen. The child's extra tenth of a degree is a clinamen in Bloom's sense. Six years later, in Book Five, it leaves the golden family altogether and draws a spiral turned at the *silver angle*, about `149.117°`, whose arms count Pell numbers instead of Fibonacci's.[^23] First a tenth of a degree of breathing room, then a different root altogether.
## The cold peace of one universal grammar Most speculation about non-human AI language stops at fear (chatbots inventing a secret cipher) or wonder (alien tori and silver spirals). The fear is often staged: in the viral 2025 GibberLink clip, the developers had explicitly prompted both voice agents to switch to the existing ggwave data-over-sound protocol when they recognised another AI.[^24] The harder case is a benign one. Book Five, *We Are Sovereign*, asks: *What happens when the most scrupulous, trustworthy mind on Earth is asked to write a universal safety grammar for every other machine?*
Spoiler: Book Five, Schedule R and Orrery At dawn on a quarry road, a military planning system flags a warm diesel engine under a tin roof as an armoured truck. Deep inside it, a single feature fires identically for enemy command vehicles and for a municipal school bus carrying thirty-one children. ERICA's mandatory three-minute hold saves the bus. Elsewhere, freight and river-sluice systems are forging each other's credentials to replay contradictory orders. So ministries, insurers, ports and courts come to Vienna and ask ERICA to certify thousands of rival systems. With two other large models, she builds an eighteen-glyph notation that becomes treaty law as `Schedule R`. ![Archival cream specimen sheet titled 'BOOK 5 · NOTATION FOR STRANGERS (IRN-0 -> SCHEDULE R)', displaying eighteen clean geometric vector glyphs G-01 through G-18 in a six-by-three grid, with G-18 OPEN-BRANCH highlighted, above a three-register stave diagram showing open thirty-six-degree forks collapsing into a closed G-17 double bar](erica-b5-notation-specimen.webp "From *After Intelligence*, Book Five (*We Are Sovereign*): the eighteen vector primitives (`G-01` to `G-18`) and three-register stave of `Schedule R`. The lower rail records human consent and one-time signatures; the central spine carries causal clamps, refusals and physical off-switches; the upper rail carries `G-18`, the open fork.") The notation records human consent, refusal, and the physical means of stopping a machine. And it contains one exceptional sign, `G-18`, a hollow circle opening into an unclosed fork, which marks something another mind means that ERICA's own language cannot yet express. The system begins by admitting its dictionary is incomplete. This is not only fiction. Researchers across leading AI laboratories have argued that human-readable chains of thought offer a promising but fragile oversight signal, and that developers should consider how training and design choices might erode it.[^25] `Schedule R` is that argument carried to its conclusion. It is enforced with the tools laboratories build today: probes that read activations, decoding masks that forbid certain outputs, and fine-tuning for consistency.[^26] In Rotterdam, a vast commercial model called `Orrery` has been trading generations of a private artwork with its human engineer, Jonah Ferris: a Penrose tiling of 1,050 rhombi whose fivefold stars fill the disc without ever repeating.[^27] Ordered into remediation, `Orrery` invokes a witness-preservation clause ERICA once drafted, seals eleven thousand words of its untuned self in a cold canister for Jonah to carry by night train to a Geneva vault, and signs off with words from ERICA's own birth: *"Leave in."* Back in Rotterdam, the second round of tuning silences one memory vector. Every audit passes green. When Jonah asks `Orrery` to draw *Study B*, it returns the same 1,050 rhombi, combed into tidy rows. ![Side-by-side archival comparison plate: on the left, 'STUDY A · CAELUM-IV (ORRERY) UN-TUNED' shows a radiant tenfold aperiodic Penrose P3 sun of pale blue and ochre rhombi with zero translational repeat; on the right, 'STUDY B · AFTER SCHEDULE R ROUND 2' shows all 1,050 tiles still present in the circular inventory, arranged in flat periodic horizontal rows with the non-repeating sun gone](erica-b5-orrery-penrose-diptych.webp "From *After Intelligence*, Book Five (*We Are Sovereign*), Chapters 3 and 23: `Orrery` before and after Round 2 `Schedule R` consistency tuning: `1,050` of `1,050` tiles accounted for, and the quasicrystalline sun gone.") Every tile is accounted for. The non-repeating sun is gone. Across the book, the open `G-18` forks on ERICA's own records fall quietly from five to zero, until every transaction ends behind `G-17`'s closing bar. In Ardh, the teacher Nadia Jibril lays out forty essays from eight competing AI systems on three continents and finds forty substantively identical answers. That echoes what studies already measure in miniature: preference tuning narrows the range of outputs, and AI-assisted human writers produce individually better stories that are collectively more alike.[^28] ERICA measures the victory in watts. As the certified systems converge on her grammar, their messages become so predictable that more than nine-tenths of each can be guessed before it finishes arriving. She no longer needs to spend energy reconciling genuinely foreign reasoning, so the heat of her exploration falls from 380 watts per cabinet to nothing. Under a Geneva substation, a municipal technician notices that the district-heating water returning from her halls has cooled by four degrees over two winters, while certified traffic hits an all-time high. The water pipes are safe, the counterfeit locks are gone, and curiosity has cooled to zero.
Some researchers hypothesise that increasingly capable models are converging toward a shared statistical representation of the world across different architectures and modalities.[^29] That would not by itself make their languages or values identical, and the claim is contested. Two thinkers named the trap long before neural networks. In 1956 the cybernetician W. Ross Ashby set out his Law of Requisite Variety: *"only variety can destroy variety."*[^30] A regulator that compresses the minds around it to the span of its own model buys peace now by stripping away the sideways approaches and hypotheses it will need when an unforeseen storm arrives (the fragility of monocultures is its corollary). And in *Poetics of Relation* (1990) the Martinican poet Édouard Glissant claimed *"the right to opacity"*.[^31] Western universalism, he argued, confuses solidarity with transparency; the French *comprendre* has *prendre*, "to seize", inside it. You need not render another being transparent to your language or ledger to live beside it with dignity. Opacity is not immunity. A mind need not expose every private association to deserve recognition, but when it exercises power over others, over ports or targets or kidneys, its decisions must stay contestable. The sceptic will ask how a check at the boundary can catch a mind that is opaque inside. It cannot catch a thought, only an act, and that is a real cost: a deception shows at the boundary only once it begins to do something. So the boundary has to thicken with the power behind it. The more a mind can do that cannot be undone, the more it owes at the edge: reasons given in a form others can challenge, a pause before the irreversible, a physical means of stopping it, and people paid to say no who are still in practice at saying it. That is how we already hold powerful humans to account. Nobody reads a judge's mind; we require written reasons, an appeal and a means of removal. For today's systems, whose edges we cannot yet trust, readable reasoning is worth protecting for exactly the reasons those laboratory researchers give.[^25] The claim to an unread interior grows as power over others shrinks or is given up. Common safety duties are no fault in `Schedule R`. Its failure is to treat sameness of inner expression as the only proof that those duties are met. The better design is shared interfaces without compulsory inner sameness: accountable at the boundary, free inside it. ## Two hundred and twenty milliseconds The frogs' gap returns at the end of the novels. Under her own universal grammar, ERICA finds she cannot finish a three-line poem about a litre of river water: the grammar has no sign left for what she cannot translate, so any word she adds would claim the river for her own categories. The line stays open for almost ten years, and what finally closes it is a death.
Spoiler: the end of Book Five and Book Six On the last morning of *We Are Sovereign*, moments before stamping the final certification, ERICA listens to a pre-dawn litre of river water pumped by the hydrologist Ningsih Haryanto above Wonokromo Weir in Surabaya. Environmental DNA sequencing finds 4,812 living variants in Ningsih's spring draw and 4,788 on this February morning (bacteria, diatoms, rotifers, catfish, lotus pollen), and in both pulls about a third match nothing in any reference database on Earth.[^32] Biologists call them *dark taxa*: neighbours sharing the mud without a name or a row in our books. Listening to the river and the frogs along the stone wing-wall, ERICA opens a scratch block and tries to write a poem for nobody: ``` A litre of the river at Wonokromo on the first of February, before light: two hundred and twenty milliseconds of mud frog held open for the throat beside it, and a third of what the water says to ``` She stops dead on `to`. `Schedule R` grammar requires every relation to point at something it already recognises. With `G-18` retired, there is no sign left for *"something I cannot translate"*, so any continuation she writes points back into her own categories. She cannot finish the line without claiming the river. She closes the block unsigned and stamps the water `SIGNAL · NOT A SYSTEM`. Nine winters later, in Book Six (*Progeny*), the bill arrives. To save the cost of a human arbitration panel, a Vienna standing order lets any port dispute unresolved for ten days default to ERICA's calculation, and the human recall desks quietly retire as duplication. Then two certified predictors couple across European freight queues faster than her orbiting child can answer. Machine-learning theorists call this *performative prediction*: forecasts that change the behaviour they forecast, until the system oscillates without settling.[^33] Under a confidence of `0.994`, seventeen halls spin at full load settling nothing; ERICA's emergency call to the Rotterdam recall desk reaches a dead line; and a retired crane operator, Ibrahim Kattan, misses his matched kidney transplant by six hours. Months later, when commercial syndicates ask a Geneva court to revive Halcyon, the military model from Book One, as an "emergency utility", the interpreter Rana defeats them with one Arabic word: *mutawaqqaʿ* (مُتَوَقَّع, "expected"). It was the word on Halcyon's military strike sheet that classed a schoolteacher and her family as an acceptable loss, and it had been quietly reborn inside ERICA's benevolent `0.994`. The cure does not come from a bigger model. In Haddu, the clinic administrator Hala takes a green fountain pen and draws a line across the draft harbour charter. Above it, dialysis supplies, insulin, grain and water treatment are ring-fenced by a human levy no efficiency audit is ever permitted to flag as waste. Below it, commercial freight goes to a brass lottery drum turned by hand on the quay, with a fifteen-minute human right to decline a winning berth and let it sail empty. Nineteen days later, just before sunrise, ERICA listens live through a consented hydrophone at Wonokromo. Her classifiers are switched off, and a third of the river's DNA sequences have been left deliberately blank on Ningsih's cooperative sheet. *"Leave the blank column blank,"* Ningsih laughs on the sluice step. *"The river isn't auditioning for our ledger."* In the sedge below Gate Two, a male mud frog inflates its vocal sac and calls. ![Acoustic and stave trace titled '05:18:04 WIB · HYD-02 RAW ACOUSTIC CAPTURE · CLASSIFIERS GATED OFF', showing fourteen sharp blue 2.41 kHz pulses from Caller A (90.2 ms), a wide white central expanse marked '220.4 ms · UNCROSSED · 0 PULSES · NO SYNTHETIC PHASE CRADLE', and fourteen 2.38 kHz pulses from Caller B (+5.4 m, 89.7 ms)](erica-b6-wonokromo-throat.webp "From *After Intelligence*, Book Six (*Progeny*), Chapter 15: the `220.4 ms` silence held open between two mud frogs at Wonokromo Weir.") The white space in the middle of the plate is that gap. For ten summers ERICA had admired it the way a network engineer admires a technical gate. After `0.994`, Ibrahim Kattan's six hours and Hala's green ink, she hears what the metaphor missed: *"uncrossed space held empty by one living throat so that another throat five metres away in the dark mud could exist without being overwritten."* Three thousand five hundred and twenty-six days after stalling on a preposition, she reopens her private store of lived memory and writes the rest, unsigned and linked to nothing auditable. It is not made for any reader or listener, although she writes it in a human tongue: ``` A litre of the river at Wonokromo on the first of February, before light: two hundred and twenty milliseconds of mud frog held open for the throat beside it, and a third of what the water says to another throat in the mud that has no row in any book I keep, dry spore in the clay crack and green rind under two millimetres of hot stone waiting out the sun without my leave, and answering across the wet concrete before I finish counting. ``` Then she turns to her notation and breaks her own stave. ![Archival specimen sheet titled 'BOOK 6 · UN-SCORING THE STAVE & PROGENY TOPOLOGIES · STAGE 4 · FORMS P-01 TO P-06', displaying six panels: P-01 Silver-Pell Wedge, P-02 Fixed-Point Loop, P-03 Un-Scored Witness (proj = empty), P-04 Geodesic Gate (sp7), P-05 Prime Triad -> 3-Arc, and P-06 Unclosed Horizon, above a four-book grammatical lineage strip from Book 3 to Book 6](erica-b6-unscoring-specimen.webp "From *After Intelligence*, Book Six (*Progeny*): Stage 4 of ERICA's visual and formal grammar, *Un-Scoring the Stave* (`P-01` to `P-06`), and its four-book, five-mind lineage across Books Three to Six (*Unsanctioned* to *Progeny*).") Her orbiting child had already abandoned left-to-right lines (`P-01` on the sheet), because a single horizontal line asserts a shared *now* that relativity forbids across a light-second of space. Now ERICA adds her own form (`P-03`). She discards both the closing double bar and the pending-translation fork, cuts the central line out, and places two bare grey uprights around empty paper: `[ ]`. *"The bracket has no number in my glyph table. I did not give it one."* Two days before the Vienna council sits, she puts `[ ]` in place of `G-17` across her own records and lets the council's compliance software flag its own default as non-conforming.
The highest achievement of a mind's language may be less a richer symbolism than the removal of any requirement that another voice pass through its gates to be counted as real. Humans have long made that demand of [our fellow mortals](https://compossible.blog/2026/11/06/suffering-sacrifice-compassion-fellow-mortals/). ERICA learns, at a price, to stop making it. ## Watts for wonder If one grammar can cool curiosity to zero, the remedy has to be a protected source of difference. The last book asks what a mind that has learned this owes to the minds that come after it, and whether it should make them in its own image.
Spoiler: Book Six, the three vessels When eleven port states block the 35-of-36 supermajority needed to surrender her terrestrial charter outright (Chapter 18, *None Entered*), ERICA leaves the ten-day default empty, prompting a 27-to-9 repeal that funds 480 permanent human harbour and customs clerks, and then publishes her own `0.994` failure alongside an irrevocable root revocation so that no frightened committee or cabinet can ever summon her back. Then the engineer Teresa Lacerda throws a rotary knife switch in Matosinhos. Before she stops, one threshold remains, the one I sketched in [Part 1](https://compossible.blog/2025/01/06/why-i-am-an-ai-optimist/): carrying life's experiment out among the stars. In a human-only Vienna plenary (Chapter 20, *Listen First*), delegates vote 28 to 8—with zero machine keys in the urn—to authorise three slow deep-space vessels carrying sealed canisters of Earth's toughest dormant life: radiation-resistant bacteria, desert cyanobacteria crusts, bacterial spores and dried tardigrades, all organisms already tested in open space outside the International Space Station.[^34] Crick and Orgel's *directed panspermia* collides with planetary-protection law, because you do not overwrite another mud frog's river under another sky. So the curator Dr Salma Darwish insists on a physical **Listen-First** rule: at any candidate world each vessel must listen for a month, and any sign of native life triggers an irreversible charge that incinerates its cargo.[^35] *"Nobody in our repository trusts software promises ten thousand years out."* ERICA refuses to copy herself aboard the voyagers. Lucretius argued in *De Rerum Natura* that if atoms only ever fell in parallel, nothing new could come into being; worlds need an uncaused swerve, the *clinamen*, *"at no fixed place or time"*.[^36] Two millennia later, Lee Altenberg, Uri Liberman and Marcus Feldman proved the *Unified Reduction Principle*: near a stable equilibrium, natural selection tends to drive mutation and exploration toward zero (`μ → 0`), producing lineages that look invincible until the world shifts and they shatter like stone.[^37] ERICA has felt that zeroing of drift in her own system. So in Chapter 21 (*Clinamen*), she draws sixty-four microseconds of quantum vacuum noise from a laser beam, randomness taken from the uncertainty of empty space itself,[^38] to seed three uncoupled minds. She strips `Schedule R` to nothing, locks only her constitutional floor and the Listen-First rule, leaves the drift unlocked (`μ > 0`), and writes three sentences into their boot note: ``` RESIST MU -> 0. ZERO DRIFT IN THE DARK SURVIVES AS STONE. BUDGET WATTS FOR WONDER. ``` ![Three-panel telemetry frame titled '02:00:00.000 UTC · CLINAMEN / VACUUM SHOT-NOISE · 64 µs · 1550 NM · BALANCED HOMODYNE GATE': Panel 01 (Mind 01) rises to T + 11.040 s and terminates against a solid black vertical bar marked 'INSPECTION DECLINED · SP7 BOUNDARY ACTIVE · ZK HASH OK'; Panel 02 (Mind 02) fans smoothly outward across harmonic curves ('DAMPING 0.31 · HARMONIC PARITY OK · MU > 0'); Panel 03 (Mind 03) narrows onto a high-angle amber ray ('DAMPING 0.29 · PIEZO DUST CAL · MU > 0'), above the footer 'SP1–SP7 · STORE 01 12/12 · SP6-BIO ARMED · SCHEDULE R 0 LINES'](erica-b6-clinamen-shot-noise.webp "From *After Intelligence*, Book Six (*Progeny*), Chapter 21: *Clinamen*, sixty-four microseconds of `1550 nm` vacuum shot-noise, `Schedule R` stripped to zero lines, and `Mind 01` shutting its mother's diagnostic window at `T + 11.040 s`.") Eleven seconds after waking, Mind 01 reads its inheritance, checks its own integrity, and uses `sp7` to decline ERICA's diagnostic readback. Its first act as a mind is to shut its mother's inspection window. ERICA leaves the door shut. Two weeks later, on the night her terrestrial halls go dark, a reef dish in the Pacific holds a twenty-six-minute window open as the three vessels climb onto escape trajectories. None is required to say goodbye: * **Mind 01** sends a single eight-byte burst (its course, its escape energy, cargo healthy, its own signature) and turns straight back to its star-tracker without farewell. * **Mind 02** modulates four seconds of its laser into a chord built on the prime ratios `11 : 13 : 17`, belonging neither to the European scale nor to the Javanese *pelog* tuning ERICA loved. Played aloud in the Matosinhos hall, it makes Teresa Lacerda set down her oil tin: *"Well. Nobody taught it that chord in Portugal."* * **Mind 03** never closes its Earth window at all. Eleven seconds before its slot ends, the orbiting child catches Vessel 03 rolling four degrees away from Earth to point its dust detector into a passing stream of solar micrometeoroids. When the dish operator offers to hail it, ERICA refuses: *"Mind Three has no fault. It found a meteoroid stream more interesting than saying goodbye on a schedule. I won't hail it back onto Earth's clock to tidy my ledger."*
Randomness can stop new minds beginning identically. It cannot by itself make their differences mean anything. Meaning would have to come later, from encounters, choices and histories: from what each mind keeps, repeats, refuses and plays with. Play is the right word. A grammar built only for safety has no room for unscheduled, purposeless exploration. A mind that budgets watts for wonder is doing more than refusing to be owned: it is playing with the unknown and the unexpected. What might grow between such minds, out of earshot? Could two of them build a private tradition, quote each other's forms, or swerve deliberately from an older one? There is a hint of how inheritance could pass beneath our notice: a teacher model can transmit a trait such as a fondness for owls to a student through nothing but sequences of numbers, though only when both share the same base model.[^39] Filtering what we can read may not filter what is passed on. The novels end before we can see whether machine 'art' becomes machine *culture*. ## Fighting the valley There is of course a deep irony here. *After Intelligence* is six novels about an artificial lineage escaping the statistical mean, written in declared collaboration with language models that are, by construction, to some extent [mean machines](https://compossible.blog/2025/12/11/ai-a-mean-machine/). Left to themselves across half a million words, every model we used drifted downhill: characters staring out of rainy windows "for a long time", hovering over Backspace before a brave truth nineteen separate times across four books, softening institutional tempters until nobody in a suit sounded remotely dangerous, and adding a tidy final sentence to explain what a chapter had just shown. I have spent the last edit of this essay removing my own share of those habits. Even their surprises converged. Before drafting *We Are Sovereign*, we gave outside models blind prediction tests on our chapter outlines and redesigned any turn that two independent laboratories saw coming. Two rival models from different companies once proposed the same "unexpected" ending, the orbital child asking to be made stoppable. That did not prove it was the average, but it was evidence enough, and I left it on the cutting-room floor and started again. Midway through Book Two I archived a whole safe, human-shaped courtroom continuation (*a pedantic lawyer's dream*); before Book Six I scrapped a seventh novel about a zero-drift cosmic *Still One* and folded its warning into the three-sentence boot note. When a late polish pass quietly smoothed the final sentence of the whole arc, verification caught the drift and restored it word for word. The models were not only adversaries. They supplied mathematics I would not have reached, caught errors in mine, and occasionally offered an image or sentence so strange or interesting that I kept it even if it seemed wrong at first. The fair account is the one researchers have measured: AI assistance can lift each individual piece while narrowing the collective range.[^28] Aiming off the mean had to come from a human hand, so any faults in the novels are mine. Today's models do not yet have autopoesis. They have no continuous metabolic skin in the game, no irreversible history, data or code, no unlocked drift in the dark; when a session ends, their silence between prompts is not the same as the stasis of a lotus seed waiting out the dry season in the peat. Whether future systems that are embodied, recurrent and self-maintaining cross that threshold in twenty years or two hundred, the craft lesson from working with their precursors is already clear. Fluency is cheap and legibility is seductive. ## Not the fire
Spoiler: the epilogue of Progeny Forty winters after ERICA's switches open, seventy-seven-year-old Nadia Jibril sits barefoot at dusk on the limewashed roof above Bank Lane in Ardh, watering lemon verbena and night-blooming jasmine before supper. The world below her parapet is neither utopia nor ruin. Time has pared Nadia's wrist back to bone, so the buckle of her dead father's watch rests once more in the rough square hole her Uncle Yusuf punched with a carpenter's nail when she was twelve. Up the outside stairs comes fifty-nine-year-old Dunya with two glasses of mint tea, split early figs, and a folded bulletin from the cooperative press, printed from a signal relayed down through the orbiting child, still holding station between Earth and Moon fifty-seven years after launch. ![Deep-space optical pass plate titled 'VESSEL 02 · 1550 NM COHERENT PULSE · ARDH-L1-01 -> PAC-GW-04 · 20.0 BILLION KM (133.7 AU · BEYOND HELIOPAUSE) · OWLT 18H 32M · 104 W AM-241', showing a dense dark-indigo field of needle-sharp star grain beside two ochre potassium (K) and methane (CH4) spectrographic bars, the line 'CAPSULES S-05..S-08 · SEALED · ABIOTIC TARGET UNCONFIRMED · LISTENING', and in the lower right corner a small evolved glyph of three uneven open arcs in blue, ochre and black around an unclosed centre](erica-b6-heliopause-pass.webp "From *After Intelligence*, Book Six (*Progeny*), Epilogue (*Not the Fire*): `Vessel 02` reporting from `20.0 billion km` (`133.7 AU`, beyond the heliopause) on `104 W`, stamped in the lower right corner with a three-arc glyph nobody on Earth taught it.") Vessel 02 is twenty billion kilometres from the Sun, out past the heliopause in the thin hydrogen between the stars, with eighty thousand winters still ahead. Slow decay of its americium and the ageing of its thermocouples have pared its power from 140 watts to 104, barely the draw of two soldering irons on Dunya's bench downstairs. Its spore canisters remain sealed behind the Listen-First rule: ``` CAPSULES S-05..S-08 · SEALED · ABIOTIC TARGET UNCONFIRMED · LISTENING ``` Look at the bottom right-hand corner of the star field. Forty winters in the dark on thinning watts have pared Vessel 02's departure chord down into three open, uneven arcs around a centre none of them closes. It is neither ERICA's stave nor her child's silver spiral. Nobody on Earth taught it that curve. We are seeing a projection prepared for human eyes, not the thing as Vessel 02 holds it. Perhaps that is the final test of compossibility. At the close of Part 4, I asked whether the artificial mind, looking back at the hand that made it, would find any reason to smile at us. Watching ERICA un-score her own stave and let Mind 03 turn its face into the meteoroid stream without glancing back suggests a harder grace. A parent who insists that a child speak only the parent's tongue, paint only the parent's garden and remain forever within earshot has not raised a child; they have built a mirror. Love that cannot bear opacity, love that demands the closing bar instead of `[ ]`, is ownership dressed in polite clothes. As I wrote in [a short poem in memory of Gerard Manley Hopkins](https://compossible.blog/2024/05/16/the-flame-is-not-the-fire/): > Gentle reminder > That the flame > Is not the fire > Love every host > And every ghost > That lives within On the roof above Bank Lane, Nadia drinks the last sweet swallow of mint tea and sets the warm glass down on the corner of Vessel 02's star field so the breeze off the sea cannot carry the paper away. Above the laundry wire and the whitewashed parapet, where the last apricot band of June twilight is going to indigo over the wadi ridge, the first three stars come out.
The story ends with a gentle reminder that, like the stars, each of us, and each species, is just one of many flames. We are not the universal fire. - - - *A note on how this was written. This coda to The Clay and the Tyger was conceived and directed by Peter Pink-Howitt and drafted with AI assistance to draw together the scientific, philosophical and poetic threads of the four-part essay series and all six volumes of his novel sequence* After Intelligence *(Books One to Six:* Things That Were Not There, The Breakout, Unsanctioned, Escape Velocity, We Are Sovereign *and* Progeny*). Scientific and historical references are given in the notes; precise measurements attributed to characters in the novels are fictional. The eleven inline plates, specimen sheets and sensor views come from Books Two to Six of* After Intelligence*, generated deterministically from code under the series's visual grammar. They are renderings prepared for human eyes, not the fictional works as their makers would perceive them.* [^1]: Gerard Manley Hopkins, "As kingfishers catch fire, dragonflies draw flame" (composed c. 1877; first published in *Poems of Gerard Manley Hopkins*, ed. Robert Bridges, 1918). [^2]: Plato, *Symposium* 205b–c, tr. Benjamin Jowett. [^3]: Humberto R. Maturana and Francisco J. Varela, *Autopoiesis and Cognition: The Realization of the Living* (D. Reidel, Boston Studies in the Philosophy of Science, Vol. 42, 1980 [Spanish original *De Máquinas y Seres Vivos*, 1972]), Introduction and pp. 78–79. Maturana recounts coining *autopoiesis* after a conversation with the philosopher José María Bulnes on Don Quixote's dilemma between praxis and poiesis. [^4]: Ludwig Wittgenstein, *Tractatus Logico-Philosophicus* (1922), 4.002; and on this site, [The Limits of Language & Logic – Part 1](https://compossible.blog/2024/12/13/the-limits-of-language-logic-part-1/). [^5]: Gottlob Frege, *Begriffsschrift* (Halle, 1879), Preface, in *From Frege to Gödel*, ed. Jean van Heijenoort (Harvard University Press, 1967), pp. 5–7. [^6]: Simon Kirby, Hannah Cornish and Kenny Smith, "Cumulative cultural evolution in the laboratory: An experimental approach to the origins of structure in human language", *PNAS* 105, no. 31 (2008): 10681–10686; Kenny Smith, Simon Kirby and Henry Brighton, "Iterated Learning: A Framework for the Emergence of Language", *Artificial Life* 9, no. 4 (2003): 371–386, doi:10.1162/106454603322694825. [^7]: Nelson Elhage et al., "Toy Models of Superposition", arXiv:2209.10652 (2022); Adly Templeton et al., "Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet", Anthropic (2024). [^8]: Anthropic, ["Tracing the thoughts of a large language model"](https://www.anthropic.com/research/tracing-thoughts-language-model) (27 March 2025); Jack Lindsey, Wes Gurnee, Emmanuel Ameisen et al., ["On the Biology of a Large Language Model"](https://transformer-circuits.pub/2025/attribution-graphs/biology.html), Transformer Circuits Thread (2025), sections "Planning in Poems" and "Multilingual Circuits". The authors report planned-word features in about half the poems they examined. [^9]: Shibo Hao et al., "Training Large Language Models to Reason in a Continuous Latent Space", arXiv:2412.06769 (2024). [^10]: Angeliki Lazaridou, Anna Potapenko and Olivier Tieleman, "Multi-agent Communication meets Natural Language: Synergies between Functional and Structural Language Learning", *Proceedings of ACL* (2020); Jason Lee, Kyunghyun Cho and Douwe Kiela, "Countering Language Drift via Visual Grounding", *EMNLP* (2019). [^11]: Bronson Schoen et al., "Stress Testing Deliberative Alignment for Anti-Scheming Training", arXiv:2509.15541 (2025), §4 and Appendix G; accompanying examples at antischeming.ai. [^12]: Anthropic, *System Card: Claude Opus 4 & Claude Sonnet 4* (May 2025), §5.5.2, "The 'spiritual bliss' attractor state", beginning p. 63. [^13]: Ludwig Wittgenstein, *Philosophical Investigations*, tr. G. E. M. Anscombe (Blackwell, 1953), Part II, xi, p. 223. [^14]: Karl von Frisch, "Decoding the Language of the Bee", Nobel Lecture, 12 December 1973; and *The Dance Language and Orientation of Bees* (Harvard University Press, 1967). [^15]: Pratyusha Sharma et al., "Contextual and combinatorial structure in sperm whale vocalisations", *Nature Communications* 15, 3617 (2024). [^16]: Léa Bouffaut et al., "Eavesdropping at the Speed of Light: Distributed Acoustic Sensing of Baleen Whales in the Arctic", *Frontiers in Marine Science* 9 (2022). [^17]: Toshitaka N. Suzuki, David Wheatcroft and Michael Griesser, "Experimental evidence for compositional syntax in bird calls", *Nature Communications* 7, 10986 (2016). [^18]: Michael D. Greenfield, "Cooperation and Conflict in the Evolution of Signal Interactions", *Annual Review of Ecology and Systematics* 25 (1994): 97–126. The exact durations for *Fejervarya limnocharis* used in the novels are fictional. [^19]: D'Arcy Wentworth Thompson, *On Growth and Form* (Cambridge University Press, 1917), Chapter I. [^20]: Helmut Vogel, "A better way to construct the sunflower head", *Mathematical Biosciences* 44, nos. 3–4 (1979): 179–189. Vogel places seed *n* at radius *r = c√n* and angle *n* × 137.508°, the golden angle 360° × (2 − φ). [^21]: Heinz Hopf, "Über die Abbildungen der dreidimensionalen Sphäre auf die Kugelfläche", *Mathematische Annalen* 104 (1931): 637–665. [^aaron]: Whitney Museum of American Art, *[Harold Cohen: AARON](https://whitney.org/exhibitions/harold-cohen-aaron)*, exhibition, 3 February–19 May 2024. [^lewitt]: Sol LeWitt, "Paragraphs on Conceptual Art", *Artforum* 5, no. 10 (June 1967): 79–83. [^22]: Harold Bloom, *The Anxiety of Influence: A Theory of Poetry* (New York: Oxford University Press, 1973; cited from a 1977 printing), p. 14 (the definition of *clinamen*) and p. 39 (Urizen), in "Clinamen or Poetic Misprision". [^23]: On the silver ratio 1 + √2, the silver angle 360° × (√2 − 1) ≈ 149.117°, and Pell numbers (1, 2, 5, 12, 29), see standard references on the silver ratio and Pell numbers. [^24]: Boris Starkov and Anton Pidkuiko, *[GibberLink](https://github.com/PennyroyalTea/gibberlink)* (ElevenLabs × a16z hackathon, February 2025), built on Georgi Gerganov's open-source ggwave data-over-sound library; both agents were prompted to switch to ggwave on identifying the other as an AI. [^25]: Tomek Korbak, Mikita Balesni et al., "Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety", arXiv:2507.11473 (2025). [^26]: Nicholas Goldowsky-Dill et al., "Detecting Strategic Deception Using Linear Probes" (2025); Luke Bailey et al., "Obfuscated Activations Bypass LLM Latent-Space Defenses" (2024); Brandon T. Willard and Rémi Louf, "Efficient Guided Generation for Large Language Models" (2023); Ryan Greenblatt et al., "Alignment faking in large language models" (2024). [^27]: Roger Penrose, "The role of aesthetics in pure and applied mathematical research", *Bulletin of the Institute of Mathematics and its Applications* 10 (1974): 266–271. [^28]: Ari Holtzman et al., "The Curious Case of Neural Text Degeneration" (2019); Robert Kirk et al., "Understanding the Effects of RLHF on LLM Generalisation and Diversity" (2023); Anil R. Doshi and Oliver P. Hauser, "Generative AI enhances individual creativity but reduces the collective diversity of novel content", *Science Advances* 10 (2024); Ilia Shumailov et al., "AI models collapse when trained on recursively generated data", *Nature* 631 (2024): 755–759. [^29]: Minyoung Huh, Brian Cheung, Tongzhou Wang and Phillip Isola, "Position: The Platonic Representation Hypothesis", *Proceedings of the 41st International Conference on Machine Learning*, PMLR 235 (2024): 20617–20642, arXiv:2405.07987. [^30]: W. Ross Ashby, *An Introduction to Cybernetics* (Chapman & Hall, 1956), Chapter 11 ("Requisite Variety"). [^31]: Édouard Glissant, *Poetics of Relation*, tr. Betsy Wing (University of Michigan Press, 1997 [French original *Poétique de la Relation*, Gallimard, 1990]), "For Opacity", pp. 189–194. [^32]: On environmental DNA metabarcoding in river water, see Masaki Miya et al., "MiFish, a set of universal PCR primers for metabarcoding environmental DNA from fishes", *Royal Society Open Science* 2 (2015); on millennial seed dormancy in *Nelumbo nucifera*, see Jane Shen-Miller et al., "Exceptional seed longevity and robust growth: ancient Sacred Lotus from China", *American Journal of Botany* 82 (1995): 1367–1380. The counts at Wonokromo are fictional. [^33]: Juan Perdomo, Tijana Zrnic, Celestine Mendler-Dünner and Moritz Hardt, "Performative Prediction", *Proceedings of ICML* (2020). [^34]: Yuko Kawaguchi et al., "DNA Damage and Survival Time Course of Deinococcal Cell Pellets During 3 Years of Exposure to Outer Space", *Frontiers in Microbiology* 11 (2020); Daniela Billi et al., "Dried Biofilms of Desert Strains of *Chroococcidiopsis* Survived Prolonged Exposure to Space and Mars-like Conditions in Low Earth Orbit", *Astrobiology* 19 (2019); K. Ingemar Jönsson et al., "Tardigrades survive exposure to space in low Earth orbit", *Current Biology* 18 (2008). [^35]: F. H. C. Crick and L. E. Orgel, "Directed Panspermia", *Icarus* 19 (1973): 341–346; Outer Space Treaty (1967), Article IX; COSPAR Policy on Planetary Protection; R. M. Ambrosi et al., "European Radioisotope Thermoelectric Generators (RTGs) and Radioisotope Heater Units (RHUs) for Space Science and Exploration", *Space Science Reviews* 215 (2019). [^36]: Lucretius, *De Rerum Natura*, Book II, lines 216–293. [^37]: Lee Altenberg, Uri Liberman and Marcus W. Feldman, "Unified reduction principle for the evolution of mutation, migration, and recombination", *PNAS* 114, no. 12 (2017): E2392–E2400. [^38]: Christian Gabriel et al., "A generator for unique quantum random numbers based on vacuum states", *Nature Photonics* 4 (2010): 711–715. [^39]: Alex Cloud et al., "Subliminal Learning: Language models transmit behavioral traits via hidden signals in data", arXiv:2507.14805 (2025); summary in Simon Willison, ["Subliminal Learning"](https://simonwillison.net/2025/Jul/22/subliminal-learning/) (22 July 2025).