We built a tiny evolutionary art engine, gave it three different “vocabularies” of shapes, andasked a question no one seems to have asked generative art before: not “how fast” or “howpretty,” but a question a linguist asks a language. Here is exactly what we did, what wefound, and — just as importantly — what we did NOT find.
The setup, in plain words
Imagine a machine that paints pictures — not with a brush, but from parts, like buildingblocks. Each “genome” is three small mathematical expression trees, one for each colorchannel (red, green, blue). Each tree maps every pixel’s position to a value. The treesmutate and cross over across generations, and a fitness function — a kind of “attention” —decides which survive.
The one thing we varied was the vocabulary: the set of primitive operations from whichevery shape is built.
Analytic — many small, transparent operations ( add , sub , mul , sin , cos , abs ). Likea language that builds meaning from many separate words placed side by side.
Fusional — a few operations, each one internally fusing several transformations (aripple that folds sine, add and cosine into one). Few units, much fused meaning,opaque from outside.
Agglutinative — one simple binary combiner plus a chain of transparent one-stepmodifiers, one meaning per bead.
These three mirror the three great strategies by which human languages build meaning —analytic (like Chinese), fusional (like Ukrainian), agglutinative (like Turkish). The parallel is aresonance, not an identity — we will be strict about that throughout.
To keep the comparison honest, the fusional operations are built as compositions of theanalytic primitives — so the “content” is roughly held constant and only the “packaging”varies. (One honest exception: the agglutinative set contains a clip operation the analyticset lacks.)
The open data behind the fitness
We did not want the “attention” to chase an arbitrary target, so we pointed it at a real,citable open dataset: the Keeling Curve — atmospheric CO₂ concentration measured atMauna Loa Observatory, published by NOAA and the Scripps Institution of Oceanography.Three anchor points are directly verifiable: about 316 ppm in 1959, the first daily crossing of400 ppm in 2013, and an annual mean of 424.6 ppm in 2024. (The intermediate values weused are a smooth illustrative interpolation, not the official annual table — for a realinstallation one would download the exact series from NOAA’s public archive.)
A crucial framing, which is the whole point of doing this responsibly: the resulting image is“a form generated from this dataset through this algorithm” — not “the sound” or “theshape” of the atmosphere. We are translating a rhythm into a form, not reading a hiddentruth out of nature.
Experiment 1 — Do the three vocabularies paint with differentaccents?
Same fitness, same budget (population 14, 8 generations, 56×56 pixels), same random-seed policy — 8 seeds per vocabulary, 24 runs in all. Only the vocabulary differs. Then wemeasured the images along two kinds of dimension: ones the fitness constrained (meanbrightness, variance — a convergence check), and one it left free (local gradient — the“roughness” or texture of the image).
The result was clean. In the free dimension — the one attention was NOT holding — thethree vocabularies separated with no overlap at all:
agglutinative: local gradient 1.12–2.55 (mean 1.55) — smooth, still water
analytic: 4.49–20.55 (mean 8.90) — gently rippled
fusional: 21.13–41.04 (mean 34.90) — turbulent, sea in wind
Meanwhile the constrained dimension behaved exactly as a control should: varianceconverged into one narrow band across all 24 runs (1266–1378). The accent appearedprecisely and only where attention left the shapes free.
Here are the actual evolved forms — three rows, one per vocabulary, eight seeds each. Youcan see the accents with your own eyes: the top rows are fine-grained and busy, the bottomrows smooth and calm.
The honest caveat, stated in the report itself: 8 seeds is a pilot, not a proof. And with thelarger sample a fragility appeared that the first tiny pilot had hidden — the boundarybetween analytic and fusional nearly closed: the highest analytic run (20.55) sits just 0.58below the lowest fusional (21.13). The separation on that border is real but delicate; a largersample could yet make them overlap. The agglutinative vocabulary, by contrast, isseparated widely and stably.
Experiment 2 — Can attention force an accent that isn’t there?
This is where it gets interesting. In the first experiment attention left texture free. What if wemake attention actively hold texture — demand a specific roughness (target 12.0) — andpush each vocabulary toward a state it did not choose on its own? We pre-registered ourpredictions before running, then ran 6 seeds per vocabulary.
The verdict split in two — as an honest verdict should.
Where a vocabulary could follow, it did. The fusional vocabulary, whose turbulence hadlooked like fixed nature, calmed dramatically: its floor dropped from 21.13 down to 9.67. Its“character” turned out not to be destiny, but merely where attention had not been calling it.The analytic vocabulary settled with its mean landing almost exactly on the target (12.35 ≈12.0). The old fragile analytic/fusional border dissolved — they now overlap.
But the agglutinative vocabulary could not scream. Attention pulled it toward roughness12 — and it did not move a hair: still 1.12–2.55, mean barely shifting from 1.55 to 1.57. Itsvocabulary simply has nothing to answer that call with. The quietest language did notbecome turbulent, however directly and insistently attention asked.
This gave us a distinction we could measure: two different kinds of failure. One vocabularyemerged from a blind spot — its distinctions were there, attention had just not fallen onthem. The other hit a ceiling — it truly had no word. The experiment separated “has themeans but doesn’t look” from “has no means at all.” 
Attention can awaken what alreadysleeps in the vocabulary;It cannot conjure what the vocabulary cannot think.
Experiment 3 — What shape is the space of all possible forms?
The first two experiments asked “what parts?” This one asks a deeper question: “whatshape is the space the forms live in?” Take one evolved form and many of its neighbors —images differing by one small genome change. If the neighbors stay close, there’s a smoothvalley around the form. If one change tears the image apart, that’s the edge of a cliff.
We pre-registered two predictions. One held; one did not — and the one that failed taughtus more than the one that held.
Prediction confirmed: the median roughness of the terrain follows the typology —agglutinative smoothest (median neighbor distance 30.3), then analytic (47.2), then fusional(55.6). One mutation step in the fusional vocabulary moves the image nearly twice as far asin the agglutinative, because a fused operation changes more at once. The same vocabularythat paints a wilder accent also lives in a steeper landscape.
Prediction NOT confirmed: we expected the distances to be “heavy-tailed” — mostly near,a few far cliffs. They weren’t. At this mutation strength almost every neighbor was alreadyfar (the “valley” fraction was 0–5%), and the tails were lighter than predicted (top-decile-to-median ratio only 1.46–1.81, where we’d expected above 2). The honest conclusion: atthis step size the neighbors are not “near” — the terrain is fairly uniformly steep rather than“valley-plus-cliff.” The real valley likely lives at smaller mutations — which became the nextexperiment, whose predictions we had not yet written.
What we take from all this
Put together, the three experiments say one thing: the strategy of a vocabulary shapes notonly the look of the forms, but the shape of the space evolution wanders through. A fusionallanguage lives in steeper mountains; an agglutinative one on gentle hills. The materialdetermines the landscape.
And the most useful moment of the whole project was a failure honestly reported — theheavy-tail prediction that didn’t hold. It didn’t embarrass the work; it pointed to where thereal structure was hiding (at smaller mutation steps). A prediction that survives a test and ahypothesis that honestly falls are both worth more than applause.
Reproduce it yourself
Everything here runs from one script on an ordinary computer — no GPU, no cloud. Eachexperiment is a single command with resume-after-interruption support, so the samecommand picks up where it left off:
python genetic_art_synthesis.py typology --seeds 8 # Experiment 1python genetic_art_synthesis.py attnswap --seeds 6 # Experiment 2python genetic_art_synthesis.py topology_land # Experiment 3
Each writes its own statistics file and comparison grid. Every number in this article was readdirectly from those files, not from memory.
The honest boundaries (the part that matters most)
This is a closed generative system of mathematical trees. Nothing here provesanything about human languages, human minds, or human perception. The parallel tolinguistic typology is a resonance of structure, not an identity of mechanism.
All three experiments are pilots — small seed counts, one fitness signal, a reducedbudget. Ranges and overlaps are reported honestly instead of dressed up assignificance tests.
The vocabularies are specific instantiations of each strategy; the results hold for theseparticular operator sets, and generalizing would need several independent sets perstrategy.
The internal units (image distances on a 0–255 scale) are comparable only within thiswork.
But numbers that rest on a real experiment, honestly bounded, are already a foundation —not a fog.
This is the first entry in Anti-curated Experiments — a column for showing the actual runs,including the ones that didn’t go as predicted. If a result here ever looks too clean, ask tosee the file. We kept them all.
Supporting Data — The Keeling Curve (as used in the experiment)
Companion file for “Do Vocabularies Have Accents?” · Anti-curated Experiments
What this is
The fitness function in all three experiments points at atmospheric CO₂
concentration — the Keeling Curve — measured at Mauna Loa Observatory,
Hawaii, and published by NOAA (Global Monitoring Laboratory) and the Scripps
Institution of Oceanography.
The values used (parts per million)
|Year|CO₂ (ppm)|Status |
|----|---------|----------------------------------------|
|1959|316.0 |anchor — directly sourced |
|1970|325.7 |illustrative interpolation |
|1980|338.7 |illustrative interpolation |
|1990|354.4 |illustrative interpolation |
|2000|369.7 |illustrative interpolation |
|2010|389.9 |illustrative interpolation |
|2013|400.0 |anchor — first daily crossing of 400 ppm|
|2018|408.5 |illustrative interpolation |
|2024|424.6 |anchor — annual mean, NOAA 2025 report |
Honesty note
Only the three anchor points (1959 ≈ 316 ppm, 2013 = first 400 ppm crossing,
2024 = 424.6 ppm annual mean) are directly sourced and verifiable. The
intermediate values are a smooth illustrative interpolation for
demonstration — they are NOT the official annual table. For any real
installation or scientific use, download the exact annual series directly from
NOAA’s public archive:
https://gml.noaa.gov/ccgg/trends/data.html
Framing (this is the point)
The image evolved from this series is “a form generated from this dataset
through this algorithm” — not “the sound” or “the shape” of the atmosphere.
The experiment translates a numeric rhythm into a visual form; it does not read
a hidden truth out of nature. The dataset is used as an arbitrary, real,
citable signal — nothing more is claimed.
We are.

typology comparison n8

fig1 typology 

fig2 attention 

fig3 topology

creol grid