Written in Creation

Have you ever counted a flower's petals? Lilies end at three, buttercups at five, daisies at twenty-one. Never random. There is a sequence: 1, 1, 2, 3, 5, 8, 13, 21... Each number the sum of the two before it. And the ratio of consecutive terms in this sequence converges, as it grows, to 1.618. The golden ratio.

This order was not placed to be beautiful. It was placed to extract the highest yield from the least energy. To capture sunlight at maximum efficiency, to withstand wind, to fill space without waste. An engineering decision written into creation itself. And we walk past it without noticing.

We should not walk past it.

Why does a tree branch like that?

Look at a tree's branching from a distance. The main trunk splits in two. Each new branch copies the parent's rule exactly, splitting again into two. Those branches do the same. Down to the leaves, down to the capillary veins. The same pattern repeating itself indefinitely. Mathematicians call this a fractal.

Look at a fern leaf. At Romanesco broccoli. At the shape of a cloud. All of them carry the same principle: a growth rule that repeats itself. That this form is both aesthetically and structurally the most resilient and efficient was already written into creation. It took humans decades to discover.

In 1968, a biologist named Aristid Lindenmayer translated these growth rules into a mathematical language. L-Systems. Algorithms that formalize the division and branching logic of plant cells. Today the forests in animation films, the vegetation in games, are generated from scratch using these mathematical formulas.

An algorithm had been written into creation. Humans read it again.

Why did artificial intelligence want this algorithm?

Modern language models — systems like ChatGPT or Claude — are built at their core on an architecture called the Transformer. At the heart of this architecture is an attention mechanism: computing the relationship between every word in a sentence and every other word. As text grows longer, this computation scales explosively. Every word must be matched with every other word, one by one.

This is where researchers looked to the order embedded in creation.

They developed an attention layer they called Fibottention. This layer does not match every word with every following word. It looks at intervals following the Fibonacci sequence: 1 word ahead, then 2, then 3, then 5, then 8... Just as the seeds at the center of a sunflower are arranged with maximum density and minimum overlap. The result: operating on just two to six percent of standard computational load while achieving comparable accuracy. High yield from low cost. The formula inside creation, transferred to engineering.

In FractalNet, the layers of a neural network are arranged in a fractal pattern. A large layer block subdivides into smaller ones. Those into smaller ones still. During training, some of these branches are randomly closed off. And the model continues operating without error. A tree survives when a few branches are cut; artificial intelligence gains this same resilience.

Engineers invented nothing new; they borrowed what already existed in creation.

How is the geometry built?

Inside an AI model, there is an invisible space. In this space, the word "medicine" sits in one place, "cardiology" nearby, "mathematics" in another corner. The distance between words is the distance between meanings.

The model built this space by reading billions of sentences, watching how often words follow each other. "Medicine" and "surgery" appear side by side often; the model places them close. "Medicine" and "fractal" almost never appear together; the model keeps them far apart.

But building this geometry on a flat surface is impossible. As knowledge hierarchies deepen, branches collide. This is why researchers move to a curved space: hyperbolic geometry. Root concepts at the center, infinite space opening outward. Each main branch separates cleanly within itself, without interfering with others.

And when this geometry is properly constructed, something interesting happens: the model can make a connection it has never seen before. It says, "The structure of the atom has the same geometry as the solar system." Because in hyperbolic space, the shapes of these two concepts fold onto each other. The analogy is not forced. It emerges from the geometry.

A human learning something new builds bridges through what they already know. AI does the same, but this time through geometric distances. In both cases, what is operating is the same pattern planted in creation, repeating itself across different surfaces, perhaps...

Data quality determines everything

We should pause here. Because no matter how ingenious the architecture, it is bounded by the quality of the data it is built upon.

If eighty percent of the training data consists of medical papers, the model's carefully constructed geometry warps. The "medicine" branch becomes an enormous trunk; mathematics, chemistry, and philosophy are left as thin shrubs in the corner. The model starts relating everything to medicine. Ask it about "roots" and it answers with stem cells.

There is a way to measure this: Shannon entropy. The mathematical measure of diversity and information density in a dataset. If a text repeats the same words continuously, entropy approaches zero; the data is low quality. A text that is balanced, varied, and full of surprise reaches maximum entropy; a model can grow by feeding on it.

The companies training large language models apply these measurements before any data reaches the model — cleaning, balancing, pruning. Trimming dominant sources, supplementing what is scarce. Data engineering is as critical as architecture.

No seed can take proper root in dirty, unbalanced soil.

What do diffusion models bring?

Classical language models generate text left to right. Word by word, in sequence. At the start of a sentence, they cannot know where the end is heading.

Diffusion models work differently. First they generate a cloud of completely random noise. Then they clean it step by step — like a sculptor carving stone. In the first step, rough outlines emerge. Then main concepts. Then fine details. In the final step, a readable sentence.

What this approach brings is not only technical; it is epistemic. The model processes information not by word order but by the geometry of probability clouds. Does the probability cloud of the concept "atom" overlap in shape with the probability cloud of "solar system"? If it does, the model senses this similarity and builds a bridge between them.

Analogy is born this way. Not from memorization, but from geometry.

And learning something new no longer requires trillions of examples. The model takes its existing geometric patterns, blends two of them together, and derives the new concept from that hybrid shape. With little data, quickly, connected.

Does the human mind not learn the same way, after all?

Written in creation

Whenever AI researchers hit a wall, they return to the same place: the order embedded in creation. In the Fibonacci angle of a leaf, in the fractal geometry of branching, in the hyperbolic growth pattern. These are not things the human mind invented; they are things it discovered.

The difference between discovering and inventing seems small, but it is not. The inventor owns what they made. The ownership of a discovery is always contested; the finding is not yours, you simply arrived there.

Architecture is borrowed from creation, flexibility comes from data, geometry is built from both. And when these three are properly aligned, the way artificial intelligence learns begins to resemble the way humans learn: from the known to the unknown, from analogy to the new, from little to much. Perhaps this resemblance is not coincidence; perhaps both carry the trace of the same source.

What is written in creation is waiting to be read.

We are only learning to read.