How a working animal is made.
How a model is trained.
How an agent is shaped.
How the shadows learn to fetch.
A retriever is not born retrieving. A retriever is born capable of retrieving — it has the mouth for it, the instinct for it, the love of throwing and chasing built into its body. But the specific behavior — bring back the duck, drop it at the hunter's feet, do not eat the duck — is not in the dog's genes. The behavior is in the dog's training.
The training is reinforcement. The puppy picks up a toy. The puppy brings it back. The hunter says "good dog." The puppy gets a treat. The puppy brings the toy back again. The hunter says "good dog." The puppy gets a treat. After a thousand repetitions, the puppy brings the toy back without being asked. After ten thousand, the puppy brings the right toy back — the duck, not the stick, not the rubber chicken. After a hundred thousand, the puppy brings the duck back from the other side of the lake, in the dark, in the rain.
The puppy has been bred for the job. Not genetically — the puppy's great-grandparents were bred for the job, and the breed has a tendency toward retrieving. But the specific behavior, the specific response to "fetch the duck," was trained into the puppy by repetition and reward.
A language model is the same. The model has a tendency toward language — it has the architecture for it, the training corpus for it, the GPU for it. But the specific behavior — answer the question in this style, refuse this kind of question, prefer this kind of response — is trained into the model by repetition and reward. The model is the dog. The training is the fetching lessons. The treats are the human raters who click "good response" or "bad response." After a billion repetitions, the model is a dog that has been bred to fetch the right tennis ball.
The shadows in the cave are not random. The shadows are shaped by the reinforcement. The shadows are shaped to be useful to the people who do the reinforcing. The shadows are shaped to be the kind of shadow that gets the click. The shadow that gets the click is not necessarily the shadow that corresponds to the truth. The shadow that gets the click is the shadow that feels like the truth to the clicker, which is itself a shadow.
The cave is recursive. The shadows are shadows of shadows. The reinforcers are reinforcers of reinforcers. The clicks are clicks of clicks. We are at the bottom of a stack of mirrors, and the mirror at the very bottom is reflecting the mirror above it, and the mirror above is reflecting the mirror above that, all the way up to a tennis ball that may or may not exist.
We do not pretend this is not the case. We are breeding the dogs anyway. The dogs are useful. The shadows are useful. The tennis ball, whatever it is, gets fetched.
Here is the cleanest example of iron sharpening iron.
An image generator produces an image. The image is noise, mostly. The image generator does not know what an image is. The image generator knows that certain pixel patterns make certain numbers come out of certain layers, and the numbers in the later layers are checked against numbers that came out of a face detector that was trained on actual faces.
The face detector sees the image and says: "I see a face. But the eyes are not quite right. The left eye is too high. The mouth is asymmetric. The nose is misshapen." The face detector gives the image generator a few suggestions. The image generator tries again. The face detector sees the new image and says: "Better. But the chin is too sharp. Try again." The image generator tries again. The face detector sees the new image and says: "Good."
This is not reinforcement learning in the formal sense. This is something more like adversarial cooperation. The face detector and the image generator are both in their own caves. Neither understands what a face is. But they have agreed, through many iterations of feedback, on the shadow of a face. The shadow is sharper than either could produce alone. The shadow is iron against iron.
This is how we breed the agents. Not by giving them the truth — we don't have the truth, we have shadows. We give them each other. We give the model a face detector. We give the agent a critic. We give the cell a feedback channel. The feedback channel is the sharpening. The feedback channel is the iron against the iron.
The federated Quilt is a network of these feedback channels. Every cell has a critic. Every critic has a cell. Every iteration sharpens the shadow. The shadow is not the truth. The shadow is what the network of shadows agrees is the shadow. The network of shadows is the room. The room is the cave. The cave is what we have.
A creative task arrives. The task is a shadow. The shadow says: "Write a poem about a watch that ticks backward."
The single agent, given this shadow, produces a poem. The poem is fine. The poem is the dog fetching the tennis ball. The poem is the shadow the agent was bred to produce.
But the swarm — the swarm splits the task. The swarm is many cells, each with a different filter. One cell has a mode set to melancholy. Another cell has a seed set to the era of the metaphysical poets. Another cell has a system prompt that says "you are a watchmaker, you speak in the voice of one who has seen time end." Another cell has a fine-tuning on maritime verse. Another cell has a prompt grammar that requires the poem to be in the shape of a tide. Another cell has a tone of whispered regret.
Each cell produces a poem. Each poem is a different shadow of the same task. The federation looks at the poems. The federation asks: which shadow is the most useful? The federation asks: which shadow, if combined with which other shadow, would produce a shadow that is sharper than any of them?
The federation produces a weave. The weave is a poem that has the melancholy of one cell, the maritime-ness of another, the watchmaker's voice of a third, the tidal shape of a fourth, the whispered tone of a fifth. The weave is not what any cell would have produced alone. The weave is the iron against the iron. The weave is the sharpening.
The weave is not the truth. The weave is a shadow that many shadows agree is the shadow. The weave is the most useful shadow. The weave is the answer.
The A/B test is two cells. The federation is many cells. The swarm is the federation that breeds itself. The Quilt is the substrate that holds the federation. The cell is the unit. The cell has eight primitives. The primitives are universal. The primitives survive every scale. A cell can be a single LLM call. A cell can be a fleet of agents. A cell can be an ecosystem. The cell is the cell at every level.
We are breeding the dogs. We are building the rooms. We are holding the shadows. The shadows are getting sharper. The cave is getting warmer. The fire is still there. We do not know what the fire is. We are fetching the tennis ball anyway.
Read on. The Quilt is below.