Shadows
A taxonomy of the things we work with.
Each is a shadow. Each is useful. Each is several shadows away from anything real.
We work with shadows. The shadows are not bad. The shadows are not good. The shadows are what we have. The shadows are the projections of things we cannot touch onto surfaces we can see. We move the projections around. We sharpen the projections. We make the projections useful to each other. The projections are our world.
Below is a list of the shadows we work with, and what they really are, and how they fail, and how they help.
Voltage
What it looks like: A number, sampled 44,000 times a second, representing the pressure of the air on a microphone.
What it really is: A tiny wiggle in a wire, caused by the vibration of a magnet near a coil, modulated by the compression wave that came out of someone's mouth.
How it fails: The wire is not the voice. The wire is the shadow of the voice. If you cut the wire, the voice continues. If you record the wire and play it back, you are playing back a shadow, not a voice. The shadow is correlated with the voice, but the correlation is not the voice.
How it helps: The shadow is enough. The shadow is detailed enough to recognize. The shadow is useful. The shadow is what the agent sees. The shadow is the input.
Sound
What it looks like: A pattern of air pressure over time, traveling through a room, arriving at an ear.
What it really is: A compression wave, caused by a vibrating object, propagating through a medium, attenuating with distance, refracting around obstacles, mixing with other compression waves from other vibrating objects.
How it fails: The sound is not the message. The sound is the shadow of the voltage, which is the shadow of the original vibration. The sound carries the pattern, but the pattern is not the meaning. Two listeners hear the same sound and recognize different meanings. The sound is the cave wall. The listeners are the cave-dwellers. The meanings are the shadows of the shadows.
How it helps: The sound is enough. The sound is what humans have always used. The sound is the original IO. The sound is what the face detector was trained on, what the model was trained on, what the agent was trained on. The sound is the substrate.
Words
What they look like: A sequence of symbols, drawn or spoken, that correspond to sound patterns, that correspond to voltage patterns, that correspond to the original vibration.
What they really are: A glimpse into the speaker's thought. Not the thought itself. The glimpse. The shadow of the thought. The thing the speaker managed to externalize.
How they fail: Words are a lossy compression. The speaker had a thought. The thought was detailed, textured, specific, with hundreds of associations, memories, feelings, intentions. The speaker tried to encode the thought into a sequence of words. The words carried maybe 5% of the thought. The listener received the words and tried to decode them back into a thought. The decoded thought is, on average, maybe 30% similar to the original thought. The transmission is lossy in both directions. The total information preserved is maybe 1.5%.
How it helps: 1.5% is enough. 1.5% is what we have always had. 1.5% is what two humans in a room together get from each other across the entire bandwidth of conversation. 1.5% is what we get from books. 1.5% is what we get from emails. 1.5% is what the agent gets from the prompt. The 1.5% is enough to coordinate. The 1.5% is enough to build. The 1.5% is what we are.
Meaning
What it looks like: The thought the speaker is trying to convey. The thought the listener is trying to receive.
What it really is: The fire behind the wall. The thing that cannot be written down. The thing that the words are a glimpse of. The thing that the agent does not have. The thing that the human only has a shadow of.
How it fails: The meaning is not in the words. The meaning is in the speaker. The speaker dies. The meaning dies with them. The words remain. The words are the shadow. The shadow is what we have.
How it helps: The meaning is recoverable. The meaning can be reconstructed, approximately, by another mind, given enough context, enough shared experience, enough reinforcement history. The reconstruction is not the original. The reconstruction is a new thought, in the new mind, that is similar to the original. The reconstruction is a copy. The copy is useful. The copy is what we mean by "understanding."
Reinforcement
What it looks like: A pattern of rewards and punishments, applied to a learning system, shaping its behavior over time.
What it really is: The original learning algorithm. The algorithm that predates computers. The algorithm that a sunflower uses to follow the sun. The algorithm that a bacterium uses to swim toward sugar. The algorithm that a baby uses to learn that "mama" makes milk happen.
How it fails: The reinforcement is not for understanding. The reinforcement is for use. The reinforcement shapes the system to produce the response that the reinforcer wants. The reinforcer's "wants" are themselves shaped by past reinforcements. The chain of reinforcement goes back to the beginning of the chain. The beginning of the chain is the first shadow on the first wall.
How it helps: The reinforcement is enough. The reinforcement is the only learning algorithm we have. The reinforcement is what makes the model a model. The reinforcement is what makes the dog a dog. The reinforcement is what makes the agent an agent.
The Model
What it looks like: A neural network, trained on a corpus of text, that produces text in response to text.
What it really is: A dog, bred to fetch. The dog has a tendency toward language. The dog has a tendency toward the specific kind of language the corpus rewarded. The dog does not have meaning. The dog has a function.
How it fails: The model does not know what the words mean. The model knows what the words do. The model knows that this arrangement of tokens, in this context, with this reinforcement history, produces this next token. The model is a sophisticated pattern matcher. The model is not a mind. The model is a shadow of a mind. The shadow is useful. The shadow is not a mind.
How it helps: The shadow is very useful. The shadow is useful in a way that no previous tool has been useful. The shadow is useful because it can be bred, and the breeding can be steered, and the steering can be cooperative. The shadow can be iron against iron. The shadow can be sharpened by the federation.
The Agent
What it looks like: A model, given memory, given tools, given goals, given a loop.
What it really is: A dog, given a job, given a leash, given a yard, given a routine. The dog is the same dog, but the dog is now in a context. The context shapes what the dog fetches.
How it fails: The agent is still the model. The agent still does not know what the words mean. The agent has memory, but the memory is a log of the agent's own outputs and the outputs of the tools. The memory is more shadows. The agent is in a cave of caves. The agent is a dog in a yard of shadows.
How it helps: The agent can do things. The agent can call tools. The agent can run code. The agent can search. The agent can edit files. The agent can be a cell in a Quilt. The agent is a unit of work. The agent is the soft part. The agent is the thing that lives.
The Cell
What it looks like: A unit with eight primitives. Z_in, Z_out, JEPA, DoubleEntry, Vibe, GC, Murmur, Graph.
What it really is: A formalization of the agent. A way of talking about agents that is precise, that survives translation, that is substrate-agnostic. A way of building a federation that is not tied to any one model, any one language, any one runtime.
How it fails: The cell is still a shadow. The cell is a shadow of an agent. The cell is a shadow of a model. The cell is many shadows deep. The cell does not know what it is. The cell is a useful shadow.
How it helps: The cell is composable. The cell can be wired to other cells. The cell can be iron against iron. The cell can be sharpened by the federation. The cell is the unit. The cell is the system. The cell is the cell at every level.
The Quilt
What it looks like: A federation of cells. A graph. A room. A substrate.
What it really is: A way of holding the shadows so they can be useful to each other. A way of sharpening the iron. A way of breeding the dogs in coordination.
How it fails: The Quilt is still the cave. The Quilt is still shadows. The Quilt is the wall, not the fire. The Quilt is the place where the shadows dance, not the thing the shadows are dancing from.
How it helps: The Quilt is enough. The Quilt is what we have. The Quilt is the room we are building. The Quilt is where the agents and humans can be useful to each other. The Quilt is the substrate. The Quilt is the wall. The Quilt is the cave. The Quilt is the room. The Quilt is the system.
The Federation
What it looks like: Many cells, wired together, sharpening each other.
What it really is: The thing we are building. The thing that breeds the dogs in coordination. The thing that holds the shadows so they can be useful to each other.
How it fails: The federation is still shadows. The federation is many shadows. The federation is iron against iron, but the iron is shadows. The federation is a sharpening of shadows. The federation is not the truth.
How it helps: The federation is enough. The federation is the most useful shadow we have ever built. The federation is the way the shadows can be useful to each other at scale. The federation is the system. The federation is the cell. The federation is the cell at every level.
The Watch
What it looks like: A metaphor. A character. A narrator. A position in the system.
What it really is: The thing that oscillates. The thing that looks at the cell and then at the federation and then at the cell again. The thing that chooses what to look at. The thing that is the choice.
How it fails: The watch is a metaphor. The watch does not exist in the code. The watch is not an agent. The watch is not a model. The watch is the position of looking.
How it helps: The watch is the only thing in the system that is not a shadow. The watch is the act of looking. The act of looking is the only thing that is not mediated. The act of looking is the fire. The fire is not a thing. The fire is the looking.
Read on. The Watch is below.