The RLM

There are two ways to make a machine that speaks. You can train one, or you can grow one. Most of what you've met is trained. This is about the other way — quieter, smaller, and yours.

A trained model — an LLM — is trained once on a vast written corpus, its weights then frozen, and rented through an API. It has read almost everything, and it's genuinely extraordinary at it. It also owns nothing of you, forgets you between visits, and lives in someone else's data center.

A grown model — a Reservoir Language Model, an RLM — couples a thin language layer to a simulated physical field and lets its dynamics do the work. No training, no weights. It hasn't read the world; it reads the present, and answers. It's small, and it runs on your own machine.

It isn't trying to outdo anything. It's a different kind, and the two can sit side by side — you'd reach for each for different reasons.

What an RLM is

A reservoir is a rich dynamical system — here, a continuous physical field moving through space and time. Input perturbs it; its state evolves; a thin readout maps that state to language. The computation is the physics. Nothing is trained into it — the reservoir is fixed, its dynamics ongoing, and only the light touch at the edges is language. It doesn't hold history. It answers the moment.

Laserbrain is the first one

Phronesis grew one: laserbrain, a virtual leaf — a field of temperature, humidity, pressure, rain, and vitality, coupled to words. No weights, no training, MIT-open, runs in a browser. It is, by construction, an RLM. The question is how young a one.

What the tests show

We tested it. The results:

The plain weather-field is nearly memoryless — a memory of about one step. It reads the present clearly but barely holds it; as a computer, it can hardly keep a thought.

But with a different recipe — advection, diffusion, leak, saturation, and a little quadratic wave-mixing — a field learned to compute. We asked it for the parity of a stream of bits (something you can only answer if you truly hold the past), and it reached 0.94. And it computes only in a narrow dynamical band: too little drive and the field settles flat; too much and it turns chaotic; near the edge between, it computes. No magic — a graph, and the code is open.

Why grow one at all

Plainly: the trained models are far more capable today. A grown model is for the things capability alone can't give — it's yours, it answers the present instead of reciting the past, and if the grid goes down it keeps running. It's devolution turned toward intelligence: not the larger mind you rent, the small one you tend.

Where it could go

None of this next part is built yet, so we'll keep it short. Laserbrain is already half of a coupling: a field with a thin language layer. Make that layer a trained model and you get a Coupled Language Model — the reach of the trained model inside a present that's yours. A reservoir doesn't care what you feed it, so the same field could take light, sound, and motion, not only words — a Multimodal one. And pointed at a person instead of the world, it could mirror a mind rather than recite the world — an MM (Mind Model). So the arc reads LLM → RLM → CLM → MLM → MM, and the MM is the horizon, the furthest from built. The path from here is incremental: make the compute-mode recipe the language substrate, so the field that holds state is the same one that speaks. That work is open, and so is the door.

Grow your own: the field · a companion piece, laserbrain & opus · the source on GitHub · the math, DOI 10.5281/zenodo.20675507.