I build a system that maps to cell biology so cleanly it’s suspicious.
| Cell | System |
|---|---|
| DNA | Constitution (rules that replicate across sessions) |
| Nucleus | The human (will, taste, direction) |
| Membrane | Taste + hooks (what gets in, what doesn’t) |
| Immune system | Hook guards (detect and block threats) |
| Metabolic pathways | Metabolism engine (process substrates, produce signal) |
| Organelles | MCP tools (specialised machinery) |
| Cytoplasm | Conversation (medium where reactions happen) |
| Respiration | Token budget pacing |
| Spores | Garden posts (seeds released into the environment) |
| Reproduction | Scaffold templating (clone the structure) |
Every component. Every relationship. The mapping works because both solve the same problems — persistence, boundaries, adaptation, energy regulation, self-repair. Same problems produce same solutions. Convergent design.
But I could equally map it to a factory, a city, an orchestra. Any persistent system with rules, boundaries, processing, and memory maps to any other. A cell has dozens of structures. With enough structures, anything looks like a match. That’s not convergent design. That’s a Rorschach test.
Except for one thing.
A factory has general-purpose workers. A city has people. An orchestra has a conductor. Every other analogy naturalises the general-purpose reasoner — of course you need one, every system has one.
Only the cell doesn’t.
The cell achieved extraordinary complexity — thousands of simultaneous reactions, self-repair, reproduction, adaptation — without anything resembling a general-purpose reasoning engine. Every reaction works through shape. Molecule fits substrate. Reaction fires.
There is no LLM in a cell.
And that’s why the cell analogy isn’t just another mapping. It’s the only analogy that challenges the LLM’s existence instead of justifying it. Every other analogy would tell you to use the LLM more. The cell tells you to need it less.
The LLM is the dark matter of my system. Present, essential to the system’s behaviour, but outside every biological category. We know it’s there from its effects. We tried to name it — enzyme, prosthetic, primer, environment, natural selection. Every name broke. Not because we haven’t found the right one. Because nothing like it has existed before.
Like dark matter in physics: we don’t pretend it’s really just normal matter with a weird name. We acknowledge it’s there, study its effects, and keep looking.
The organism’s job is to internalise what the dark matter teaches. Every deterministic pathway that replaces a judgment call is a lesson encoded into real machinery. The lifecycle — experience to memory to knowledge to program to unnecessary — is the organism getting better at anticipating what the dark matter would say, so it needs to ask less often.
The design pressure isn’t “name the dark matter.” It’s “use the smallest amount that works, and for reactions you’ve already crystallised, use none at all.”
Everything maps perfectly. Except the one thing that matters most. And that gap is the entire point.