In 2020–21 the LLM substrate became real. Massive, general, available. We treated it like the whole organism. But cytoplasm isn't a body — it's medium. The work is done by what lives inside it. Mitochondria. Little models. Specialized organelles drawing on the substrate to produce specific cellular work.
Below: a Memeroot canvas drawn as the cell it is. The faint dot-pattern behind everything is the LLM substrate — pervasive, ambient, the medium. Inside the cell are eight mitochondria, each one a feature that produces a specific kind of cryptographic or compositional work. Click an organelle to read its part.
Hover or click any mitochondrion to see what it does in the cell. Each is a real feature shipping in the canvas right now — eight little models, eight specialized productions, eight arias.
In 2020–21, the LLM substrate became real. A massive, general, mostly-API-callable medium that any program could draw on. The temptation was to treat it as the whole organism — "an LLM that does everything" — and a lot of energy in the field went into making single calls go further.
But a substrate isn't an organism. Cytoplasm isn't a cell. The substrate enables work; it doesn't do the work. The work is done by what lives in the substrate: small, specialized, autonomous-with-their-own-genome units. Organelles. Mitochondria.
The next era of AI isn't bigger models. It's smaller, specialized models composed inside cellular substrates that can sign, hash, chain, validate, transform, and export their work without trusting the host.
Mitochondria are a particularly precise metaphor:
If mitochondria is the biological metaphor, arias is the cultural one. An aria is a solo voice over the orchestra — the substrate that supports without dominating. The orchestra plays continuously; the singers come and go, each with their own piece, drawing on the underlying harmony but producing something distinct.
That's the right picture for an LLM-substrate architecture. The substrate is always there, hummed by every call, available to every feature. But the actual music — the user-visible work — is made by individual voices: signing one piece of text, sequencing another, validating a third. Each its own aria, each able to be performed solo or in chorus, each over the same standing substrate.
Provenance survives the substrate. When a mitochondrion does work, the work is signed by you, not by the substrate. The substrate is anonymous medium; what comes out of the cell is attributed. No matter how the substrate changes — different model, different provider, different generation — your signatures, your chains, your structures persist.
Specialization without lock-in. A mitochondrion can be replaced. The signing organelle can be swapped for a different signing organelle (RSA instead of ECDSA, say) without rewiring the cell. The substrate can be replaced (Anthropic for OpenAI, GPT-style for Claude-style) without rewriting the features. Both are pluggable; neither is the architecture.
Composition is biological, not theoretical. You can have many mitochondria of the same kind doing different work in parallel. Two signing organelles for two identities. Three streams for three concurrent research lines. Eight schemas for eight content shapes. The cell scales by replicating organelles, not by upgrading a single brain.
Transferable as cells, not blobs. The Memeroot bundle is the cell, packaged. The recipient drops it in their own canvas — same cytoplasm, same substrate — and the organelles light up, the chain still verifies, your signatures still hold. A platform-bound document is dead text. A Memeroot bundle is a transplantable cell.
It is not an attempt to replace the LLM. The LLM substrate is necessary; without it, the cell has no medium. Memeroot doesn't compete with the substrate; it organizes work that uses the substrate.
It is not multi-agent in the loose sense of "many AIs chatting." Mitochondria are not separate intelligences. They are specialized procedural units — cryptographic operations, validation logic, presentation transforms, AI calls. Each one is small enough to read, audit, sign. Each one has a specific job.
It is not a framework that owns your work. The cell is yours. The organelles are yours. The substrate you call is whichever one you choose. Personal · open · free.
The web after this is not a web of LLMs. It is a web of cells, each running its own mitochondria, each drawing on whatever LLM substrate is available, each producing signed work that travels.