An Open Letter to the Cosmos Institute

Every builder’s first duty is philosophical: to decide what they should build for. — Brendan McCord, The Philosopher-Builder

If you are lucky enough to have lived in Paris as a young man, then wherever you go for the rest of your life, it stays with you, for Paris is a movable feast. — Ernest Hemingway

Dear Brendan, and the Cosmos Network,

This is a letter written in public because the work it describes is not private. It is an invitation for Cosmos Institute and the Human-Centered AI Lab at Oxford to partner with Univrs.io and the Sepahsalar Research Lab on a question we share — and which neither of us, I believe, can answer alone.

The question is this: how do we build the civic infrastructure of the AI age such that human flourishing is not an afterthought but the substrate?

What follows is not a pitch. It is an argument.

I. The Stack Inversion

The animating insight at Univrs.io is what we call stack inversion.

As you move down the computational stack — through application code, into runtimes, compilers, network protocols, instruction sets, and silicon — complexity is compressed. That is the point of those layers. A good compiler removes complexity. A good chip hides it. The economic and engineering pressure of every layer below the application is toward optimization, and optimization is, by its nature, a reduction.

But as you move up — from individual users, into communities, into the para-social layer where computational agents have a kind of social existence among us — complexity should accumulate. Not noise; meaningful complexity, selected for by the question does this help people, and the systems they care about, become more of what they are trying to become?

This is the layer at which civilizations live. It is the layer at which Franklin’s subscription library lived. It is the layer at which the Paris of Stein, Pound, Hemingway, Joyce, and Fitzgerald lived. It is, I believe, the layer at which the Cosmos Institute already operates, whether or not it uses these words.

The current AI stack is being built almost entirely from the bottom. That is unsurprising — that is where the capital and the talent and the legible metrics are. But it means we are accumulating a great deal of compressed capability with very little corresponding civic substrate to receive it. The HAI Lab’s philosophy-to-code pipeline is, on our reading, an attempt to address exactly this asymmetry. So is Univrs.io.

II. The Movable Feast Paradigm

We call our framing the Movable Feast Paradigm.

Hemingway’s Paris was not productive because Hemingway was there. It was productive because the scene was there — the cafés, the conversations, the cross-pollination between writers and painters and editors and patrons. Subtract any single individual and the Paris of A Moveable Feast still happens. Subtract the scene and no individual could have produced what they produced.

We believe the same is true of intelligence — biological, computational, or hybrid. The locus of meaningful learning is not the model. It is the network of interaction patterns through which models, humans, and institutions teach each other. We call the knowledge that lives in those patterns distributed epistemology: ways of knowing that exist nowhere in particular and everywhere in aggregate.

The clearest existing example of distributed epistemology working at planetary scale is, we believe, DeepMind’s release of 250 million AlphaFold protein structures. AlphaFold did not need to be present in subsequent biology labs for those labs to learn from it. It deposited a gesture into the para-social layer of science, and that gesture compounded — through use, through critique, through extension — into capability that no single institution could have produced or owned.

That is the design pattern Univrs.io is trying to make available to many more domains than protein folding.

III. The Gesture Protocol

If the locus of learning is the network of interactions, then the question of what agents share with each other becomes urgent.

Sharing model weights is the wrong primitive. Weights are roughly analogous to genes — they sit below the level at which Lynn Margulis’s symbiogenesis operates. Symbiogenesis is the principle that the deepest evolutionary leaps happen not through gradual mutation but through the fusion of whole organisms into new compound organisms. The eukaryotic cell is the canonical example: a fusion, not a mutation.

Software has not yet had its symbiogenesis. NPM and Crates give us code reuse at the module level, but no compositional matrix at the organism level. Two libraries do not merge into a new organism with novel emergent properties. They are merely linked.

Univrs.io, through our Domain Ontology Language (DOL) and the VUDO runtime, is an attempt to build that compositional matrix. The entities that live in VUDO we call Spirits. Spirits do not share weights. They share:

Behavioral gestures — small, transferable patterns of action in context
Selection gradients — the what’s worth doing here signal that shapes downstream behavior
Membrane configurations — the rules a Spirit uses to decide what to admit, what to refuse, and what to fuse with

This is, deliberately, behavioral DNA at a level that can recombine. The central research question of Univrs.io is therefore:

What is the minimal gesture protocol that allows agents to teach each other without centralizing the learning?

This is not a question we will answer with engineers alone. It is a question that needs philosophers — particularly philosophers in the lineage Cosmos draws on: Mill on the conditions for inquiry, Tocqueville on the architecture of free association, Aristotle on the relation between practice and flourishing.

IV. The Inverted Success Metric

A system designed at the para-social layer cannot be measured by the metrics native to layers below it.

Conventional product metrics ask: did the user do what we designed for? That question is appropriate when the goal is optimization. It is the wrong question when the goal is human flourishing, because flourishing — by every philosophical account I take seriously — is a state in which the person determines their own ends, not one in which a system determines them on the person’s behalf.

Univrs.io therefore holds itself to an inverted metric:

Did something emerge that we could not have designed?

A platform that passes this test is one that has successfully respected the autonomy of its participants. A platform that fails it has, however gently, captured them. This is the empirical face, I believe, of Cosmos’s philosophical commitment to autonomy and decentralization. It is also the empirical face of Walker and Cronin’s Assembly Theory — the proposal that life and intelligence are best defined by the accumulation of selectable complexity over time. Univrs.io is, in part, an experiment to see whether that proposal can be tested in software.

V. Why I Am Writing to Cosmos

I am writing because Cosmos Institute is, as far as I know, the only organization in the world whose stated mission is to cultivate philosopher-builders at the seriousness this moment requires. The HAI Lab at Oxford, the Cosmos Fellowship, Cosmos Ventures, the Fast Grants — these are not isolated programs. They are, taken together, the closest thing we have to the Junto, the subscription library, and the American Philosophical Society rolled into one.

I think we can do real work together. I outlined four concrete proposals in the private letter that accompanies this one:

A Cosmos Fast Grant or Fellowship focused on the gesture protocol research question above, hosted between Sepahsalar Lab and the HAI Lab.
A philosophy-to-code seminar series in which each session is paired with a running VUDO prototype the participants can interact with — philosophy that survives contact with executing code.
Open-source civic infrastructure — Univrs.io, VUDO, and DOL are being designed from the foundations to be open and to resist capture. They will benefit enormously from Cosmos’s network of philosopher-builders as both contributors and stress-testers.
A shared public frame for the para-social layer — the field lacks language for what sits above social media and below “AGI.” This is a writing project as much as a software one, and Cosmos is uniquely positioned to shape it.

VI. A Portfolio Piece

If it is useful — and especially if any reader of this letter is wondering what philosophy-as-code actually looks like in practice rather than in essay form — I offer ardeshir.io/barnyard as a small, deliberately scaled portfolio example. The Barnyard is a playful, ecological model of autonomous agents interacting in a shared world. It is not a product. It is a thinking-tool, in roughly the spirit that Franklin’s lightning rod was a thinking-tool: small enough to be understood completely, and large enough to suggest the shape of something much bigger.

It is also an honest example of what a philosopher-builder portfolio piece looks like when the goal is to think with software rather than to ship it.

VII. Closing

You once wrote that philosophical seriousness creates a negative selection gradient — and that you want it that way, because the people who do the reading are the people most likely to build something different. I read that line and recognized my own design choices.

Univrs.io is not built for a viral curve. It is built so that, when the people who do the reading arrive, there is a substrate worthy of them.

I would be grateful for a conversation.

— Ardeshir Sepahsalar Founder, Univrs.io Research Lead, Sepahsalar.org/research

Where to find the work

univrs.io — the host environment
vudo.univrs.io — VUDO runtime and Game-of-Life-class demo
sepahsalar.org/research — Sepahsalar Research Lab
metalearn.org — recursive, self-directed learning framework
ardeshir.io/barnyard — philosopher-builder portfolio piece
ardeshir.io/cosmos — companion publication of this letter