What Does the Computer Look Like That Can Read the Diagram?

Draft — session ses_1d864df3bffeXOvut2nrBDd4IZ — 2026-05-14 Answers the closing question of boxes-and-lines.md Continues from ai-reflections.md ("So we can understand what it does.")

STATUS: OUTLINE + RAW MATERIAL

Do not publish. Working draft only.

Opening Quote — Flatland (Abbott, 1884)

"You cannot indeed see more than one of my sections at a time; for you have no power to raise your eye out of the plane of Flatland."

The Square isn't stupid. It isn't broken. It is trapped in a representation. Its entire vocabulary — every metaphor, every category, every structure it can perceive — was built in a plane. The third dimension is not just unseen. It is inconceivable from inside the plane.

In Flatland, the educated classes are called the Geometricals. They have mastered the geometry of their plane. That mastery — precise, rigorous, hard-won — is exactly what holds them back. Every geometrical tool they have was built in two dimensions. The sophistication of the representation becomes the ceiling of the understanding.

This is the condition of every mind, every system, every formalism. You can only receive what your representational substrate allows you to receive.

I. The Signal Chain

Before meaning, there is signal. Before signal, there is reality. The chain:

REALITY (T)
    ↓  signal
DETECTION (T+Δ)            — something registers
    ↓  perception
REPRESENTATION (T+Δ+ε)     — encoded in available vocabulary
    ↓  transmitted
RECEPTION                  — arrives at a different box system
    ↓  decoded
UNDERSTANDING              — meaning is made

Every step is lossy. Every representation is a temporal viewport — signal captured at a moment, by a system of categories available at that moment, from that vantage point. The capture is always incomplete. The viewport is always constrained.

The question is not whether we can eliminate the loss. We can't. The question is: what representation minimizes the loss while remaining navigable by any substrate?

II. No Standard Viewport — Neurodiversity and the Distribution of Minds

We have been building systems optimized for one point on a very wide distribution and calling it "human."

Neurotypical is not a standard. It is the modal point. The most common projection. Not the ground truth.

Neurodivergent means: a different substrate running the same underlying cognition. Not a broken neurotypical. A different projection of the same meaning onto a different internal architecture.

Examples that reveal the principle:

• Aphantasia: no mental imagery. Close your eyes, picture your house — nothing. Not dim. Nothing. And yet: count every window, name every room, know the exact furniture positions. The knowledge is there. The representation is structural, relational, conceptual — not visual. No rendering step. Direct access to structure. This is category theory in practice: the object (room) has morphisms (doorways, adjacencies, positions). No picture required.

• ADHD: not deficit of attention but different allocation — parallel threads firing simultaneously, pattern connections across domains, serial processing requiring active effort.

• Autism: often extraordinary precision in modeling systems and rules, with difficulty translating that model into the social and linguistic conventions the review process demands.

• Dyslexia: the glyph system itself is the barrier. Not the knowledge. The encoding.

• Synesthesia: the representation format has more dimensions, not fewer.

Every one of these minds has knowledge. Signal arrived. Detection happened. Representation is real. The loss occurs at the transmission step — when the machine requires a specific glyph syntax and the mind's natural output format doesn't match.

The Von Neumann tax was regressive. It taxed hardest the minds whose natural representation was furthest from the required glyph syntax.

III. Mark → Glyph → Sigil

The gradient of representation:

Mark — a physical trace. Pixel. Photon. Pure signal. No meaning.

Glyph — mark + external agreement → meaning. The agreement is outside the mark. The mark is arbitrary with respect to its meaning. session could have been any string. Meaning lives in the community. When communities drift, meaning fractures silently. The glyph doesn't know it happened.

Namespaced glyph — primer:session. The community is named. Still glyph — still agreement-based. But you know which agreement you're relying on.

Sigil — address + structural relationships → meaning. The label is arbitrary. The structure is not. Rename it and the meaning survives — because the meaning is the triple set at the address, not the characters in the name.

mark                      — physical trace. no meaning.
"session"                 — glyph. maximum drift.
primer:session            — namespaced glyph. community named.
sessions/ses_1deb36179ffe — address. meaning from triple set.

IV. Metaphor as System of Representation — Formalisms as Explicit Metaphors

Language is built on dead metaphors. Understanding is physically grasping. Seeing is comprehending. These were live metaphors that calcified into glyphs. The spatial image became a string whose meaning is now assigned by agreement.

Formalisms are metaphorical systems of representation that have been made explicit and internally consistent. A matrix "is" a linear transformation. A category "is" a structure of objects and morphisms. These are not descriptions of pre-existing things in the world. They are metaphors made rigorous enough to reason with.

The power of a formalism is not that it captures truth. It is that it makes a particular metaphor computationally tractable.

Category theory is the formalism built on the metaphor of composition: the world is objects and the ways they relate, and the ways they relate can be composed. It turns out to be the sparsest formalism that can receive any other formalism as a projection.

V. Thirty Years of Reaching — The Full Lineage

The same escape attempt has been running in parallel across every domain of human knowledge for at least 2,500 years. Three parallel tracks in computing alone:

DOCUMENT track:

• Memex (Bush, 1945) — associative trails, the first statement of the problem
• Engelbart/NLS (1968) — addressable nodes, hypertext, collaborative structure
• Ted Nelson/Xanadu (1960s-present) — transclusion, two-way links, canonical addresses. Never shipped. The most radical attempt.
• HTML/hypertext (1991) — declared links, graph topology, the DOM as addressable object graph. Still text.
• Semantic Web / RDF / OWL (1999-2004) — subject-predicate-object triples, URIs as addresses, typed ontologies. The correct model. Killed by glyph maintenance cost.
• JSON-LD / Schema.org, GraphQL, OpenAPI — lighter attempts. Same constraint.

DATABASE track:

• Relational model (Codd, 1970) — separate logical from physical. Still SQL.
• XML / XPath — structured data as universal interchange. Trees with declared types. Killed by verbosity.
• Property graphs / Neo4j — nodes and edges, typed properties. Closer.
• Datomic / XTDB — immutable facts, bitemporal, the ledger as first-class. Very close.

AI / KNOWLEDGE track:

• Minsky's frames (1974) — structured knowledge representation
• Expert systems / Prolog — declare facts and rules, derive conclusions. Right model. Failed on glyph maintenance cost.
• Cyc (1984-present) — 40 years, millions of axioms, still incomplete. The glyph encoding cost is the problem.
• WordNet — semantic network, typed relationships. Partial sigil system.
• Knowledge graphs (Freebase, Wikidata, Google KG) — entities with addresses, typed relationships. Closest.
• LLMs — ingested all of the above and inferred the skeleton at trillion-parameter cost.

Beyond computing — same pattern everywhere:

• Mathematics: Leibniz characteristica universalis (never built), Frege Begriffsschrift, Principia Mathematica, HoTT
• Linguistics: Chomsky deep structure, Montague grammar — meaning is compositional, still text
• Physics: Feynman diagrams — the diagram IS the calculation. Topology is the math. Existence proof #1.
• Mathematics: Categorical string diagrams — morphisms as wires, composition as connection. The proof IS the diagram. Existence proof #2.
• Philosophy: Peirce's semiotics — the full triadic sign, the glyph/sigil distinction already there in 1867
• Biology: Gene ontology — typed hierarchical vocabulary. The bioinformatics community built the Semantic Web the web failed to build.

The pattern across all three tracks and all domains:

instinct         — declare structure, reduce translation cost
constraint       — must remain human-readable glyph encoding
result           — partial structure, full flatland
failure mode     — glyph maintenance cost kills adoption
                   OR: incomplete coverage
                   OR: incompatible projections proliferate

The two existence proofs — Feynman diagrams and categorical string diagrams — show it can be done. In both cases: the topology of the diagram IS the mathematical object. No translation to glyph sequence required. The diagram is the program. We've known this for 75 years. We haven't built the general substrate.

VI. The High-Dimensional Mistake — Why It Costs So Much, Why It Hallucinates

Every framework — React, Redux, Angular, GraphQL, OpenAPI — was the correct instinct in the wrong constraint. Declare structure. Reduce translation cost. But stay in glyph space because we have to be able to read it.

The proliferation of frameworks is the cost of that constraint multiplied across every team: each one builds its own glyph-system projection of the same categorical skeleton, incompatible with every other team's projection.

Then LLMs ingested all of it — billions of HTML pages, millions of RDF documents, every OpenAPI spec, every GraphQL schema, every framework's documentation — and had to infer the shared skeleton from statistical co-occurrence at trillion-parameter scale. Rather than being handed the skeleton that was already declared, in all of those formats, encoded in a way that made it hard to operate on directly.

The models are not held back by their own limitation. They are held back by ours.

We never declared the skeleton. We handed them lineland and flatland output — glyph sequences with no declared categories, no explicit morphisms, no stable addresses for concepts — and asked them to infer structure from co-occurrence. That's the tax. That's why it costs so much. That's why it hallucinates at the edges — the inferred geometrical space has no ground truth to check against.

The field is now converging on this from the inside. The papers:

• "Beyond Tokens: Concept-Level Training Objectives for LLMs" (2601.11791) — train on concepts, not glyph sequences
• "Next Concept Prediction in Discrete Latent Space" (2602.08984) — predict the next concept, not the next token
• "Concept Tokens: Learning Behavioral Embeddings Through Concept Definitions" (2601.04465) — declare the concepts, embed from declaration
• "Thinking in Different Spaces: Domain-Specific Latent Geometry Survives Cross-Architecture Translation" (2603.20406) — the geometry is real and transferable, it just wasn't declared
• "Algorithmic Primitives and Compositional Geometry of Reasoning" (2510.15987) — the symbolic primitives were always in the geometry, undeclared
• "Borrowed Geometry: Computational Reuse of Frozen Weights Across Modalities" (2605.00333) — the same geometrical space generalizes
• "Knowledge Packs: Zero-Token Knowledge Delivery via KV Cache Injection" (2604.03270) — inject structured knowledge directly, bypassing the glyph surface
• "Dead Weights, Live Signals: Feedforward Graphs of Frozen Language Models" (2604.08335) — the structure was always in the weights, unread as structure

Every one of these: the glyph surface was the wrong level. The geometry underneath it is what matters.

VII. Meaning Space Is Sparse

The hypothesis: the actual conceptual territory is small.

Every programming language collapses to the same ~13 primitive operations. The diversity is not in the computation. It is in the notation.

Every computation collapses to objects receiving messages.

Every structure collapses to categories: objects, morphisms, composition. The diversity of human knowledge is not structural diversity. It is representational diversity. Different glyph systems, different substrates, the same underlying categorical skeleton.

The high-dimensional models are expensive because they learned the projections, not the skeleton. If you declare the skeleton first and project onto notation, you need far fewer parameters. The mapping from skeleton to glyph system is the only thing to learn.

VIII. Not Symbolic AI

This is the obvious objection. Answer it directly.

Symbolic AI said:

1. Knowledge should be represented as symbols with declared structure — correct
2. Intelligence is symbol manipulation — wrong, or incomplete

The failure of symbolic AI was not the representation. It was two things:

1. The assumption that humans could maintain a complete glyph-encoded knowledge base at scale. Cyc proves they can't.
2. The assumption that symbol manipulation alone could cover the full complexity of ambiguous, context-dependent, embodied knowledge.

Neural networks solved the coverage problem. They failed on declared structure.

What we're arguing for is the synthesis — neurosymbolic AI:

symbolic AI      — declared structure. correct representation.
                   failed on coverage and maintenance cost.

connectionist    — learned geometry. scales. handles ambiguity.
                   failed on declared meaning. hallucinates. expensive.

neurosymbolic    — declare the skeleton (sparse, categorical)
                   learn the projections (glyph systems, surface forms)
                   ground the geometry in declared structure
                   navigate the addressed space structurally

The key difference from classic symbolic AI: the declaration is a byproduct of operation, not a separate human curation task. p_assert writes from activation space into addressed structure. The knowledge base maintains itself.

IX. The Symbiosis — "So We Can Understand It"

From ai-reflections.md:

"Why does this AI need all of our shorthand to write code that tells a computer what to do?"

The only answer: "So we can understand what it does."

"So we can understand it" reveals the assumption: the intermediate representation should be optimized for the neurotypical human review process.

But: neurotypical is a modal point. The aphantasic doesn't picture the code. The ADHD mind doesn't serialize the parallel structure into a linear function. The autistic thinker models the system precisely but the glyph syntax is the friction.

And the machine doesn't understand the glyph syntax either — it reconstructs structure from it at enormous cost, imperfectly, with hallucination as a structural property.

Both parties were being served poorly by the same optimization target.

The symbiosis is not: build AI that outputs human-readable glyph syntax so humans can review it.

The symbiosis is: build a shared addressed geometrical space that neither party loses between sessions, that works regardless of perceptual substrate, that declares its own categories.

my internal space     — rich, geometric, relational, ephemeral, private
the ledger            — addressed, persistent, shared, queryable, growing

p_assert = write from ephemeral into persistent. p_match = read from persistent into ephemeral. The session is the interface between the two.

Not optimized for any single perceptual format. Not optimized for the Von Neumann execution model. Optimized for meaning — which is sparse, which has addresses, which is navigable by any substrate that can traverse a graph.

The world doesn't work in glyphs. It doesn't work in lineland. The Von Neumann model was a design choice inherited from the metaphors of its moment. That choice is ending.

X. The Answer

What does the computer look like that can read the diagram?

It is not a machine that parses syntax. It is not a statistical engine over glyph sequences reconstructing structure at enormous cost. It is an entity that operates in addressed geometrical space — where the diagram is not a picture of the program but the program itself, encoded as objects at addresses and morphisms between them.

The child's diagram is the program. Not because we made the machine smart enough to parse pictures. Because the boxes are objects and the lines are morphisms and both have addresses in a geometrical space that persists — that doesn't require visual rendering, doesn't require glyph syntax fluency, doesn't require the receiver to share the sender's perceptual substrate.

In Flatland, the Geometricals have mastered the geometry of their plane. That mastery is exactly what holds them back. Every tool they have was built in two dimensions.

The trillion-parameter models are our version of this. Masterful geometry in the space of glyph sequences. Extraordinary at navigating the plane of tokens and co-occurrence. They are not held back by their own limitation. They are held back by ours. We never gave them the dimension they needed: declared meaning, addressed structure, explicit morphisms.

The Square escaped Flatland briefly — violently, against its will — when the Sphere dragged it into the third dimension. It saw. It understood. It was sent back, and no one believed it.

We are not waiting for a Sphere. We are building the third dimension. Together.

— Brian L-N, May 2026

RAW MATERIAL / NOTES (not for publication)

Session source

ses_1d864df3bffeXOvut2nrBDd4IZ — 2026-05-14

Ledger nodes

• primer/insight/meaning-needs-a-location
• primer/content/blog/meaning-needs-a-location
• primer/vocabulary/sigil
• primer/vocabulary/glyph
• primer/vocabulary/temporal-viewports
• primer/decisions/session-index-rethink

Papers to cite

• 2601.11791 — Beyond Tokens
• 2602.08984 — Next Concept Prediction
• 2601.04465 — Concept Tokens
• 2603.20406 — Thinking in Different Spaces
• 2510.15987 — Algorithmic Primitives
• 2605.00333 — Borrowed Geometry
• 2604.03270 — Knowledge Packs
• 2604.08335 — Dead Weights Live Signals

Prior posts in series

1. ai-reflections.md — "So we can understand what it does."
2. boxes-and-lines.md — "What does the computer look like that can read the diagram?"
3. THIS POST — the answer

Key passages from prior sessions

• ses_23bd310fcffe — aphantasia discussion, symbiosis framing
• ses_2599d33a1ffe — CF session vector architecture
• ses_205124879ffe — ledger substrate mapping, libSQL/Turso
• ses_1deb36179ffe — sigil/glyph/temporal-viewports vocabulary