Matthew, here. For the last 15 years, I've dreamt of creating a hieroglyphic writing system for English. The idea might have first come to me while I was studying Japanese in high school. I loved being able to look at a Japanese character and feel what it meant before I fully understood it.
Chinese hanzi and Egyptian hieroglyphs work the same way. The water radical 氵 tells you you're dealing with a lake or an ocean. The insect radical 虫 tells you a hanzi might be a mosquito or a beetle. The characters carry semantic meaning visually, independent of sound.
I felt stifled by our alphabet. Every word is just a sequence of sounds with no visual richness. You basically know it or you don't.
Locus is my attempt to fix that — a visual notation system built on Wikipedia's knowledge graph, where every glyph maps out where a concept lives in the universe of human knowledge. Every point is a real Wikipedia article. Give it enough room and that point becomes a glyph of its own.
These are live Locus glyphs for Jazz, Plato, Albert Einstein, Buddhism, and the Roman Empire — painted by the same canonical renderer as the reader.
You'll learn to read them. Scroll to find out how.
There is one canonical Locus glyph, not a family of visual fonts. It keeps the same grammar at every scale. Read the shell first, then the field.
The outer ring gives the broad territory at a glance: Arts, Earth, Human, Human History, Life, Mathematics, Philosophy, Science, Society, or Technology. Ten fixed colors. Learn them once.
The pale notch sits at that domain's permanent clock position. Rose is always 12 o'clock; Science is always around 8:30. Hue lets you spot it quickly. Position lets you verify it — even in grayscale.
The current silhouette distinguishes a person, place, event, organism, created work, organization, object, product, or abstract concept. It must follow Wikidata's is-a hierarchy rather than exact-match a short whitelist. The cue is useful; its placement is still open. The final center may become the most defining subglyph or disappear, with noun class retained as a smaller mark.
Every interior point is a Wikipedia article. Its core uses that child's primary-domain color. Its halo alone carries that child's secondary-domain mixture. Primary identity stays in the core; supporting context stays around it.
A few neutral tethered anchors form the subject's defining sentence. They are chosen for orientation: structured is-a heads first, then the highest-weight real linked articles, with conservative category heads only filling genuine gaps. Obvious anchors are allowed to be load-bearing. Country belongs inside China; rap music belongs inside Tupac.
The subglyphs slider is a true count-versus-size control: move left for fewer, larger recursive clues; move right for more, smaller ones. Ten is the current default and the reader can expose up to the available 48-anchor supply. Rarity size can enlarge less-familiar clues, while familiar depth keeps common learned tokens visible but recursively shallow.
Beneath the sentence, ordinary articles remain compressed noun marks at their stable atlas coordinates. They are fingerprint texture and Wikipedia addresses, not headline clauses. Background dust is off by default. Sentence terms are always protected, even at zero dust.
Every article keeps one frozen anchor everywhere. Below 3 pixels it is a plain domain dot; around 3–5 a known noun seed and clock tick resolve; around 5 pixels the full ring and fixed notch join. A legacy node whose noun class has not loaded remains an honest domain dot rather than pretending to be a concept. Selected sentence anchors open into recursive fields from roughly 12 pixels onward; they are magnified at constrained sizes and may fan a short distance to avoid overlap, with a tether pointing home. Ordinary field marks stay compressed until explicitly entered, so geometric luck can never masquerade as meaning.
The single cream arc carries the lifespan or historical span in the outer black rim band. A dark keyline and two rounded endpoint beads separate it from the domain ring; the fixed domain notch is always repainted above it. A small open circle with four cardinal crosshair ticks places geocoded subjects on the flattened world map. The translucent mark survives at every metadata-bearing recursive scale, while its growth is capped on the largest glyphs.
You do not begin by memorizing thousands of pictures. You make one inference at a time, from the most stable cue to the most specific.
Locus glyphs can also sit inline beside the words they annotate — like kanji in Japanese text, or footnotes that carry their own semantic weight. At that scale the shell is enough; the same glyph reveals its field when opened.
Notice how your eye moves through the passage differently. The glyphs are load-bearing walls — but they are not a second decorative font. Each is the same recursive object shown with fewer pixels.
Early versions made surprise the hero. They muted obvious same-domain neighbors and amplified unexpected associations. The pictures could be beautiful while the reading was wrong: an entertainment glyph could hide entertainment; Tupac could omit rap; weak category backlinks could outrank the subject's identity.
The new law is orientation before surprise. Structured P31/is-a heads receive a floor. Real backlink neighbors then keep their stored weight order. Conservative category heads may fill a sparse sentence, but they may not replace stronger evidence.
Sparse truth is part of the language. If an article has no trustworthy neighborhood row, Locus keeps its domain, type, time, place, and structured orientation anchors—and leaves the rest empty. It never manufactures a cloud merely to fill the circle.
The system is honest. Concepts that occupy similar neighborhoods should share a family resemblance. But the shell, noun shape, frozen anchor markers, tethered fan-out, and relationship halos give you several independent ways to tell them apart.
The first 48,729 vital articles form a frozen atlas. The same article keeps the same coordinate everywhere and across renderer versions. A visual shell may fan away from a collision, but its tether points home; changing the number of visible subglyphs never repacks the map.
Beyond the atlas, a one-card generator extends the long tail. Generated cards use the same orientation contract and enter an epoch-scoped cache. The server prewarms the Sacred roots first, then the articles most often reused as subglyphs, in durable 500-card batches.
A private local mirror can keep raw cards, compressed immutable batches, a manifest, and checksums. The mirror is data only: it does not replace the visual grammar or move the frozen atlas.
In 1679, Gottfried Leibniz dreamed of a characteristica universalis — a symbolic language where every concept's sign would encode its relationships to all other concepts. You could read the sign and know where the idea lived in the universe of knowledge.
He didn't have Wikipedia. He didn't have a knowledge graph built by millions of editors over twenty years. He didn't have color screens.
Locus is built on Wikipedia's graph — the largest collaborative map of human knowledge ever assembled. The glyphs aren't individually designed. They're derived. Jazz begins with the facts and strongest links that make it Jazz; every one of those articles can then open into a field of its own.
That recursion is not decoration. It is what turns a static fingerprint into a navigable language: one frozen anchor system, one readable fan-out rule, and one grammar from dot to shell to field. The goal is simple: nothing should be completely foreign. Every unfamiliar concept should have a neighborhood and a path back to things you already know.
"The system is at an early stage. The weights are imperfect. But already — after twenty minutes — you start to recognize patterns. You start to read."
— Matthew, Bangkok, 2026
Built on Wikipedia's graph: Wikidata type hierarchies, weighted backlink neighborhoods, conservative category fallbacks, frozen coordinates, and recursively generated cards. The center, reticle, physical-size law, comparison view, watchface grammar, and learning system remain active design questions. An early proof of concept — feedback welcome.