FROM PARTICLES TO MEANING · AN OPEN EXPERIMENT
A field that learns its shape.
Many information systems depend on explicit schemas, indexes, and retrieval logic. The Semantic Field explores another route: label the data, give each word a vector, and let relationships form through interaction. The work of forming the field is intensive. The promise is a durable structure that can be saved, revisited, and used again.
The particle premise comes from The Vacuum Experiment (book 7, page 15). The dot, pulse, container, and blue/red/green vocabulary comes from The Nyarindo Lens (book 1, pages 7–10).
01 · THE IDEA
How the field finds its shape
This is a single-file browser experiment that runs locally without a server. Each word becomes a node carrying a vector: either from the hand-coded six-dimensional set or from an uploaded GloVe vocabulary. Cosine similarity influences how strongly two nodes attract. A Lennard-Jones-style force pushes nodes apart when they get too close.
You do not manually place every relationship. Training lets the nodes negotiate positions through attraction and repulsion. Clusters and bonds emerge from those interactions. The display begins in two dimensions and can climb to another dimension when doing so improves the green count.
The experiment makes the negotiation visible through two different color languages. Motion-state colors describe what nodes are doing; bond colors describe the relationship between nodes. A dot and a line carry different meanings even when they share a color.
Motion state · dots
Bond strength · lines
02 · FORMATION AND RETURN
The aluminium idea: intensive once, durable after
Think of aluminium. Its first extraction is expensive and energy-intensive. Once formed, the metal is durable and can serve again and again. The analogy is the shape of the cost: a demanding first formation, followed by repeated use of something that has already been made. (book 2, page 13)
Training is the extraction. For the same vocabulary and configuration, the field can be trained, saved, and loaded again without paying the full training cost for every read. The settled positions and payloads travel in the exported field.
This does not make retrieval free. New or changed data can require more work, and each query still performs computation. The aim is to make the expensive formation reusable rather than repeat it for every encounter.
A query is meant to disturb only a small part of the field. In the design described here, a temporary query node locates related words, wakes its neighbourhood, and is removed when the field settles again. This local-return idea is the direction of the experiment; the current browser implementation still has a wider query pass, described under Honest limits.
03 · THE NYARINDO LENS (book 7, pages 8–13)
Deep equilibrium and the stillness that only looks deep
Book Seven, The Vacuum Experiment, returns to a question that matters here: what does it mean for a field to settle? In the book’s particle simulations, very different systems could show green on the equilibrium counter. The counter alone could not tell whether a structure was deeply held or merely appeared still. (book 7, page 8)
The distinction emerged under pressure. When density increased and particles were brought closer, some apparently settled structures broke apart. More strongly bound structures held. In the book, deep equilibrium is not simply the absence of visible movement; it is a balance formed through relationship that can endure a changed container. (book 7, pages 8–9)
That gives the Semantic Field a useful question to carry forward. A stable green count or a stale field is a sign that the simulation has paused; it is not, by itself, proof of deep equilibrium. A stronger test would ask whether useful relationships persist as the vocabulary grows, the field is revisited, and its conditions change. (book 7, page 9)
The book’s vacuum experiment and this semantic database are separate simulations. The connection is conceptual: the Lens asks us to look beneath apparent stillness and test what holds when the container changes. This browser database has not yet demonstrated the book’s stress test. (book 7, page 13)
04 · IN PRACTICE
Train, write, read, store
- Train: Enter words and train the field. Watch the motion-state colors move as the nodes negotiate. Inject a word to wake nearby nodes.
- Write: Insert a key with a value. The key is embedded as a node and the value travels with it as payload. Batch insert accepts one
key :: valuepair per line. Vocabulary upload can overwrite or extend the field. - Read: Query a word already in the field, choose how many neighbours to return, and review similarity scores. The minimum-similarity control filters weaker matches. A word the field does not contain returns no result instead of a guess.
- Store: Export saves the settled field, its positions, and payloads as JSON. Import restores a saved field.
05 · WHAT THE EXPERIMENT DOES AND DOES NOT SHOW
Honest limits
- Training is O(n²) per tick. The field can take real work to form.
- Retrieval is not O(1). The current code checks for an exact label, scores all current nodes by cosine similarity, sorts them, and then wakes the selected results. Scoring is linear in node count; sorting adds work. The local physics then runs for the woken nodes.
- The current code uses a matching node already in the field; it does not yet create and remove a separate temporary query node. The temporary-node and local-only query model described above is a direction to implement.
- The supplied project description reports testing close to 500 nodes in the browser. That is a reported test size, not an independent benchmark here.
- GloVe represents co-occurrence patterns; that is not the same as grounded meaning. Browser words outside the vocabulary can be hash-placed, which does not give them a semantically grounded vector.
- The hand-coded six-dimensional vocabulary is limited to its authored entries. An uploaded embedding brings its own coverage and assumptions.
These limits shape the next experiment. A field that looks settled should be tested under added data and changed conditions. A query that wakes a few nodes may still have done work across the whole field to rank them.
06 · POSSIBLE DIRECTIONS
Where the field could lead
A memory that deepens on return
Conversations, notes, and stories could become semantic nodes. Each return could add a new observation and strengthen links to earlier material, while keeping the raw report distinguishable from the story built around it. A memory system could retrieve a bounded neighbourhood to help form a context window, rather than repeatedly passing the whole archive. Bounded prompt size is a design goal; constant retrieval cost has not been demonstrated. Privacy, user control, and the ability to inspect or remove stored material would be essential.
AI training and data preparation
A field might help arrange related examples, reveal clusters, or select material for training. Whether that improves training quality, cost, or efficiency would need to be tested against other approaches.
A semantic operating system
Payloads could eventually hold programs, commands, or other fields and help route work by relationship. That is one possible path toward a new operating-system architecture. This prototype stores and displays payloads; it is not an operating system and does not execute them.
The Semantic Field is an experiment in making relationships visible and reusable. Its future depends on what survives careful testing, repeated return, and the pressure of a larger field.