Chart forms borrowed from Goodfire’s “already clean” essay, rebuilt on the live corpus.
The corpus as a city kept
Pattern: “A map of every neuron, drawn as a city” + “recursive explainability” (semantic zoom)
Districts are the three roots of the taxonomy, and blocks are structures sized by how many papers report them. Each small square inside a block is one paper, and darker squares are newer. Click a block to zoom in. Where the structure has a shape (a circle, helix, torus, simplex and so on), the zoom draws that shape with every paper as a point on it, oldest to newest along the shape. The placement is schematic: it orders papers by date and does not use measured coordinates. Drag to rotate.
Model families as a city new
Same form, other axis: who gets studied
Districts are model domains, and blocks are model families sized by how many papers analyse at least one of their models. Each square is a paper, and darker squares are newer. Click a family to zoom in.
Structure genealogy moved from /genealogy · dots
Patterns: per-row unit dots + “every paper, one dot” crowd underneath, linked by hover
One row per structure, ordered by first appearance with the oldest at the bottom. Each dot is a paper reporting the structure, placed at its arXiv month, and colour is the structure’s role (roles ordered by when they first entered the map, so warmer arrived later). A ★ marks a structure reported by two papers that share no author. The crowd underneath has one dot per paper. Hover a row to light up its papers in the crowd, and click a name to open the structure. This does not claim that one structure was derived from another, because the schema doesn’t encode lineage.
Findings by mathematical role revised
Pattern: “Five circles” (angle = category, radius = rank or time, click to isolate)
Each sector is a structure type (the classification’s type axis), and each dot is a finding placed by its first structure. The radius is time. By rank gives each ring an equal share of findings, so the crowded recent years get room. Newest in centre flips the direction. Click a sector or a domain to isolate it.
When each structure arrived new
Pattern: unit dots on a real time axis, one swarm per row
One row per structure with at least four findings. Each dot is a finding at the date of its earliest paper, and dots push apart vertically instead of overlapping. Stretch recent uses a log scale back from today, so the last two years are as wide as the decade before. Click a dot to open it.
Structures that appear together new
Pattern: edge motifs and attribution-graph hover (highlight one node’s edges)
Each node is a structure with at least three papers, sized by paper count. An arc joins two structures reported in the same paper, and thicker arcs mean more shared papers. Hover a node to see its neighbours, or click it to pin them.