Mike's Paper Reviews — Latent Space

An interactive 3D map of 236 daily paper reviews written by Mike Erlihson. Reviews on similar topics end up near each other in space.

Pipeline

1. Fetch the 236 English-language review markdowns from merlihson/scientific-resources.
2. Embed each review (title + body, capped) with all-MiniLM-L6-v2 (384-d).
3. Project to 3D with UMAP (cosine metric).
4. Cluster in the 384-d space with KMeans (k=10) — projection is lossy, so we cluster on the embedding, not on the 3D coords.
5. Label each cluster by asking gemma4:e2b (Ollama) for a 3–6 word human title over its 25 centroid-nearest members.
6. Render with Three.js — glow sprites, neighbor-web, fog, orbit controls.

Controls

• Drag to rotate · scroll to zoom · right-drag to pan.
• Hover a point → tooltip · click → detail panel with the full review.
• Sidebar cluster row: click to isolate, shift-click to toggle.
• Search: full-text over title and body.

Source + pipeline: shmulc8/mikes-latent-space · Sibling project: Agent skills latent space