About this library

A working library of AI and LLM research, curated by Adrian Chan at gravity7.com. 1,170 synthesis notes drawn from 1,652 whitepapers, organized into 9 thematic clusters and an interactive knowledge graph.

What this is

This is a personal research library — opened up. For years I have been reading AI and LLM papers and excerpting the passages I found important into a private Obsidian vault. The synthesis layer on top of those excerpts — the questions, findings, and cross-paper connections you see on this site — was generated by Claude using the arscontexta.org Obsidian plugin.

Each entry on this site is one of two things:

• Whitepaper excerpts — passages I selected from arxiv papers as the most quotable or claim-bearing. My curation, lightly tagged in Obsidian. Not full text. Follow the arxiv link on each paper page to read the original.
• Synthesis notes — short claim-shaped distillations of what the literature has shown or asked. Written by Claude via the arscontexta plugin, working from the excerpts I curated. The plugin's prompts are designed to surface claims that span multiple papers and to preserve fidelity to the underlying sources. Each note links back to the papers it draws from.

Where it came from

I curate papers in Obsidian. Each one goes into a topic file as an excerpt with light tagging. Over time those topic files grew into a body of work too large to traverse by hand — connections between excerpts were visible only one or two papers at a time.

The arscontexta.org Obsidian plugin uses Claude to read across the vault, surface cross-paper questions and findings, and write atomic synthesis notes that capture patterns the curation alone could not. As new papers arrive, the plugin suggests reweavings so older notes can be updated against newer evidence. The knowledge graph emerges from those notes and their connections, built inside Obsidian.

This public site lifts that work into a queryable form. Notes link to their source excerpts; excerpts link out to arxiv; everything has a semantic neighborhood you can explore.

How to navigate

Three entry points

1. Knowledge graph — 9 clusters arranged in a force-directed layout. Click a cluster to focus its sub-topics; click a sub-topic to see what is in it; click a note to read it. The graph is interactive — pan, zoom, search the box top-right.
2. Synthesis notes — every claim or question I have written, browsable by cluster.
3. Whitepapers — every excerpted paper. Each links to its arxiv source and to the synthesis notes that discuss it.

Two cross-currents

• Topics — the Obsidian folder structure, preserved. Each topic file is the original excerpt set; some have been renamed for public display (e.g. "Flaws" → "LLM Failure Modes").
• Search — semantic search across notes, papers, and topics. Searches by meaning, not just keywords. Try "parallel reasoning" or "agents collaborating".

What is on a note page

Every synthesis note page has:

• A synthesis title and a 1-2 sentence framing
• A Question / Finding toggle — see the same content as a research question or as a research finding
• Related papers — top semantic matches from the full collection, not just the source topic
• A concept map — local explorer of nearby notes; click to walk the graph
• Open in graph ↗ — see the note's home cluster in the full knowledge graph

A few caveats

• The library contains whitepaper excerpts, not full papers. The excerpts are what I selected as the most quotable or claim-bearing passages. The arxiv link on each paper page gets you the original.
• Synthesis notes are LLM-generated. Claude wrote them via the arscontexta plugin, working from the excerpts I curated. The plugin's prompts constrain the model toward fidelity to source material, but each note still carries an interpretive layer. Every note links back to the underlying paper(s) so you can verify.
• Some recently-added papers carry an auto-summary · synthesis pending badge. These were added directly from an arxiv URL through a separate one-step pipeline (not via arscontexta), so their Q-form and F-form are generated from the abstract alone. Useful for discovery; the deeper synthesis happens when the paper goes through the normal Obsidian-then-arscontexta flow.

What it runs on

• Synthesis layer: Claude via the arscontexta.org Obsidian plugin — writes the synthesis notes themselves, working from my curated excerpts
• Reader layer: Claude Haiku — derives the public-facing Q-forms, F-forms, blurbs, and mini-titles from the synthesis notes
• Embeddings: bge-large-en-v1.5 (1024-dim), local
• Clustering: K-means over note centroids, 9 final clusters
• Graph rendering: force-graph.js
• Hosting: Fly.io
• Primary source: arxiv, with some excerpts from blog posts, technical reports, and conference proceedings

Built and maintained by Adrian Chan. Questions or feedback: adrianchan66 at gmail.