About Exodia

Distilling SotA open-endedness through open-endedness.

Exodia is a self-updating pipeline. When the upstream reading list changes, it re-ingests the entries, enriches them with abstracts, full text and citation counts, analyzes the consensus concepts and themes, asks an automated scientist for new ideas, and rebuilds this site — the loop shown on the home page.

A note on the name: exodia tracks and distills open-endedness research. It is inspired by an open-ended loop, but it is not (yet) an open-ended system itself in the technical sense — perpetual novelty and learnability, as formalized by Hughes et al. (2024). See Future work for how it could get there.

Method

1. Gate. Poll the upstream commit SHA; run only when it changed.
2. Ingest. Parse the upstream README into structured entries.
3. Enrich. Add abstracts, full-text PDFs and video transcripts, and citation counts (arXiv + Semantic Scholar; rate-limited, cache-first). The full text is mined for analysis only and never redistributed.
4. Analyze. A curated concept gazetteer plus TF-IDF + clustering surface the field's concepts and consensus themes (tfidf_kmeans).
5. Ideate. Invoke AI-Scientist-v2 to generate & summarize ideas.
6. Render. Build this site, plots, and the changelog; deploy to Pages.

Credits & attribution

• The knowledge base is distilled from awesome-open-ended, curated by Jenny Zhang. The upstream list publishes no license, so we store only factual metadata (titles, authors, venues, links) and link back to the originals — we do not republish upstream prose.
• Idea generation and summaries come from AI-Scientist-v2 (Sakana AI), used under its source-code license. Machine-generated content is labeled as such wherever it appears, as that license requires.
• Abstracts are retrieved via the arXiv API and shown with attribution and a link back. Full-text PDFs and video transcripts are fetched into a local, git-ignored cache for analysis only — they are never re-hosted or republished; the site always links back to the source. Citation counts come from Semantic Scholar.

This project's own code is released under the MIT License; no third-party source code is vendored. See the repository NOTICE file for full details.

Future work

The loop is intentionally left open. A future step could close it by invoking AI-Scientist-v2's full experiment-and-write-up pipeline to draft an actual paper from the highest-consensus ideas. That step is not implemented here and appears as a dashed branch in the loop diagram.

Closing that feedback is also what could make exodia genuinely open-ended rather than a pipeline that merely tracks the field: generated ideas seeding work whose results re-enter the knowledge base and reshape the next round, steered by a model of interestingness toward the novel-but-learnable frontier, with novelty and learnability measured across runs (in the sense of Hughes et al., 2024). That is the project's intended direction, not its current state.