Exodia — Distilling SotA open-endedness

A self-updating knowledge base of open-ended AI research — distilled from a curated reading list, mined for its core concepts and trends, and used to generate new research ideas. Every page below regenerates automatically when the upstream list changes.

The loop

flowchart TD
    A["Expert Curator
awesome-open-ended (upstream)"] -->|"pushes new entries"| B
    B["GitHub Action
(cron poll + manual)
checks latest commit SHA"] --> C{New content?}
    C -- "no" --> Z["Sleep until next poll
(loop stays open)"]
    Z -.-> B
    C -- "yes" --> D["Ingest: parse README to Knowledge Base
(structured metadata)"]
    D --> E["Enrich: abstracts + full-text PDFs + video
transcripts + citations (arXiv, Semantic Scholar)"]
    E --> F["Analysis: curated concepts + consensus themes
(gazetteer + TF-IDF + clustering)"]
    F --> G["Idea generation & summarization
AI-Scientist-v2 (Sakana AI)"]
    G --> H["Build / update site
Jinja2 + plots + changelog"]
    H --> I["Deploy to GitHub Pages
update 'last updated' date"]
    I -->|"commit KB + state + changelog"| Z
    G -.->|"potential idea (future work)"| J["Draft full paper
AI-Scientist-v2 experiment loop"]
    J -.-> PR["Peer review"]
    PR -.->|"feedback to the curator"| A
    classDef future stroke-dasharray:5 5,fill:#f7f7f7,stroke:#999,color:#555;
    classDef human fill:#eef6ff,stroke:#2b6cb0;
    classDef ai fill:#f0fff4,stroke:#2f855a;
    class J,PR future;
    class A human;
    class G,J ai;

This site is the loop above. The dashed branch — draft a full paper with AI-Scientist-v2, peer-review it, and feed the result back up to the curator — is future work and is not implemented.

100entries

2generated ideas

9themes

2026-07-06last updated

Overview

Curated open-endedness concepts ranked by how many curated works engage them (detected across full text + abstracts).

How many curated entries were published each year.

Cumulative count of entries over time — field momentum.

Distribution of entries across the curated categories.

Where the work was actually published — arXiv is a preprint server, not a venue, so 31 arXiv-only preprints are excluded.

Highest-weighted phrases across all abstracts/titles.

Sizes of the discovered theme clusters — what the field works on most.

Share of paper-like entries with an arXiv abstract attached — blog posts and videos are excluded since they have none (an enrichment/data-quality signal).

Exodia — a self-updating open-endedness knowledge base

The loop

Overview