AI Papers

A static webapp (formerly "Frontier AI Research Tracker") that tracks research publications in the interpretability / AI safety / capabilities / reasoning space from Anthropic, OpenAI, and Google DeepMind — with mini summaries, topic + time filtering, and analytics.

Data sources

• transformer-circuits.pub (all papers + Circuits Updates)
• anthropic.com/research and the Alignment Science blog
• openai.com research index
• deepmind.google/research/publications
• OpenAlex (affiliation-indexed works for all three labs — catches arXiv papers never posted on lab sites)
• arXiv API topic sweeps (mech interp, alignment, reasoning, evals, agents) + a daily firehose of the latest cs.LG/cs.CL/cs.AI/stat.ML submissions, theme-gated — so yesterday's relevant papers appear without waiting for OpenAlex indexing
• alignment.openai.com (OpenAI Alignment Science blog) and the Alignment Forum (via the GreaterWrong mirror; community research posts)
• Bulk topic corpus (scripts/collect-topics.mjs): ~35k topic-relevant papers from all orgs, 2012-present, via ~85 OpenAlex phrase queries derived from the interest themes — gives the analytics full historical bell curves instead of a lab-recent sample. Records carry an importance score (0-100: lab affiliation, citations, timeline membership) and the feed defaults to Featured ordering so frontier-lab work stays on top.
• Field baseline (scripts/fetch-baseline.mjs): quarterly counts of all ML papers (OpenAlex) — the denominator for share-of-field interest curves.

How it works

• docs/ — the webapp (vanilla JS + Chart.js, no build step). Served by GitHub Pages.
• data/papers.json — the merged, deduplicated dataset.
• scripts/merge.mjs — merges data/raw/*.json collector output into data/papers.json.
• scripts/update.mjs — daily refresher (runs in GitHub Actions every day at 13:00 UTC / 06:00 PT, see .github/workflows/update.yml; also triggerable manually from the Actions tab): scrapes/queries all six sources — OpenAlex (per-lab affiliation), arXiv topic sweeps, transformer-circuits.pub, the alignment blog, openai.com (sitemaps + RSS), and deepmind.google (sitemap + JSON-LD detail pages) — dedupes, spam-filters, classifies paper-vs-post, and appends new papers with rule-based topics and abstract-derived summaries.
• data/timeline.json — editorially curated breakthrough-papers timeline (LLM-curated, not citation-ranked), shown in the Timeline tab.
• Records carry kind: "paper" | "post" — company announcements are excluded from Analytics and from the default Papers view (toggle under the Type facet).

Local development

node scripts/update.mjs          # refresh data
cd docs && python3 -m http.server 8000

Summaries are auto-generated; verify claims against the papers themselves.