Apeirron

apeiron (ἄπειρον) — ancient Greek for "the infinite, the boundless, the undefined origin of all things."

An interactive knowledge graph mapping the biggest questions humanity asks — consciousness, ancient civilizations, the nature of reality, hidden power structures, the cosmos and many more — as interconnected nodes in a visual web.

Every idea is a node. Every node links to others. Every connection has a reason. The result is a web of thought where nothing exists in isolation.

How it works

The site is a force-directed graph. Each node is a topic — written as a narrative deep-dive, not a Wikipedia summary. Click a node to read it. Follow [[links]] in the text to fall deeper into the rabbit hole. The graph grows as contributors add new nodes through Pull Requests.

All content lives as Markdown files in the content/nodes/ directory. The graph, connections, and site are generated automatically from these files at build time. No database, no CMS — just Markdown and Git.

RSS feed

New nodes are published to an RSS feed at /feed.xml, ordered by date added with the newest node on top. Subscribe to follow the graph as it grows.

Books

The same content is also available as a typeset edition: seven EPUB and PDF volumes, one per category, generated from the same nodes. See books/ for the build pipeline and details.

Audio narration

Every node can be narrated. The speech/ directory holds a small local text-to-speech pipeline built on Kokoro-82M (Apache-2.0): it normalizes a node's Markdown into clean, speakable prose and reads it in a natural voice. Synthesis runs locally on the CPU — no API keys, no cost, and no text ever leaves your machine — and the 82M model generates roughly 12× faster than real-time.

Published narrations are stored as MP3s in a Cloudflare R2 bucket, served from audio.apeirron.com — not committed to git. A node page shows a Listen player automatically once the node's id appears in public/audio-manifest.json; nodes without a published file fall back to the browser's built-in speech synthesis.

Setup

cd speech
brew install espeak-ng        # system dependency for pronunciation
uv sync                       # installs Kokoro + the spaCy model (en_core_web_sm)

Generate and publish

# 1. preview exactly what will be spoken — fast, no synthesis (catches bad text)
uv run python generate.py --check ../content/nodes/consciousness.md

# 2. generate the MP3 locally and listen to it
uv run python generate.py ../content/nodes/consciousness.md consciousness.mp3

# 3. upload it to R2 and add the node to the manifest (needs R2 access — maintainers)
uv run python publish.py consciousness

# 4. commit the manifest + waveform and redeploy so the site shows the player
git add public/audio-manifest.json public/audio-peaks/ && git commit -m "feat(audio): narrate consciousness"

Choose a narrator with --voice (suggested: am_michael (default), am_puck, bm_daniel, bm_fable, bm_lewis); run generate.py --list-voices to see them all. Anyone can generate and listen locally; only publishing (step 3) needs access to the project's R2 bucket.

Text normalization. The cleaner (speech/clean.py) strips frontmatter, the Sources bibliography, Markdown syntax and footnotes; turns [[wikilinks]] into spoken words; and fixes what the TTS model mis-reads (e.g. the Greek gloss in phi (Φ), superscripts, °C). It intentionally leaves numbers, years, decades, currency, and initials alone — Kokoro's G2P already handles those — and it preserves paragraph breaks so Kokoro chunks sentences without breaking names at their middle initials.

Gotcha — the Hugging Face network check. On first run the model weights download from Hugging Face and are cached. After that, huggingface_hub still pings the Hub each run to check for updates — a metadata check about the public model only, never your content — which surfaces as a unauthenticated requests to the HF Hub warning. To skip it and run fully offline once the model is cached, prefix the command with HF_HUB_OFFLINE=1 (the only catch: offline mode can't fetch a voice you haven't used before).

Contributing

Apeirron is open to contributions. You can:

• Add a new node — write a deep-dive on a topic and submit a PR
• Improve an existing node — better writing, more connections, factual corrections
• Propose a topic — open an issue if you have an idea but don't want to write it yourself

Every node must include verifiable sources — books, papers, videos with timestamps, or official documents. PRs without sources will not be merged.

Read the Contributing Guide for details on how to write a node, how connections work, and what makes a good submission.