Show HN: Hallucinopedia

Overview & Concept

  • Hallucinopedia is an AI-backed “encyclopedia of things that do not exist,” written in a deadpan Wikipedia style while the content is absurd and fictional.
  • Articles are generated once per URL slug, cached, and then served as if they were permanent entries.
  • The system explicitly forbids real-world facts and repurposes real terms into fictional entities.

Implementation & Performance

  • Uses a single-page React frontend, Cloudflare Workers, and Cloudflare KV for storage, yielding very low latency.
  • New pages stream from an LLM (Gemini 2.5 Flash Lite); previously generated pages are served from cache.
  • Requires JavaScript; server-side rendering is planned to improve crawler access and non-JS usability.
  • Some users note occasional generation failures tied to exhausted LLM credits or high load.

Humor, Tone, and Comparisons

  • Strong positive reaction: described as Borges-like, Discworld- or SCP-esque, with “Monty Python” and absurdist vibes.
  • Compared to earlier AI “encyclopedias” and joke sites; several argue this one is distinguished by its detailed, carefully designed system prompt and internal lore-building.
  • Some users experience LLM-content fatigue despite enjoying the concept.

Internal Consistency & Feature Ideas

  • Prompt enforces dense cross-linking with per-link “context” to seed canonical lore.
  • Later updates add use of referrer-page context so linked pages cohere better; users notice improved rabbit-hole depth.
  • Requested features include: fake search that “finds” any article, talk/controversy pages, backlinks, graph dumps, better link markup, and avoiding self-referential links.

Impact on Web and AI Training

  • Debate over whether it “poisons” LLM training:
    • Some see it as a humorous data-poisoning honeypot for careless scrapers.
    • Others argue serious LLM training pipelines curate data and will ignore such obviously fictional, labeled content.
    • A few say if the web is vulnerable to clearly satirical sites, it is already failing.

Abuse, Moderation, and Safety

  • Open generation via arbitrary URLs led to mass vandalism: hateful, especially antisemitic, titles and spam, visible via “Stumble” and “All entries.”
  • Users call this “why we can’t have nice things” and report the site quickly becoming unsuitable for kids or schools.
  • Proposed mitigations: limit creation to linked titles, add flagging and captchas, filter or reject offensive titles via heuristics/LLM, hide or prune defaced entries.
  • The maintainers subsequently report adding moderation; details of its effectiveness remain unclear in the thread.