Show HN: ChatGPT UI for rabbit holes

Original Article ↗ Hacker News Discussion ↗

A new web app reimagines ChatGPT as a branching, hypertext-style interface for exploring “rabbit holes,” where each highlighted concept opens in its own side-by-side panel. Users praise the fast, uncluttered UI and liken it to a dynamic, personalized Wikipedia or note-taking system, while requesting features like tree or canvas overviews, link sharing, keyboard-only navigation, and support for personal or local API keys. At the same time, several comments stress that LLM hallucinations and lack of sourcing limit its reliability for factual research, underscoring how early current UX patterns are for deep learning and exploration with AI tools.

Overall reception

  • Many commenters find the UI fresh, fun, and “more than a gimmick,” praising how easy it is to fall into “rabbit holes” similar to Wikipedia or TVTropes.
  • Several say they could see themselves using it regularly for learning and research, calling it addictive and “infinite hyperlinks.”
  • Others see it as an important UX experiment in how humans might better interact with LLMs.

UX & interaction model

  • Core idea: branching, column-based “cards” where clicking highlighted terms spawns new panels, preserving context.
  • Strong comparisons to:
    • Wikipedia rabbit holes.
    • Obsidian / personal wikis.
    • Andy Matuschak’s “stacked notes” / Miller columns.
    • Mind maps and git-like branch graphs.
  • Many ask for:
    • Tree / map / zoomed-out view of all branches.
    • Parallel branches instead of replacing columns.
    • Back/forward navigation and keyboard-only usage.
    • Ability to resize tiles and keep multiple branches visible.
    • Highlight-to-delve or manual link creation for arbitrary text.

Visual design & usability

  • UI praised as “crisp,” “snappy,” and uncluttered; speed is widely noted.
  • Critiques include:
    • Link styling is too subtle; requests to use classic blue/purple link colors.
    • Confusion that suggestions are just examples and any topic can be entered.
    • Mixed opinions on onboarding: some want a guided walkthrough; others fear it would be annoying.

Accuracy, learning, and limitations

  • Some users happily use it to summarize books, explore technical topics, or learn domains quickly, accepting that truth might be imperfect.
  • Others are wary: LLM hallucinations make it risky as a primary learning or discovery tool versus Wikipedia, which is seen as more reliable.
  • There are examples of clearly hallucinated facts; one person concludes LLMs are better at “doing” than “thinking” for you.

Technical/model questions & sustainability

  • Users speculate about API usage, caching, and which LLM is behind it; reports mention OpenAI 4o, Anthropic, and possibly Groq, but this is unclear.
  • Many request:
    • Ability to use their own API keys or local/OpenAI-compatible endpoints (e.g., Ollama).
    • Open sourcing the code or exposing the UI as a reusable component.
  • Concerns about API cost and rate limits; some sessions reportedly stopped answering.