Launch HN: Undermind (YC S24) – AI agent for discovering scientific papers

Hacker News Discussion ↗

An AI-powered literature search tool for scientific papers is drawing praise for uncovering hard-to-find, highly relevant studies that many users missed with Google Scholar, arXiv, and other research platforms. Commenters highlight its strength on complex, topic-level queries and meta-analyses, while noting trade-offs such as slow, compute-heavy searches, reliance on abstracts rather than full texts, limited access for non-institutional users, and relatively high pricing. Comparisons with tools like Elicit, Semantic Scholar, and Exa focus on accuracy versus speed, as well as calls for features like APIs, PDF export, better ranking by importance, and iterative refinement of search results.

Overall reception

  • Many commenters are impressed; several say it found papers they had missed with Google Scholar/arXiv and plan to keep using or subscribing.
  • Users from diverse fields (CS, medicine, marketing, animal advocacy, neuroscience, etc.) report relevant results, sometimes discovering genuinely new papers.
  • Some are “very impressed” by quality but emphasize it should complement, not replace, traditional systematic search methods.

Search quality and capabilities

  • Strengths:
    • Handles complex, detailed queries and refines them via follow-up questions.
    • Often surfaces niche and obscure but relevant work with fewer false positives than generic search.
    • Produces structured outputs: ranked lists, topic-match scores, brief notes, and an estimated coverage (“% discovered”).
  • Weaknesses:
    • Currently relies mainly on abstracts; missing full-text-only signals and some gray literature, theses, and key theoretical papers.
    • Ranking sometimes overweights topical similarity/recency and underweights perceived “importance” or seminal status.
    • False positives still present; some users report high precision, others note ~50% noise.

Architecture and design choices

  • Uses multi-stage, LLM-heavy retrieval with high-quality models, trading latency and compute cost for accuracy.
  • Citations are used to explore the graph but not primarily to rank final results, unless explicitly requested.
  • Core dataset is from a large academic aggregator; open-access full texts are planned, paywalled full text would require publisher deals.

Positioning vs. other tools

  • Compared to Elicit, Scite, Consensus, Semantic Scholar, Exa, etc., Undermind is framed as:
    • Slower but more accurate and suited for complex topic discovery.
    • Less focused on fast summarization and more on deep, agentic search.

Access, pricing, and UX feedback

  • Some frustration with institutional-email and signup requirements; a special HN link bypasses this.
  • Requests for:
    • Student or lower-cost tiers, pay-per-query, or rate-limited cheaper plans.
    • API access for integration into in-house tools and VS Code extensions.
    • Better follow-up questioning (e.g., multiple choice), “refine” as well as “extend” searches, importance-weighted ranking, richer citation formatting, and easy PDF/export options.
  • Concerns that long-term availability of results depends on startup survival; users want robust offline saving.