Ask HN: How soon can we have LLMs that read textbooks and answer queries?

1 point by baapercollege 1 year ago | 5 comments
Any specialist needs to read many high-level textbooks. How far are we from having a chatbot where we can simply drop the 1000 page textbook pdf and ask questions to the bot and get answers from the content preferably with references to appropriate page number? Can this already be done?
  • ChildOfChaos 1 year ago
    I use Gemini pro 1.5 for this as it has a large context window.

    https://aistudio.google.com/app/

    Feed it the pdf and prompt it that you only want answers from the supplied text with references to appropriate page numbers and quotes from document.

    Use a free USA vpn chrome extension if not available in your country.

  • user90131313 1 year ago
    When they fix the hallucination problem.
    • warrenm 1 year ago
      This has been possible (at a mildly simplified level) in Acrobat for decades
      • baapercollege 1 year ago
        :) i'll tell you the difference between search and language processing. if you have 10 chapters each highlighting 10 unique points, search works great as the book is basically a sub book of 100 topics.

        but for books like medical texts, all chapters have similar sub-headings, so the book is like a table with medical conditions as rows and approaches like pathology, physiology and pharmacology as columns in chapters. this is why simple search throws up massive false positives and is mostly useless.

        the complexity in medical field stems not from existence of vast array of unique features but due to having a fixed vocabulary of features in different permutations for distinct conditions. so word level searches are far less efficient than paragraph level searches where the meaning is inferenced by considering every word in its vicinity