Show HN: Learn LLMs LeetCode Style

(github.com)

172 points | by Exorust 1 day ago

13 comments

  • janalsncm 1 day ago
    This is decent for what it is. Some of the problems are pretty open ended which has pros and cons, but that is very different from leetcode, which has very specific data and test cases.

    For example, implement linear regression but the example solution uses a random number generator without a fixed seed. It’s fine, reproducibility isn’t the point, but leetcode problems are more structured.

    In leetcode they usually don’t tell you exactly what data structure you must use, only that it must pass certain test cases. By analogy this might not tell you which architecture to use but require that it passes certain eval metrics.

    • Exorust 19 hours ago
      I hoped that it would be a little open ended as most questions in ML in real life are open ended.
      • janalsncm 3 hours ago
        Most ML problems in real life don’t constrain you to use linear regression or a CNN either. But there will be some metric you need to optimize.

        What would take this repo to the next level is to have a reproducible data generation function for each exercise as well as a reasonable metric which must be passed. I don’t see anything that requires my classification auc to be over 0.5 which would be a basic criteria of bug-free code.

        • Exorust 1 hour ago
          It's also what most people ask when they go for interviews.

          I was reverse engineering the ML interview pipeline for myself and that's how I stumbled upon all this.

          I think the data aspect does make sense tho. I might add that as the next thing to do

  • gerroo 1 day ago
    Cool idea, will try. Since it seems mostly llm generated you could publish the process and prompts for transparency.
    • Exorust 1 day ago
      I'll do that. I'll also add a disclosure that I did use Gpt to generate it.
  • only-one1701 1 day ago
    > Avoid using GPT. Try to solve these problems on your own. The goal is to learn and understand PyTorch concepts deeply.

    I mean...this entire project appears to be mostly GPT-generated?

    • mumbisChungo 1 day ago
      One time my teacher used a computer to make a math test for me, but then told me I couldn't use my computer during the exam. I dropped out of school immediately.
      • only-one1701 1 day ago
        Great analogy my brother there’s minimal difference between a word processing software and an LLM
    • YeBanKo 1 day ago
      Why do you think it I GPT generated?
  • pj_mukh 1 day ago
    What are people's other "go try to build this thing, perfectly aligned to your noob-level" ways of learning lower-level ML Tools (PyTorch, CUDA etc.)?
  • cwlcwlcwlingg 9 hours ago
    It helps me a lot!!
  • oezi 1 day ago
    Is it just me or does anyone else find the red squiggly lines under Pytorch and Leet hilarious in the heading picture?
  • arnab_optimatik 1 day ago
    super helpful, thanks for sharing!
  • lihaciudaniel 1 day ago
    [dead]
  • saltserv 1 day ago
    [dead]
  • thehamkercat 1 day ago
    [flagged]
  • NetRunnerSu 1 day ago
    [flagged]
    • raincole 1 day ago
      ^ This is a bot/crackpot account whose only purpose is promoting jumbled words on HN. How/where to report it?

      [0]: https://news.ycombinator.com/submitted?id=NetRunnerSu

      • dotancohen 1 day ago
        Reminds me of some of Xah Lee's earliest posts on the Python mailing lists. Many, myself included, thought he was a clever bot.
      • anonymoushn 1 day ago
        Accounts with high karma have a "flag" link in the header of comments and submissions for indicating that it may not be the right kind of content for the site. Human moderators actively monitor the site, so your comment about this user's behavior will be seen by one of them soon.
      • satvikpendem 1 day ago
        Email hn@ycombinator.com