AI Has a Discovery Problem

(mhacevedo.com)

37 points | by miguelacevedo 3 hours ago

8 comments

  • Vishal_Max 1 hour ago
    This is very true, but I don't think this is an AI-only problem. It is fundamentally a human problem.

    There are so many things happening around us, and one of the biggest things that separates people is knowledge. A non-technical person who puts in the effort to learn technology becomes a technical person. The only way to become good at something is to actually do it and practice.

    Someone who still doesn't know what AI is also doesn't know what is possible in the world today. My mom is a good example. She comes up with a lot of ideas and tells me about them on calls. Recently I told her, "Just tell ChatGPT. It can probably build a lot of what you need." But even that sentence isn't enough for her to understand what I mean, because she doesn't yet have a mental model of what AI can actually do. She simply isn't aware of its capabilities.

    I'm going home in a few days and will be staying with her for a couple of weeks. I'm actually going to try an experiment: teach her some of these tools and see what happens. Can someone with essentially no technical background learn to use AI effectively and actually build something meaningful within a few weeks?

    She already has a good job and doesn't really need to become technical, so there is no practical necessity for her to do this. But I think it would be a fascinating experiment.

    • da-x 1 hour ago
      I see this too. Generally, having a builder mindeset is independent from the tools available.
  • comrade1234 1 hour ago
    Does it really? Or is it that they're all pretty much good enough and so you just go with what's cheapest or what you're used to?

    For example, in IntelliJ I can switch between maybe a dozen models with a drop-down but the only time I do it is when whatever model I'm currently using has an outage.

    Another example, our product has some ai features (not coding related). I use spring-ai and so I can change models just in the config file. But I stick with deepseek because it's more than good enough and is super cheap.

    • kolinko 8 minutes ago
      It’s pretty well established by now that harness matters as much as a model. Even if a model is smart enough to work around harness’ limitations, it may take it way more time, context and money.
  • pseudosavant 16 minutes ago
    I was trying to teach a co-worker some agentic stuff today, and this is totally true. Especially if they are using gimpy tools like Github Copilot.

    The harness matters. None of my super-powered CLI + skills stuff worked like how it does in a more powerful harness like Codex or Claude Code.

    Try telling someone to imagine it could do something amazing, if they switch to a completely different tool.

  • try_the_bass 1 hour ago
    Subtle Lena[1] reference, maybe?

    [1] https://qntm.org/mmacevedo

  • qsera 28 minutes ago
    If I know the code behind the button, I don't have to press it to know what it will do.

    If I know how LLM work, and what it is trained on, I don't have to "explore" it to understand what it can do for me. The only variable is how deeply one knows how "LLMs" work, and that makes up all the on going arguments by lay people...

  • nalekberov 15 minutes ago
    The biggest problem with “AI” is that big corps are shoving it down our throats. AI has existed since nearly invention of computing, notorious example are video games, but no one marketed them with AI, why? because that was not the point, and besides, today’s AI is overloaded term, they are represented as if they are capable of starting new war, but in reality they are repeating the same stuff over and over again they “learned” from Reddit posts, books, codebases etc. in an indeterministic way.

    In most cases you don’t even need them. In some cases they give quite inaccurate answer with high confidence.

  • zerolunier 2 hours ago
    that's super true, it;s very hard for non-tech people to use ai at its peak power
    • specproc 1 hour ago
      That's our moat, right there. I'm starting doing clean-up jobs now, advising non-technical users on where they're going wrong. I had a guy recently who'd spent well over a month on Claude, running up the wrong tree on a research question.

      AI appears to be empowering non-technical users, but whether it's making them more productive or actually answering their questions is a completely different matter.

      • gloxkiqcza 28 minutes ago
        I'm also convinced software engineering tools like git and markdown are currently the best way to use AI for most things even unrelated to software development. E.g., I think using markdown notes + LaTeX + git for writing contracts is more efficient than using Word + Copilot, which are the tools I would expect a lawyer to use.
        • boorang 14 minutes ago
          totally agree with the repo idea. working with AI in a git repo and allowing it to write markdown files it slowly builds context over time. i knew a guy who had been using the same session for weeks because he felt he had invested too much time in building the context and was scared to close the session.
      • zerolunier 24 minutes ago
        it can't basically, because all the ai's are build around the tech community, doing complex coding automation etc. nobody thinking about how normal user gone use it
    • alserio 1 hour ago
      asking the AI helps
      • miguelacevedo 1 hour ago
        That helps, but people tend to ask about uses they can already imagine. The hard part is discovering what’s possible beyond that.
  • markking 44 minutes ago
    [flagged]