Comment on AI agents wrong ~70% of time: Carnegie Mellon study

• jsomae@lemmy.ml ⁨1⁩ ⁨week⁩ ago

  Right, so this is really only useful in cases where either it’s vastly easier to verify an answer than posit one, or if a conventional program can verify the result of the AI’s output.

  source

  • MangoCats@feddit.it ⁨1⁩ ⁨week⁩ ago

    It’s usually vastly easier to verify an answer than posit one, if you have the patience to do so.

    I’m envisioning a world where multiple AI engines create and check each others’ work… the first thing they need to make work to support that scenario is probably fusion power.

    source

    • zbyte64@awful.systems ⁨1⁩ ⁨week⁩ ago

      It’s usually vastly easier to verify an answer than posit one, if you have the patience to do so.

      I usually write 3x the code to test the code itself. Verification is often harder than implementation.

      source

      • jsomae@lemmy.ml ⁨1⁩ ⁨week⁩ ago

        It really depends on the context. Sometimes there are domains which require solving problems in NP, but where it turns out that most of these problems are actually not hard to solve by hand with a bit of tinkering. SAT solvers might completely fail, but humans can do it. Often it turns out that this means there’s a better algorithm that can exploit commanalities in the data. But a brute force approach might just be to give it to an LLM and then verify its answer. Verifying NP problems is easy.

        source
      • MangoCats@feddit.it ⁨1⁩ ⁨week⁩ ago

        Yes, but the test code “writes itself” - the path is clear, you just have to fill in the blanks.

        Writing the proper product code in the first place, that’s the valuable challenge.

        source

        • -> View More Comments
• MangoCats@feddit.it ⁨1⁩ ⁨week⁩ ago

  I have been using AI to write (little, near trivial) programs. It’s blindingly obvious that it could be feeding this code to a compiler and catching its mistakes before giving them to me, but it doesn’t… yet.

  source

  • wise_pancake@lemmy.ca ⁨1⁩ ⁨week⁩ ago

    Agents do that loop pretty well now, and Claude now uses your IDE’s LSP to help it code and catch errors in flow. I think Windsurf or Cursor also do that also.

    The tooling has improved a ton in the last 3 months.

    source
• Outbound7404@lemmy.ml ⁨1⁩ ⁨week⁩ ago

  A human can review something close to correct a lot better than starting the task from zero.

  source

  • DreamlandLividity@lemmy.world ⁨1⁩ ⁨week⁩ ago

    It is a lot harder to notice incorrect information in review, than making sure it is correct when writing it.

    source

    • MangoCats@feddit.it ⁨1⁩ ⁨week⁩ ago

      harder to notice incorrect information in review, than making sure it is correct when writing it.

      That depends entirely on your writing method and attention span for review.

      Most people make stuff up off the cuff and skim anything longer than 75 words when reviewing, so the bar for AI improving over that is really low.

      source
    • loonsun@sh.itjust.works ⁨1⁩ ⁨week⁩ ago

      Depends on the context, there is a lot of work in the scientific methods community trying to use NLP to augment traditionally fully human processes such as thematic analysis and systematic literature reviews and you can have protocols for validation there without 100% human review

      source
  • MangoCats@feddit.it ⁨1⁩ ⁨week⁩ ago

    In University I knew a lot of students who knew all the things but “just don’t know where to start” - if I gave them a little direction about where to start, they could run it to the finish all on their own.

    source