This is the fediverse, so this is probably going to be extremely unpopular, but yea here goes.

Not saying you have to love it, or trust it blindly. Skepticism is healthy. But there’s a difference between “I use it carefully and check everything” and “I refuse to touch it.”

Historically, refusing new tools has never stopped them. Coopers didn’t stop the steel drum by refsuing to make them. They just ended up out of work while the industry moved on.

The realistic threat isn’t AI replacing you. It’s a colleague who uses AI well doing your job faster. And the skill of using it properly (knowing when it’s wrong, veriyfing output, knowing what not to use it for) takes time to build. Starting late means competing with people who’ve had years of practice.

You can have legit concenrs about ethics, energy, or jobs and still learn the tool. Opting out doesn’t fix those problems. It just removes you from the people who get a say in how it’s used.

PS. Ethically trained models exist and you can run models locally.

  • jj4211
    link
    fedilink
    English
    arrow-up
    1
    arrow-down
    1
    ·
    1 day ago

    Finding issues, sure if you don’t mind wading through the false positives. Remediating is a bit touchier, since it will confidently declare a common pattern to address a potential issue without considering the implications. For example, recently AI suggested adding a very slow call multiple times in a call stack, that turned an instant operation into one that took 90 seconds, but without any real improvement. In another it suggested a very very slow way to get information that was available in a very quick function call. Then In another, it fabricated test reports asserting that the developer had validated the functionality in a real world test, but the change utterly broke the only thing that could have tasted the change.

    It really depends on the scenario in play, the potential stakes and how well instrumented “correct” is. If correctness is well instrumented, then you can instrument retries. If circumstances make that unrealistic, then it can accelerate manifesting code, but going full agentic becomes a nightmare of never stable slop.

    • Ek-Hou-Van-Braai@piefed.socialOP
      link
      fedilink
      English
      arrow-up
      1
      arrow-down
      1
      ·
      15 hours ago

      Spoken like someone that hasn’t used a frontier model in the last 6 months.

      They have gotten so much more powerful, they rarely make mistakes anymore on certain tasks.

      There’s no real downside in asking it to scan your open source project for security issues.

      Or to ask it for test cases you missed, so you can add more unit tests.

      • jj4211
        link
        fedilink
        English
        arrow-up
        1
        ·
        12 hours ago

        I am exposed to frontier models, my company is so enthusiastic about AI and so of course they are providing access. I am so tired of people denying my experience in this.

        They make mistakes all the time, but in many software tasks you can provide a framework for mistakes to not matter and get detected.

        There’s not much of a downside of scanning a codebase, but they absolutely raise false positives.

        Yes it can find issues you missed, and you can ignore when it is wrong, that’s what makes it particularly good at code review, the mistakes don’t matter. It’s also effective at compromising security, because the attacker does not care about the mistakes. Letting it generate code in a scale where the human can review and intervene can be useful, particularly around some use cases that are very obvious but very tedious, as long as the human is extra vigilant.

        It continues to be a risk to go full agentic and assume that the AI will sort everything out without review in scenarios where you can’t afford mistakes.