• merc@sh.itjust.works
    link
    fedilink
    English
    arrow-up
    2
    ·
    6 hours ago

    That’s what makes them shitty though.

    When I have a hard technical problem I often search for and read through a dozen different sources. Many of them are wrong, or are right but not covering exactly the situation I’m looking at. Eventually I’ll find one that’s either right and answers my problem, or gives me the clue I need so I can figure out the solution for myself.

    If I ask an LLM to solve the problem, it will make up an answer that would seamlessly blend in with all its training data. In other words, it’s most likely to produce something that’s wrong, or something that’s right but not for my particular case, or something that’s close but incomplete. That’s effectively useless. At worst it blends in with its training data enough to convince me it’s right, while not actually being right. At best it’s something that is close enough to give me the clue I need. Most of the time it’s going to be something that’s wrong and I know it’s wrong because if it were that simple I wouldn’t have had to resort to the AI bullshit generator.