Conversation

Notices

1. Embed this notice
   Tim Krief (timkrief@framapiaf.org)'s status on Thursday, 21-Aug-2025 17:55:03 JST Tim Krief

   "I asked chat GPT" is the new "I read in today's horoscope", but it kills more trees

   • Embed this notice
     djsumdog (djsumdog@djsumdog.com)'s status on Thursday, 21-Aug-2025 17:54:59 JST djsumdog

     in reply to

     • Julianoë

     There are some things that have gotten a lot better. Only a year ago most of the stuff these models generated was buggy, crap and badly formatted. I was recently trying to switch this media player frontend to use OpenGL rendering so it would work in Wayland. There's literally zero documentation of using python-mpv + Qt6 widgets. (There's one example using QQuick, but I'm not about to convert my entire application from QWidgets to QQuick/qml). Specifically the initialization, paintGL and process handlers it wrote here seemed like they were actually novel:

     https://gitlab.com/djsumdog/mediahug/-/blob/master/mediahug/gui/mpv/gl_player.py

     ..and mostly worked. (the other stuff I've written). Things that are grunt work, like a basic confirmation dialog, it can save a lot of time on. Claude Code wrote the basics of this, matched the style of the rest of my code, and mostly worked, although I still cleaned it up:

     https://gitlab.com/djsumdog/mpvbuddy/-/blob/master/mpvbuddy/gui/enqueue_files_dialog.py

     but they are literally weighted random word generators. I like the horoscope joke, because I know way too many people who swear by LLMs, even outside of coding. One developer I know seemed unbothered by lawyers using them when they generate non-existence cases. And when I really pressed him, he didn't understand the basics of transformers, embedding spaces, attention blocks .. a developer advocating for LLMs, even to the point of saying we're close to the movie Her, without understanding the fundamentals of how the machines work!

     The dangers are not in the random word generators, but those who trust them as their new oracles. The morons who ask "GPT/Grock is this true?" without doing the most basic lookups.

   • Embed this notice
     Julianoë (julianoe@mastodon.xyz)'s status on Thursday, 21-Aug-2025 17:55:01 JST Julianoë

     in reply to

     @timkrief I'm not a defender of generative AI usually. But pretending it is fully random and totally useless is missing the point and guarantees us to not see the why of its success and why people use it.

     For your PHP problem of earlier for example I'm pretty sure that with 3-5 lines of describing your contexte you would have a 95% accurate answer to your problem.
     https://mastodon.xyz/@timkrief@framapiaf.org/114441190475676584

     The question being: how many Google searches would be required to have the same answer?

   • Embed this notice
     djsumdog (djsumdog@djsumdog.com)'s status on Thursday, 21-Aug-2025 18:47:27 JST djsumdog

     in reply to

     • Julianoë

     It's not hallucinating or lying. I really hate it when articles keep using those terms. It can't do those things because it has no intent. There's no mens rea. In that sense, people need to understand it's a machine that randomly generates words, weighted to make them seem like coherent language.

     LLMs are just good at getting things that look right by accident.

   • Embed this notice
     Tim Krief (timkrief@framapiaf.org)'s status on Thursday, 21-Aug-2025 18:47:29 JST Tim Krief

     in reply to

     • djsumdog
     • Julianoë

     @djsumdog @Julianoe @ Grok is that true???

     Yes the issue is thinking it has any legitimacy.

     It's not hallucinating sometimes, it is always hallucinating, by design, sometimes the hallucinations make sense to you.