I asked GPT4 to refactor a simple, working python script for my smart lights… and it completely butchered the code and apologized mid-generation.

No amount of pleading or correction would get it to function as it did just a week or two ago.

It is so over.

  • ultranaut@lemmy.world
    link
    fedilink
    English
    arrow-up
    13
    ·
    1 year ago

    I noticed this today working on some bash scripts. Compared to a few weeks ago it’s become noticeably dumber, but also faster.

    • yads@lemmy.ca
      link
      fedilink
      English
      arrow-up
      5
      ·
      1 year ago

      There was a warning that they’ve temporarily disabled browse with bing

  • Aidan@lemm.ee
    link
    fedilink
    English
    arrow-up
    8
    ·
    edit-2
    1 year ago

    Gpt4 is not good at writing code. I think it’s because it has a lower token limit. Ask Gpt 4 to write out detailed specs for the code you want, then copy and paste that into a Gpt-3.5 session and ask it to write the code

    And if it gets cut off, paste in the last line it output successfully and ask it to continue with the line following that one. Then just copy and paste the blocks together

  • vegivamp@feddit.nl
    link
    fedilink
    English
    arrow-up
    10
    arrow-down
    3
    ·
    1 year ago

    Microsoft bought it. They’re not going to let their paying userbase of millions of coders evaporate…

    • KonaKoder@lemmy.world
      link
      fedilink
      English
      arrow-up
      5
      ·
      1 year ago

      Microsoft wants to own tools crucial to the mainstream of software development. They also want to own the cloud infrastructure on which those tools depend. Today, they might lose dimes on every LLM call. In five years, they’ll make a penny on orders of magnitude more calls. Microsoft has many flaws, including cloud capacity, but they aren’t short-sighted about investment. (I used to work in DevDiv and Azure Machine Learning.)

    • DarkenLM@kbin.social
      link
      fedilink
      arrow-up
      2
      arrow-down
      1
      ·
      1 year ago

      It’s Microsoft. Expecting them to make good and logical decisions is completely delusional.

      • Madrigal@lemmy.world
        link
        fedilink
        English
        arrow-up
        2
        arrow-down
        2
        ·
        1 year ago

        Good and logical decisions are plausible. However, expecting Microsoft to make consistent decisions and be able to work as a single cohesive team, now that’s delusional.

  • Totendax@feddit.de
    link
    fedilink
    English
    arrow-up
    7
    ·
    1 year ago

    I had their plus plan for a while but with the speed it gets worse I got rid of it.

    Nowadays latest model is without any doubt worse then what we had last year in December if you just regard answer quality.

  • Xandar437@feddit.nl
    link
    fedilink
    English
    arrow-up
    4
    ·
    1 year ago

    Noticed the same yesterday, seems like something is wrong. Gave it a simple row off numbers for 10 days. So for example day 5: 2, 4, 56, 8, 12, then asked it to give me for example day 7. And it keeps on mixing all the number for all the days. Then I correct the Ai, it apologies, gives the correct numbers for a prompt or two, then again mixes them all up…

  • aslaii@lemmy.world
    link
    fedilink
    English
    arrow-up
    5
    arrow-down
    1
    ·
    1 year ago

    Rewriting the whole code sometimes happen to me. But it just means gpt is already at token limit on one answer. You can copy the unfinished code then make it to continue.

    Try avoiding to make gpt generate a long code.

  • twelvewings@lemmy.worldOP
    link
    fedilink
    English
    arrow-up
    3
    arrow-down
    1
    ·
    edit-2
    1 year ago

    It’s worse that I thought. Both 3.5 and 4 butcher Python code by skipping words and inserting “```python” whenever you click the “Continue generating” button.

    Literally unusable now.

    Not to mention they even made the entire context window thinner and now there’s a horizontal scrollbar and code off screen.

  • GutterPunch@lemmy.world
    link
    fedilink
    English
    arrow-up
    2
    ·
    1 year ago

    It’s moderately good at in-line commenting functions and creating full function doc comments for the specific language / documentation format you need, but its code generation abilities are still not game-changing. Getting it to generate anything longer than a few helper functions is a test of patience.

    • twelvewings@lemmy.worldOP
      link
      fedilink
      English
      arrow-up
      9
      arrow-down
      2
      ·
      edit-2
      1 year ago

      This wasn’t always this case. I had zero Python experience a month ago, and managed to make a 300 line Python script that checks credit card validation, and has a beautiful UI. This would be impossible today.

        • twelvewings@lemmy.worldOP
          link
          fedilink
          English
          arrow-up
          4
          arrow-down
          1
          ·
          1 year ago

          Considering how long I’ve been using Python, and how it looked when I started, it is to me. And here is the ancient one I was previously using:

          • LoafyLemon@kbin.social
            link
            fedilink
            arrow-up
            2
            ·
            1 year ago

            Somehow this looks better to me, possibly because of less redundancy (copy buttons aren’t needed when you have ctrl+c).

  • Spzi@lemm.ee
    link
    fedilink
    English
    arrow-up
    1
    ·
    1 year ago

    Any clue what the ‘mistake’ was for which it apologized?

    • twelvewings@lemmy.worldOP
      link
      fedilink
      English
      arrow-up
      2
      arrow-down
      1
      ·
      1 year ago

      I mean, I could copy the code that works, but it’s not really the point.

      Not only did it apologize mid-codebox and not even comment it out, it started the code from the top again instead of actually fixing.

      Just so asinine and bad it’s not even worth analyzing further.

  • NotAPenguin@kbin.social
    link
    fedilink
    arrow-up
    8
    arrow-down
    8
    ·
    1 year ago

    People were saying the exact same thing a few weeks ago, and have been ever since it came out basically.

    You having issues with one prompt or one conversation doesn’t mean it’s dumb now.

    Still working fine for me.