• HiddenLayer555@lemmy.ml
    link
    fedilink
    English
    arrow-up
    19
    ·
    edit-2
    3 days ago

    “Write a program that does this.”

    “Fix this part”

    “And this part”

    “And this part”

    “Wait you fucked up the first part again, change it back to when you first fixed it.”

    “Ok now fix this last part.”

    “Damnit why do you keep changing the first part I already told you it was fine!”

    Chat length limit reached, please start a new chat.

    • ☆ Yσɠƚԋσʂ ☆@lemmy.mlOP
      link
      fedilink
      arrow-up
      5
      arrow-down
      1
      ·
      3 days ago

      My experience is that you want to keep changes as focused as possible. The smaller the context it has to work on the better the results tend to be, and the easier it is to check that it did what you actually wanted it to.

  • Anna@lemmy.ml
    link
    fedilink
    arrow-up
    8
    ·
    3 days ago

    Yeah AI is really stupid especially when it comes to bash. Once one of my colleague came to me asking a bash script not working that they got from ChatGPT. It was 20-30 lines with function and everything was a variable even the commands. I didn’t even bother reading, asked them what they were trying to do and wrote a simple one liner and it did exactly what they wanted in first try…

    • ☆ Yσɠƚԋσʂ ☆@lemmy.mlOP
      link
      fedilink
      arrow-up
      2
      arrow-down
      1
      ·
      3 days ago

      I’ve had pretty good success with DeepSeek making decent bash scripts. What was the script out of curiosity, I can try and see what DeepSeek spits out.

      • Anna@lemmy.ml
        link
        fedilink
        arrow-up
        2
        ·
        3 days ago

        Check if a CSV exists and if it does rename to the filename_date.CSV and upload to hdfs

        • ☆ Yσɠƚԋσʂ ☆@lemmy.mlOP
          link
          fedilink
          arrow-up
          3
          ·
          3 days ago

          here’s what DeepSeek gave me for the following prompt: write a small bash script to check if a CSV exists and if it does rename to the filename_date.CSV and upload to hdfs, keep it simple

          #!/bin/bash
          
          # Check if filename argument is provided
          if [ $# -eq 0 ]; then
              echo "Usage: $0 <csv_filename>"
              exit 1
          fi
          
          CSV_FILE="$1"
          
          # Check if file exists
          if [ ! -f "$CSV_FILE" ]; then
              echo "Error: File $CSV_FILE not found!"
              exit 1
          fi
          
          # Generate new filename with date (YYYYMMDD format)
          DATE=$(date +%Y%m%d)
          NEW_FILENAME="${CSV_FILE%.*}_${DATE}.csv"
          
          # Rename the file
          mv "$CSV_FILE" "$NEW_FILENAME"
          
          # Upload to HDFS
          hdfs dfs -put "$NEW_FILENAME" /path/to/hdfs/directory/
          
          echo "Successfully processed: $NEW_FILENAME"
          
          • Anna@lemmy.ml
            link
            fedilink
            arrow-up
            3
            ·
            2 days ago

            OK at least it didn’t create functions for everything. ChatGPT had a function to kinit, to rename and to upload to hdfs. And each function had 2-3 line comments. I mean for a single command how can you put 3 lines of comments.

      • WhatsTheHoldup@lemmy.ml
        link
        fedilink
        English
        arrow-up
        4
        ·
        3 days ago

        If you’re talking OpenEvolve you have to specific put in the tests and expected values for each function for this approach to work.

        • ☆ Yσɠƚԋσʂ ☆@lemmy.mlOP
          link
          fedilink
          arrow-up
          2
          ·
          edit-2
          3 days ago

          Not just OpenEvolve, I’m talking about how agents can use MCP to run code now. Cursor can now have agents iterate on a solution completely on their own.

          • jackr@lemmy.dbzer0.com
            link
            fedilink
            arrow-up
            2
            ·
            2 days ago

            Does this not collapse at some point? Like you see with running an image through image to image multiple times?

            • ☆ Yσɠƚԋσʂ ☆@lemmy.mlOP
              link
              fedilink
              arrow-up
              1
              ·
              2 days ago

              Genetic algorithms use a fitness function to converge on a solution, so it’s not just random iterations hoping one will work. It’s basically a gradient descent algorithm applied to output from LLMs.