“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.
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.
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…
Happy cake day! (If we do this on Lemmy)
This is one thing from Reddit that I don’t miss. To everyone else this is just spam.
Yeah, maybe you’re right
Hey we need a little bit of spam on the internet to filter out bots who write elaborate garbage from GPT and humans who write simple msgs.
The profile just says 2 years ago. How do you know this is the cake day? I don’t even know when I joined so maybe it is who knows.
Saw on Voyager (my client)
It also says cake day underneath the 2 years ago, or at least I see it on the web client.
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.
Check if a CSV exists and if it does rename to the filename_date.CSV and upload to hdfs
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"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.
I’ve never touched ChatGPT out of principle, but everything I hear about it makes it sound like hot garbage.
On the ai line, the purple cars are the part a human had to clean up and correct
Nowadays, the agents can iterate. It basically works like a genetic algorithm where it converges on a solution.
If you’re talking OpenEvolve you have to specific put in the tests and expected values for each function for this approach to work.
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.
Does this not collapse at some point? Like you see with running an image through image to image multiple times?
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.
