Command lint - Terminal AI

Started by sleestack, October 20, 2024, 11:51:25 PM

Previous topic - Next topic

sleestack


Hi, I've been using supercollider with Claude for a little while now and it's very good at programming super collider including gui apps if it does something wrong you just tell it the error and it fixes it and I'm talking about programming , .. terminal commands would be nothing to this thing so I think it would be really interesting if CDP Could be put into one of these LLM's and the write terminal commands for cdp to say -  I want to do this process in this way then into this process in this way and then it types it all out for you as a file and bam. ChatGPT is very good at at terminal so I wanted to say maybe that could be an option I know there's maybe moral issues some people have with some of these LLM's,  but they can provide access to people without having to spend a year reading you know and reading and reading, also. Are the workshops for working with Terminal on a Mac somewhere? I had them but wasn't sure if they've been updated or not. Thank you

sleestack


simonk

Hello,
You can find the Tutorials here :
https://composersdesktop.com/workshops.html#LM

They haven't been updated for a while, you might find the 'Learning Manual'
of more interest

-further down the same page

sleestack

I get it, hard work is the only way. But I don't think there is anything wrong with a different approach. Sometimes you get manual fatigue. It's not the only thing Im learning. Im reading about supercollider, about production, about sound design. Some people also suffer with adhd, and it takes forever to get anywhere, but are still very creative people. Its like you didn't read my post

simonk

Quote from: sleestack on October 21, 2024, 10:08:37 PMIts like you didn't read my post
I am sorry if my reply seems that way to you. I was offering information about the whereabouts of the Tutorials etc.
Also, I have no particular knowledge of AI & plenty of learning of my own to do.
The best advvice I can offer on soaking up manuals is to 'snack not banquet' when taking in information. Do a bit,do something else, do bit & so on.

As previously mentioned by one or more of the very small CDP team we ~are~ working on new GUI tools.

Simon

sleestack

go to changpt if you feel to. Ask it to do something for terminal. Anything. It is incredibly good at Terminal. All CDP can be done in terminal. Instead of a gui. Teach it CDP and let us talk to it . Ask it to do this and that, It will spit out all the information you need to make a sound file.

rwdobson

The challenge in doing such a thing lies in the requirement to "Teach it CDP". LLMs are trained over literally billions of image and text examples trawled from the internet. I doubt if there is anything like enough examples of a working CDP command line on the net to do the job. It could be something a Computer Science department at a university might fancy as a research project. Part of the problem lies in the critical evaluation stage. Given a "correct" command line, one person listens to the output and says "that's it", while another might say "no that's not it at all".

As a side example of the challenge: I am a flute player, and I see from time to time examples of AI-generated images showing a person playing a flute. It is all in photo-realistic full-colour 3D. But in literally all the cases I have seen, none of the flutes matches what a flute actually looks like, and could not possibly work. Some don't even have a mouthpiece. Others assume the player has 20 fingers. But they look, well, plausible as images. Training an AI to "do CDP" may be an even harder task, and generating a sufficient data set for it could, alone, takes thousands or even milions of person-hours.

That said, a few people well versed in Csound have got ChatGPT to generate correct working Csound instruments (albeit simple ones) -  because there is a lot of Csound text available online for ChatGPT to work with, not least in large textbooks such as "The Csound Book". I would say that across the whole extent of the CDP online documentation and tutorials, there is probably less than 0.0001% of the examples an AI would need!

The approach I would suggest might be the most likely to succeed would be to create a script for a single CDP process, that can algorithmically generate variations on a working command line (start simply - no external breakpoint files!). Unix shell scripts have all the language elements one would need for this. Run each one, and score them based on a chosen evaluation criterion (or multiple criteria?). Perhaps if this data set is passed to an AI, it can generate new good command lines for that process based on input criteria. Sounds like an idea for a PhD to me!


simonk

There is also an AI elephant in this imaginary room and somewhat philosophically challenging.

Is it acceptable for AI bots to run round scraping other artist work without payment ?

https://www.theguardian.com/film/2024/oct/22/thom-yorke-and-julianne-moore-join-thousands-of-creatives-in-ai-warning