CHATGPT on CAKEWALK

[msmcleod]

10 hours ago, Teegarden said:

@msmcleod, interesting video, he explains it very well! I did notice they are from three years ago, so by now are antique?(just kidding...)

Did ChatGPT not already do that collection of enough quality data for you? I thought you could make it to good use because it has already all that training you need for at least some useful implementations? 

ChatGPT is being used to quickly write a procedure or app while saving programmers a significant amount of time. They still need to check and finetune it, but the various examples I seen over the last few months are very encouraging and -of crucial importance if you want good results- depending on how you adjust your questions, the program quickly provides increasingly better answers). So, I could imagine that you might benefit from it with faster programming. 
Version 4 of ChatGPT is only just out and a vast improvement over the former version, acing tests where the former still miserably failed.

 

First of all, the neural network technology currently in use was developed around 50 years ago.  I remember studying it at university (and that was a long time ago too!). It's only now however, that computers are fast enough to leverage it effectively (apart from maybe applications such as AutoRoute which have been using it for decades).  But the requirement for speed is especially the case in deep learning, where the AI solution is left to work it out for itself rather than giving it pre-labelled data.

ChatGPT is designed as a chat application. It is not a catch-all AI app that does everything you need.  Fine if you're feeling lonely in the middle of a mixing session and fancy a chat, but not so useful for musical applications. All it will ever give you is an amalgamation of things it's scraped from the internet (and we all know how contradictory that can be), and the things it's learned from experts correcting the answers it gives. 

The results from ChatGPT may look impressive, but are suspect when looking at the detail. I had a lengthy "chat" with ChatGPT on a chord transformation algorithm, and the answers were best described as naive.  There were way too many assumptions regarding implementation details (e.g. how to distinguish between the fundamental pitch of a quieter sound and the harmonics of a louder sound), and when pushed further, I basically got "There's been lots of research on this but I don't know the answer".  I ended up suggesting a possible algorithm, and even then the results were flawed.

In general, there are three types of AI applications for audio:
1. Classification - this is by far the most prevalent, and includes genre classification (as used by Spotify), or chord/key recognition.
2. Transformation - where you give it audio, and it does something to it rather like a VST effect. Izotope Nectar 3 or their mastering suite is a good example here, as well as the stem separation utilities.
3. Generation - where you give it something basic, be it a melody or just some descriptions, and it generates music (either MIDI or audio) based on the input.

None of these applications are suitable for ChatGPT, and all of them need a vast amount of training data to be effective.

[Starship Krupa]

1 hour ago, Byron Dickens said:

More and more steps towards "push button, out comes song." No skill or work required.

So what? Who's going to consume that theoretical song?

So someone can press a button and create a song. Who's going to pay to listen to it or own their own copy? If the answer is "people who would otherwise have paid for my music," well, pardon me, but if I'm putting out soulless crap that can be bested by a robot, I deserve to be put out of business.

It's no skin off my nose if someone can push a button and make a song. Doesn't affect me in the slightest.

I've never cared 2 plugged poops for what mainstream music consumers want. It's been almost 50 years since I gave a rat's hiney about the Billboard Hot 100. My thing in the 80's and 90's was indie labels and college radio. My deal in this century is SomaFM and Bandcamp. I have fringe-y tastes and I like it like that. So slick saccharine overproduced crap-pop can now be created entirely without human involvement? The music in the top 10 already sounds like it was created by algorithms. Why not take it all the way?

Elvis. The Beatles. Disco. Punk. Rap. Synthpop. Rave. All spelled "the end of true musicianship." All were bellwethers of an imminent musical apocalypse.

Doomsayers will be doomsayers.

[Teegarden]

@msmcleod thanks for the detailed feedback!

39 minutes ago, msmcleod said:

ChatGPT is designed as a chat application. It is not a catch-all AI app that does everything you need.  Fine if you're feeling lonely in the middle of a mixing session and fancy a chat, but not so useful for musical applications. All it will ever give you is an amalgamation of things it's scraped from the internet (and we all know how contradictory that can be), and the things it's learned from experts correcting the answers it gives. 

I'm aware of this. 

Apart from writing things like complicated chord algorithms, I thought it might be possible is to use AI tools, not just ChatGPT but maybe better tools like Copilot or Sonic Pi to write and check code, speeding up the development for the more basic parts of CbB.

What is Github Copilot?

Here's an article written by an professional AI programmer that used Sonic Pi  to program quite complex structures to produce algorithmic music and many other things instead of Sonic Pi's initial purpose which is live coding: 
Sonic Pi - programming artificial intelligence based daw

[Milton Sica]

29 minutes ago, Teegarden said:

@msmcleod thanks for the detailed feedback!

I'm aware of this. 

Apart from writing things like complicated chord algorithms, I thought it might be possible is to use AI tools, not just ChatGPT but maybe better tools like Copilot or Sonic Pi to write and check code, speeding up the development for the more basic parts of CbB.

What is Github Copilot?

Here's an article written by an professional AI programmer that used Sonic Pi  to program quite complex structures to produce algorithmic music and many other things instead of Sonic Pi's initial purpose which is live coding: 
Sonic Pi - programming artificial intelligence based daw

Great, great !

 

[Bruno de Souza Lino]

The problem is that you'd have to bundle in already pre trained datasets which will still be several Gigabytes in size and you'll also push your system requirements to require a proprietary framework (CUDA) just to run the thing. And that assuming you even have enough computing power to run a local instance of any AI algo out there. The most optimized version of Stable Diffusion requires at least 16 GB of RAM and 10 GB of VRAM for example.

[User 905133]

Is it Real or is it AI?

 

[Bruno de Souza Lino]

2 hours ago, User 905133 said:

Is it Real or is it AI?

 

"Using artificial intelligence A.I. to identify key elements of a spirited outtake from the 'Get Back' sessions, we've attempted to create a new arrangement of this forgotten Beatles moment that we hope fans will enjoy."

So, this video is 1% AI, 112% human work. If anything, all the stuff done in the video could be accomplished without AI and the result would be more or less the same. Also, the lovely videos about AI always focus on that 0.1% where AI succeeds at a task instead of the other 99.9% where it was wrong.

[User 905133]

14 hours ago, Bruno de Souza Lino said:

"Using artificial intelligence A.I. to identify key elements of a spirited outtake from the 'Get Back' sessions, we've attempted to create a new arrangement of this forgotten Beatles moment that we hope fans will enjoy."

So, this video is 1% AI, 112% human work. If anything, all the stuff done in the video could be accomplished without AI and the result would be more or less the same. Also, the lovely videos about AI always focus on that 0.1% where AI succeeds at a task instead of the other 99.9% where it was wrong.

You busted my bubble. I thought it was pretty good and was starting to see a good side to using AI in the music creation process.  My take-away: OK, I'm 1% pro-AI in music creation (as long as human artists do the other 99%).  ?

This might be something I can get into, though:

Maybe it IS time for CbB to take the lead and develop TalentGPT for Cakewalk?

Or, how about a sampler like the masses have been asking for.  Certainly massive collections of samples can be used as a database to train AI.  Then people can type in requests for sounds / harmonic structure, sample manipulation techniques, frequency / pitch, BPM,  duration, and other musical parameters (including line, progression, key, harmony, etc.) and the AI can just make it happen.

Maybe CbB users could even seed the request with 10 to 100 original clips made 100% within Cakewalk?  How many Cakewalk users have all sorts of snippets of unfinished projects and not enough time to work them into finished compositions?  Hmmmmmm.  Food for thought.  

Edited April 4, 2023 by User 905133
word changes in the interest of accuracy

[John Nelson]

On 3/20/2023 at 3:09 PM, Milton Sica said:

1) Evaluate my mix and propose a better one considering my settings.
2)

Define "better".....

[Milton Sica]

4 hours ago, John Nelson said:

Define "better".....

Something like the mastering plugins that use references like TR5 Master Match, that is, capturing and treating my knowledge base and proposing improvements based on them.

When I created the topic, I focused on using AI for better DAW operation, such as increasing productivity beyond hotkeys, improving CAL codes, improving theme editing, etc.

Edited April 5, 2023 by Milton Sica

[Bruno de Souza Lino]

On 3/20/2023 at 2:00 PM, Byron Dickens said:

More and more steps towards "push button, out comes song." No skill or work required.

Bandlab already does that. It's called SongStarter.

https://www.bandlab.com/songstarter