if i had access to a decent image creation llm i’d prompt a mash up of boobs bums bacon sandwiches and machinedrums to convey how i’d be terrified and ashamed to see accurate results direct from my mind…
“please filter out all the animalistic instincts and obsessive intrusions” will be a required checkbox
Edit: actually that maybe exactly what I’m striving for in sounds i create….
I gave it a shot. Using the musicgen large model I generated a bunch of desert trance clips/songs. The results are very interesting. The words you use in the prompt can produce wildly different results.
I went for 3.5 minutes of output in each clip, and perhaps that was a bit much. I’m interested to see how Octatrack slices these up, and perhaps I should have made them shorter. Using the large model spiked around 9GB GPU memory, which really is not much compared to big LLMs or image/video generation.
I see a big future for this type of thing and I’m surprised Ableton is not leading the way. Look at adobe with Photoshop. You take a picture of a lake and then draw a circle in the middle and say “add some ducks” and it gives you a bunch of realistic options to paint in the ducks. This should be seamless in Ableton where you highlight an area and say “Give me a syncopated drum fill here” and it inserts a context aware clip… I.e., not some random drum sample but one tailored to what’s going on in the timeline. Maybe eventually.
AI music generation is improving all the time. There’s a service called suno.ai that can generate surprisingly decent songs, even with vocals that don’t suck if you know how to prompt and “lead” it from part to part. But there are still many of the audible artifacts, glitches in places, etc. But better sounding than most I’ve heard.
Right now, AI is most useful to me in terms of vocals. For instance, in my latest track, I used AI to generate a vocal track from my lyrics, and did several dice rolls on each part of the vocal until the melodies and timings were close to where I wanted.
Then I brought the parts into RipX where I could clean up and fix melodies and phrasings that weren’t right.
I used that “fixed” vocal in my Ableton project with all of the instrumentation I created. Once I had sliced and diced the vocals even more and added some effects, I had a pretty good track.
But the voice track was still of lower quality than I wanted, so with the sliced and diced vocal stem from Ableton, I used ANOTHER AI service (audimee.com) that can transform a vocal (real or AI) into any of their dozens of virtual singers. I found one of their virtual singers that sounded “right” for the track and had the AI effectively duplicate the uploaded original AI track, but with better quality vocals.
Then I layered the source AI vocal and new AI vocal together and with a few more fixes (weird consonant mismatches, etc), and it sounded great.
I didn’t have to sing a single thing (been dealing with a flu for the last two weeks) and got a vocal that fits my track, plus the services and apps I use are copyright free, trained on open sources or using training data from singers, etc. who were properly licensed and paid for their work.
Just be emotionally prepared for the day coming soon when you can prompt the AI with, “make me a finished club banger in the style of _________ meets __________”, and within minutes you can download a polished track ready to release. 
Just imagine YouTube ads delivered directly into your brain.
Unskipable publicity. Unavoidable consumer tracking.
Horrible. AI loops is the less harmful thing.
the only exciting thing to me is SPEED. adverts will be there, but we’ll normalise ignoring and swipe aside as we do now.
but if i can make music how i perceive it by looking at things and touching my fingers then that could be really fun. same with coding. really interested if it would be quicker than a keyboard. still not clear to me at this stage.
minority report style interfaces will be here within a year…
100% agree
Not only that, but we also tend to ascribe too much value to instrumental reasoning vs substantive/value reasoning. It’s how you get incredibly well educated STEM people with degrees from MIT or whatnot who think Jordan Peterson is the best philosopher of our age.
I’m loving Lee Gambles latest rec, ‘models’. It’s full of disembodied ai voices.
That should, at the very least, be on a t-shirt.
That fourth track is awesome.
Did you hear about that : https://www.suno.ai/ ?
i hate to admit it but (as long as ai music references shit music) it’s shit
having said that I clicked create without changing the prompt, and this is fresh:
I listenned some tracks and it’s actually not so bad. 2 years ago AI didn’t exist, it grow so fast… What will be in 5 years ?
I think there will be tons of radio music unlimited generation of any kind.
I see a video of a guy using chat GPT to create lyrics in the style of a popular french rapper, uploading some sample lyrics to have identical style. Then he created the track with Sunno, then he replaced the voice with a model of the french rapper. The result was poor quality but seem pretty accurate.
I can’t stop thinking in few years when the quality will be top, how the music industry will react ?
They make android superstars.
Whatever music AI comes up with, it’s sure to be more emotive than Taylor Swift.
Here is a thread about Suno