> I see music as "the space of all possible 5-second clips at stereo 48kHz 24bit depth".
5.6MB? That's an astounding number of combinations. 1 followed by 1733933 zeroes.
> If you think about it, that space contains the intro to Stairway to Heaven and Oops I Did It Again, and the end of either song. It contains every 5-second segment of O Fortuna, plus a previously unimagined O Fortuna Remix with MF Doom rapping the pledge of alliegance backwards. My point is, AI is getting good at searching music space for novel patterns, and that's entirely the point of music, not making a career out of being an alcoholic minstrel with a tour bus.
AI is not searching that space for novel patterns for the most part, it is taking what it has heard before and coming up with things based on that. Which isn't a dig at AI, that's pretty much how humans do things too. I don't think today's AIs would be able to come up with something like Stairway to Heaven if Led Zeppelin and the music they inspired had never existed though.
Agree. Maybe novel is the wrong term given my framing. "Listenable?" The combinatorial space of music is hyperastronomical, effectively infinite. And most of it is probably noise.
AI isn't "searching" in the standard indexing sense. But if say, Suno, is doing stable diffusion on fourier transform heat maps, and there's a finite space of configurations... it is using a heuristic approach to pick an option from a well-defined (gargantuan) set of options.