> people said "it's impossible to separate tracks (voice, bass, ...) after they are mixed". true in theory,

and true in practice too. what you get out of an ai demuxer isn't an exact match for what went into the mix. it's a plausible approximation, much like ai upscalers, upsamplers, colorizers, and such.

> but neural-networks can separate them in practice

That's a massive stretch.

They are able separate them sort of, but they are nowhere close the original quality of the individual tracks.