The combination it came up with that they then tried selling on a food stall was chocolate on pizza. It wasn't to everybody's taste but some liked it, with one describing it as "weirdly delicious".
It doesn't seem completely ridiculous that if an AI could be trained to recognise patterns of music that worked well together that it could analyse a large corpus of recorded music and come up with surprising mixes that you wouldn't think would work, but do.
Many restaurant chains sell chocolate on their pizza base as a dessert; chocolate is also known to complement many savoury dishes as a 'secret ingredient'.
Of course, there's always the possibility algorithmic composers could do even better than humans. Clarke told a story about that in 1957: https://en.wikipedia.org/wiki/The_Ultimate_Melody
Text here: https://archive.org/details/1957-02_IF/page/n71/mode/2up?vie...
So many of the selections that great DJs make are curveballs and surprising choices. This is less true in more samey genres, but true pioneering selectors will tend to surprise you with their taste and juxtapositions in a way that's hard or maybe even impossible to really capture in an algorithm.
Put another way, even in something as simple a song selection, the weirdness and contingencies in human creative decision making is a feature not a bug.
The efficiency and cleverness of algorithms can create create great powerful recommendation engines but there's nothing like the idiosyncrasies of a person as a curator.