[crazygringo]

Yes -- I think Knuth's work is a great first draft of this.

In my mind, however, the concept is less about providing parameter values (like Computer Modern), but more about actually drawing the parts. In other words it's less about variable/multiple masters, but "here's what the end of a vertical serif looks like at this weight", "here's what a full vertical stem looks like". (Though if you're extrapolating from Latin to Arabic where virtually nothing is in common, then it might be closer to a multiple masters situation.)

Maybe it's too hard to do, I can't be sure. But I've also seen deep neural-network-generated fonts posted here that have intrigued me... and even if it's too hard to simply "assemble" fonts from pieces and get a good result, I can't help but wonder if deep learning applied to 10,000's of typefaces couldn't be a tool in making that happen effectively.