The Possibility, Feasibility, and Utility of Music AI

Two challenges made me abandon music AI for the past few weeks.

The first challenge was that, depending on the metaphysical nature of things, music AI might not be possible (specifically if reductionism is false). And, depending on how complicated a program needs to be to make good music, it might be possible but not feasible.

The second challenge was that music AI might be possible and feasible, but not worth the opportunity costs. Specifically, I think that what I should do is seek enlightenment, and writing music AI is not seeking enlightenment.

As far as the first challenge goes, I’ve come to the conclusion that continuing to doubt my philosophical views is only hurting me, and not helping anybody. So for now, I believe that reductionism and materialism are false.

Here is not the place to state what I believe instead; and I can’t do so adequately, because I haven’t finished clarifying my thoughts on the topic. But I do need to say what is true about the possibility of music AI under my metaphysics.

Under my metaphysics, “beauty” is not reducible to any finite mathematical formula. So writing a generalized beauty-creating computer progam is not possible.

But it is clear and obvious that some aspects of beauty can be mathematically quantified. The Baroque and Classical era defined elaborate rules which music should follow; and these rules partially select a set of beautiful pieces of music from the set of possible pieces of music.

What about the qualitative aspect of beauty? What about that feeling of beauty which feels purely non-logical?

That feeling is an illusory perspective on an objective reality. Perhaps, at least in many cases, there is an illusory logical perspective on the same objective reality. This would be the same position as reductionism, except that it does not privilege the logical perspective as being somehow more real than the qualitative perspective.

My own hypothesis is that we can define music-making rules which pick out particular subsets of the space of beautiful music. We can define rules which produce some beautiful music, but not all beautiful music. And we can continue making better and better rule sets infinitely, without ever arriving at a final rule set.

This is similar to what is true about math under Godel’s incompleteness theorem. Under Godel’s incompleteness theorem, every axiom set yields some truths, but not all truths. And we can continue constructing better and better axiom sets infinitely, without ever arriving at a final axiom set.

So if we are writing a music AI, the task we are considering is the task of defining and implementing a rule set which picks out some nice subset of the space of beautiful music.

I feel pretty confident that this is possible and feasible. The strongest evidence that it is possible and feasible is the existence of EMI. My project differs from EMI mainly in the details of the composition algorithm, and in the fact that EMI works from an existing corpus written by humans, whereas I intend to write an AI that does completely “de novo” composition.

So that leaves us with the second problem: should I do this? I think that what I should do is seek enlightenment, and if this project doesn’t contribute to that, then I shouldn’t do this project.

What I’m realizing, however, is that seeking enlightenment needs to involve having experiences in the everyday world and distilling new insights from those experiences. The everyday world is this huge, apparently chaotic flux of experience which helps us to discover things we didn’t know about. Assembling a huge quantity of those things eventually amounts to new fundamental understandings about the nature of enlightenment and the most efficient ways of seeking it.

So I’ve come to conclude that I should engage in and care about things that, in a superficial sense, have nothing to do with enlightenment. This means that working on music AI is not incompatible with the idea that what I should do is seek enlightenment.

So it seems that music AI is possible, feasible, and something I potentially should do. There are no fundamental philosophical obstacles to my working on it.