Eno(Eno(Eno()))


 

Brian Eno is the undisputed king of musical synchronicity.  There are patterns in the void, and at his best, he is drawing them out and shaping them with a recording studio into something the rest of us can observe and maybe, if we're lucky, comprehend.  

I saw Gary Hustwit's new generative biopic about Eno last night (showing 10.2.24, Roxie Theatre SF) .  I feel the Internet's hackles rising across time and space as I type, so I need to immediately clarify that "generative" here refers to a statistical process by means of which an approximately 90 minute subset of about 30 hours of interviews with Eno and 500 hours of movie scenes are algorithmically selected for a given showing.  There is no GenAI scene generation in this movie.  With that out of the way, the particular movie we rolled last night made some fascinating connections between topics I hadn't thought enough about, including how Eno's process is in many ways the polar opposite of generative AI.  

The first big takeaway, that is very applicable to the current zeitgeist, is Eno's perspective on generative art.  He admits that he has no skill as a musician, but what he does have is a good ear.  He has judgement.  (I present a few samples as evidence).  There is an argument that those who build with generative AI are playing a corpus and the humanistic contribution they are making is in the choice of the output.  This argument relies very much on the scale of the choice.  Restricting spaces of choices help.  Western music has 12 notes, and thus there is an implicit restriction in claiming a scale to what the next note in a melody can be.  Taking an algorithm to further restrict to a "random" 8 notes instead seems entirely reasonable.  Now imagine instead that you take a motif, say 4 notes.  You restrict yourself to 10 generated sets of 4 note motifs.  This is a greater restriction, but there is still an incredible amount of space for shaping and curation: particularly if you allow yourself to select a new set if none of the sets presented satisfy.  A related quote from the movie was Eno pointing out that he doesn't play an instrument, or a panel, but he does play a recording studio.  He has a fundamental grasp of scale. 

This should generalize, but I think part of the reason it doesn't, and why there is so much uneasiness about the product of modern generative AI, is the unforeseen complexity that emerges from small scale changes in the initialization.  Another theme of the movie, and Eno's work, is the effect of judgement on emergent complexity.  In the film, Eno discusses Conway's Game of Life as the classic example of this curation of complexity.  The Game of Life is a cellular automata where cells are initialized over a fixed size grid, and a set of simple rules define the state changes at every timestep.  This can lead to a variety of fascinating transitions, and is hours of fun to play with (the code is simple to script, and you can also play it on the web here). 

While the Game of Life is very illustrative of the importance of small initial variations on long term growth, the interplay between emergence and judgement and scale is most visible in a real life example in the Matthew Effect: loosely "the rich get richer."  AI generated content is awful for the same reason that some people fail up: the judgement system relies on aggregate measures, calculated at the wrong scale, instead of any sort of smaller, aesthetic or local-environment based metric.  The generative algorithm has all of the technique to create photorealistic output, or output in any style.  What a generative AI lacks is the discernment to prefer one output over another.  What it judges best (the closest it gets to "wants") instead is to produce output that will fall easily within the space of existing boundaries.  Similarly, the lazy, but easiest, way to select a candidate for a scholarship, promotion, or oppurtunity is to determine how clearly the candidate fits in a pre-existing mold of what "successful" should look like.  The Matthew effect has been statistically shown to inflate the impact factor of research papers, as well as bias all sorts of empirical outcomes.

This gets to the third, and most revelatory to me, takeaway of the evening was the idea that the judgements that mattered were those that came from or evoked feelings.  Art as a way of interrogating emotion. An overarching theme of the biopic (itself generated and curated!) was a movement towards accepting and celebrating that aesthetic is based on feeling.  As someone who wants to diagram sentences and prove theses, this is difficult, but as I think about it more I think it's difficult because it is true.  This goes to the core of why generative AI in its current form cannot produce great art.  It is skill without feeling, judgement based purely on what an abstracted population judges.  Eno is in this sense the anti-genAI, and the world is greater for it.


Comments