Up From the Bottom (Official Video)

Need another 1393 views on this video to be eligible for a music award.  Need to reach 5000 views within 9 days!   Please watch and share!   THANK YOU!  I love you all!

The Computer as the Artist

Computers are now writing songs!   They’re now in the cultural mainstream, and getting more and more prevalent.  I’ve seen some short indie film soundtracks credited to AIs.  They’re already in the theatres!  Some programs can haul-off a piece of music after just a few mouse clicks describing the desired length, and style of the music.  Pretty soon we might see a program with a “write hit song” button. 

And critically.... the music isn't all that bad!   Maybe that's the most infuriating part: it's not utter crap.   It would’ve been easier to dismiss if it had been.  From what I've heard, it's easy to think a human wrote it.   It sounds safe and familiar.  Mostly familiar.  Listeners worldwide were spared an auditory version of an "uncanny valley", and wouldn't be the wiser to its artificiality. 

I used to feel relatively safe about not having my job replaced by a machine, or computer program.   Both as a family doctor, and also as a singer/guitarist/songwriter. 

Being a doctor is still pretty safe for now, only because we don't yet have robots that can do a physical exam.   But the diagnostic process of synthesizing data from a history and physical and matching it up against known conditions, is a task far better suited to a computer than the human mind.  Once a PC can do a physical exam.... look out! 

And as a musician I felt even more safe.   Many drummers and keyboardists were perturbed by MIDI and synths that held the potential to replace them.  But digital processes have never convincingly reproduced guitar strumming.   And vocals?  Forget about it!  And as for blistering solos......yeah, you could do that with MIDI if you put in enough hours, but doing it the real way would take WAY less time, be WAY more convincing and WAY more fun!   

But it was when I was wearing my "Composer" hat when I felt most safe from the threat of computers.   Here is a task that is purely human......engaging in the act of creativity!   This was bringing to physical life that which is intangible, often unaware of its purpose, or reason for existing.  This was a human activity if ever there was one!    And now the last of the purely human domains has become fresh meat to our AI offspring.  After having rewritten the laws of the possible in every other field, these pioneering AIs now bring their enormous impact to bear on the artistic world. 

In some ways, music composition does deal in a mathematical language, which computers can manipulate in ways our human brains just can't, and never could.  A computer could analyze 10,000 hit songs, find the commonalities, and compose a comparable song.   I'll admit it, when I'm writing a new song, I often will refer to a list of variables to check off, in my efforts to improve it.   For example, when switching sections in the song, I'll make sure to vary the phrase length or it's start point in relation to the downbeat.   I'll think about melodic contours and including interesting intervals.  And lyrically, I'll make sure to add more info to Verse 2, and put another point of view at the bridge.   So, in this way, I'm already thinking somewhat algorithmically.   A computer could make similar adjustments with its own list of variables.... variables found by analyzing more songs than a human ever could.   And what would it discover, if it analyzed the entire discography of humanity?   Would it have identified the "millennial whoop"?   Would it use it?   Steal it?   Modify it?   Perhaps it would analyze our own patterns of artistic re-use/stealing/borrowing, and recycle ideas in a similar way?   What other patterns would it recognize and recreate? 
 
And how original would it be?   Here's the really interesting bit.   Because in music especially, being too original, usually means the songs are not easy to listen to, and have limited appeal.   To be curt, too much originality sucks.  Really good music is derivative to one degree or another.   We heavily incorporate familiar motifs, to ease listeners into our musical world.   And when the AI's write songs, they rely heavily on what's come before.  That's really ALL they can do: analyze and recombine to create new works.   It’s not really all that different from the human process. 

But AI composers ARE faced with a fundamental speed-limit that would make true machine-learning impossible.    If an AI were tasked at getting better at a certain game, it could play against other AI's thousands of times, and thereby quickly develop winning strategies.  It can do this because the computer knows objectively whether game X was won or lost, and can adjust its strategy accordingly, without needing further input.    

BUT if the AI is tasked with writing better and better songs, it can't apply the machine-learning process it applies to games, because once it's written the song, there's no way for it to objectively measure how good the song is.   A computer has no inherent appreciation of music.   It can't tap its toe to the beat, or get a hook stuck in its head.   Simply put, it has no idea whether the song it has just written is any good at all, unless it seeks out that answer with human feedback.   And THAT takes time!  Such feedback-seeking would critically bottle-neck what would otherwise be a runaway process of trial-and-error improvements to which human creators could never catch up.

 

So, this gives me some hope.  But the time we creators have is limited.   I'm sure right now there are programmers out there trying to crack the code of how human appreciation of music might be broken down into quantifiable variables.   If we knew, for example, that the most compelling melodies have a third interval jump on the downward contour, being the most likely to cause gooseflesh in 70% of listeners, and 100 other similar factoids a future AI might dissect and exploit, then the computer might be able to predict and quantify your appreciation of a song before you hear it.   And then computers would be better manipulators of human emotion than we are.   

But for right now, at least.... what makes a hit a hit, is still clouded in enough mystery to keep reliable hit-making an elusive and unpredictable artform.  We don't yet understand these phenomena well enough to create amazing, immensely popular, critically-appreciated songs on demand.  It's still very much a process of trial and error.   

Even if we could find a formula that creates reliable hits, we'd get sick of the formula, and soon move on.   Our experiences and history factor too much into how we appreciate art today, and so we cannot write or listen to any music in a vacuum.  What happens outside the creative process, can immensely affect the music's impact.  There is a continually evolving relationship between art and its patrons, that impact upon one another in dynamic, ever-changing ways.   So, any given hit, may not be a hit in a different time and place.  We crave a critical blend of the novel with the familiar.  We get sick of trends.  We have 20-year cycles of appreciation.  A true computer hit-maker would be aware of all of these things, and be plugged into the cultural zeitgeist.  It would predict our cycles, and the perfect blend of the novel and familiar.  It might analyze and predict trends in tastes, styles, sounds, lyrics and melody that NO human composer has EVER been privy to, and manipulate us in ways of which we're not remotely aware.  An AI songwriter might blow away Paul McCartney and Max Martin!!  It might blow the doors off of any human that's ever lived. 

OR....they may never get it quite right.  They might end up being like the audio version of a fake Christmas tree:  dispensing what we once thought was a pleasing result, but ultimately revealing an ugliness due to a lack of imperfections.  Great music often manifests very beautiful imperfections, spoken in the language of errors.  It's like having a very complicated sort of math that you need to get wrong, in order to get the best possible result, and it may be a concept we ultimately are never able to teach our computers.   Perhaps. 

Another possibility is that they may never get beyond presenting repackaged versions of our aggregate past.  They may result in songs that confuse us, more than entertain us.... like being presented with someone you think you've met before, but just can't place.  And TRYING to place them now takes ALL of your attention, making it impossible to really hear the song with a fresh novel innocence.  Maybe that won't happen.  Who knows?   

Or maybe songwriting AI will end up being something we use only so often.... maybe to solve problems, like "computer.... analyze what I've written so far, and give me a chord progression for a bridge".  Maybe it will briefly take the industry by storm (like autotune), but then become just like every other modern musical tool.... having its place in moderation, but not a panacea, nor substitute for creative talent.   

Or AI may ultimately fail in enhancing our own creativity.  Because at its heart, creation is a very instinctive process.   We have the ability to write songs using a purely analytic method (the only method available to AI), but unlike AI, we also have the ability to just sing and play a melody/riff completely off-the-cuff without any conscious thought at all.   And we very often find that creating like that gives the best results.  Don't think about it too much, during the initial creative phase.  Don't think at all.  Maybe it's because we're human, and can only analyze a little bit at once, but we create the best art from a non-thinking, non-analytical place.  And maybe this non-thinking place is really the best way we can access our subconscious, which can synthesize way more info than our conscious minds.  Perhaps giving our subconscious more of the reigns, allows for a better result, as it's drawing from a vastly richer pool of resources.  That's a very human way of creating; one of which we know computers aren't capable. 

The scary thought here though, is that we DON'T know if that will matter.  Because the AI has so much ability to crunch through pure analysis, that it may not need the blank-mind cheat to which humans must resort, in order to process more than our conscious minds are capable of.   Since a computer can process so much anyway, maybe it wouldn't need a digital version of our "subconscious" to serve the same purpose. 

In any case, I'm very excited to see what the future holds.  I can't wait to see what a consumer-level AI songwriter plug-in can do!  Maybe it'll solve my bridge writer's block, or give me ideas for better choruses!  But then again, maybe it'll be so popular, that everyone will use it, and every song will sound like it came from the same composer.  And maybe various AI composers will end up competing with each other on the charts, until a clear and uncatchable winner emerges, and the rest of the songwriting world is helplessly looking on in slack-jawed amazement, and would fall into depression at the loss of their careers, if they weren't so sublimely entertained! 

And maybe after DEEP BLUE gets his 15 minutes of world-conquering fame, he too will have to face the harsh reality of human fickleness, as what was once hot, becomes not, and he becomes just another flash-in the pan.... a one-trick pony.... a Has-Been!   Humans can be cruel customers, and have never had to share the charts with another entity before, so things could get ugly. 

Either way, I'll be very keen to see how it all goes down!  I'll be sitting in the front row with a big bag of popcorn, and a pen and paper, just in case inspiration strikes!   And if ever inspiration abandons me and never returns.....maybe then, just maybe, I'll see what happens with a few mouse clicks.   

Nah!  Just kidding!

Steve Harley