Generative Adversarial Relationships

Human creative systems through the lens of our AI counterparts.

In the 2009 documentary “The September issue”, a creative director named Grace Coddington repeatedly butts head with notorious fashion editor Anna Wintour in the course of creating campaigns for the September issue of Vogue. Grace creates an exploration of “textures” and is forced to cut a latex outfit: 

“Smooth is not a texture, it's the absence of texture” 

She makes a vintage-style shoot using old lenses, taking a bold stance against the dogma that images need to be tack sharp to be beautiful: 

“Why are these images blurry? If we can’t sharpen them we're going to have to cut this.”

Each campaign, meticulously planned with a compelling point of view, is ruthlessly modified, cut down, or axed. Grace is fuming, Anna is coldly agitated. But at the end of the day the issue launches to great success, and Grace is surprised and proud to see how many of her campaigns have made it to print: “not half bad, actually”.

This is all well and good, but I will confess that I had to google Grace Coddington’s name. Unless you work in the fashion industry I doubt you’ve heard of her either. When Kanye West made his infamous “Famous” music video featuring nude silicone versions of celebrities lying in a bed together, Grace Coddington's naked body was nowhere to be seen—but Anna is there. 

The human animal is prejudiced towards these charismatic megafauna: larger-than-life individuals who stand in for those unintuitive networks that they have come to represent. Anna Wintour is the highly visible head of a massive creative system that works together to produce the magazine that we see in the grocery store aisle. The same can be said for all of the people in that naked pile of bodies. 

Thoughtful modern historians rightfully tend to look back at history and talk about economic incentives, changing environments, shifting social mores… all sorts of boring stuff. But the first and second-hand accounts of events almost universally focus on the personalities of the big names of history: these kings and popes and revolutionaries that shape our world.

The creative robot uprising

Computers, however, are not bothered at all by how credit is assigned: they have no egos, emotions, or even pesky consciousness itself to get in the way of cranking out creative output.

~*~* 𝓂 𝒶 𝒸 𝒽 𝒾 𝓃 𝑒   𝓁 𝑒 𝒶 𝓇 𝓃 𝒾 𝓃 𝑔 *~*~

Hark! The phrase is soft, carried on the wind. The people breathlessly whisper the words, eyes wide, mouths agape:

~*~* 𝒶 𝓇 𝓉 𝒾 𝒻 𝒾 𝒸 𝒾 𝒶 𝓁   𝒾 𝓃 𝓉 𝑒 𝓁 𝓁 𝒾 𝑔 𝑒 𝓃 𝒸 𝑒 *~*~

This software, at the center of our aspirations since the dawn of computing, represents a unique hope for our society: one that could free us from the shackles of having to actually do anything forever.

The field of artificial intelligence moves surprisingly slowly. Most of the biggest advances that have created the illusion of progress in recent years basically boil down to “what if we take old ideas from the 80s and scale them up to run on modern-day supercomputers?”

One recent approach that has proven to be a genuine advancement is the “Generative Adversarial Network” architecture. At their simplest, GANs are a collaboration between two independent neural networks designed to work directly against each other: a generator and a discriminator. The generator learns how to synthesize counterfeit panels of Garfield comics, and the discriminator learns to spot fakes. The generator iteratively learns to outsmart the discriminator, and the discriminator learns the new tricks, in a sort of Tom and Jerry style creative arms race.

Just imagine if these neural networks were human beings: the sheer frustration and bitterness that would bloom. Managing emotions would present a significant challenge. In fact, an enormous portion of art school is dedicated to “critiques”, in which students learn to harshly tear down each other's work while gracefully defending their own. Early on this can lead to bitter explosions of raw emotion, but through repetition the students slowly become perfectly desensitized to their feelings of inadequacy. Desensitization is necessary to prepare for the reality that bitter conflict is the norm in creative environments, and it is adversarial systems of human beings produce the content that we see in the world today.

We have many names for pairs of these “generators” and “discriminators” in the human world, working seemingly at odds with each other to produce higher and higher qualities of work:

Writer vs Editor
Artist
vs Curator
Director
vs Producer
(or perhaps Actor vs Director)

To be a successful creative person is to expose yourself to many sources of brutal competition, conflict, criticism, and rejection. Driven individuals who can make the cut in discriminator roles rise to the top of the pyramid, piloting their organizations and gatekeeping distribution channels. Whereas generative people—even when they work within an organization—are fundamentally outsiders who can only hope to reach legendary status by writing their own ticket.

The adversarial relationship within

“But Drew, I haven’t left my basement in 3 years and I’m already halfway finished with my epic fantasy novel about a beautiful troll named Num Lock. Surely SHE is not the product of of an adversarial system!?”

Yes, even Num Lock is a product of such a system. Because what is a brain, really, but a big squishy computer that fits neatly into the narrative that I’m constructing?

You sit down to write something and the words spill onto the screen:

“yes, yes, YES!”

You get up for a cold bottle of mountain dew code red and take a minute to reread what you just wrote:

“No... no no no.”

There appear to be multiple little people that live inside your brain. You put on your “writer hat” to write, and then later you put on your “editor hat” to rewrite and unwrite. 

Your internal discriminator is supernaturally powerful, like a little Jason Voorhees that lives inside of you: this editor hat ruthlessly cuts and slashes at your ideas in cold blood. Common and generally good advice for writers is to avoid editing while you're writing entirely.

In brainstorming sessions taking place in office parks around the world, “creative professionals” set down their Keurig coffees and engage in what we in the biz call blue sky thinking, rattling off new ideas as quickly as they spring into mind. “No wrong answers!” The success of these efforts depends on how the group lives up to those ideals in practice: can people really afford to look stupid in front of their colleagues? What ideas cannot be said for fear of damaging certain egos? Even with the bluest of skies, at the end of the exercise the discriminator awakes from its temporary slumber at the bottom of a lake: the ideas are rounded up and killed off, cut down to size, transformed, and—if they are any good—acted upon.

Ira Glass has a well-known piece of advice about this: when creators are first starting out, they are setting down that path because they have good “taste”, and they become discouraged because the amateur work they are producing doesn't live up to it. Their discriminator is a full-grown beast, but their generator is a newborn calf walking on shaking legs. Young generators are precious: they must be protected and nurtured.

Is it... getting hot in here?

Many machine learning models, including the famous GPT3, have a setting called “temperature” which controls the variance of the output. Low-temperature settings make the model output “safe” options, which are... pretty boring:

[WP] You are a time traveler from the future. You have been sent back in time to kill Hitler. You arrive in the year 1939, and you see a young man with a moustache. You know that this is Hitler, but you can't bring yourself to kill him. You decide to follow him around for a few days, and you see him do some terrible things. You still can't bring yourself to kill him. You follow him for a few more days, and you see him do some terrible things. You still can't bring yourself to kill him. You follow him for a few more days, and you see him do some terrible things. You still can't bring yourself to kill him. You follow him for a few more days, and you see him do some terrible things. You still can't bring yourself to kill him. You follow him for a few more days, and you see him do some terrible things. You still can't bring yourself to kill him. You follow him for a few more days, and you see him do some terrible things. You still can't bring yourself to kill him. You follow him for a few more days, and you see him do some terrible things. You still can't bring yourself to kill him. You follow him for a few more days, and you see him do some terrible things. You still can't bring yourself to kill him. You follow him for a few more days, and you see him do some terrible things. You still can't bring yourself to kill him. You follow [...]

This story happens to be pretty reflective of the decision-making process of the bot itself: it just can't seem to gather the gumption to actually do anything about Hitler, so it just keeps following him around.

High-temperature settings, on the other hand, cause the system to make bold and sometimes insane decisions. Perhaps in the “hot” version of the story the narrator decides to feed him rat poison in a Vienna sausage, only to discover that Hitler is actually an alien with psychic powers. Spicy. More obviously stupid things are generated, but so too is there a higher frequency of truly creative leaps and successes. This high-variance style is actually familiar enough that there’s common term for it: hit or miss

“This guy keeps adding weird jokes into his essay... some of them are funny but overall it’s pretty hit or miss.”

In advertising, you are not remembered for the 1000 campaigns that you created that are forgotten, you are remembered for the 2 or 3 campaigns that cut through the noise. And in order to do that, to really be memorable, you have to make big choices. Of course, those 2 or 3 successful campaigns were financed by the other 1000 clients who received mediocre work, but nobody said advertising isn't a grift.

Finding the perfect partner

Finding a creative partner is like finding a good spouse. It’s easy enough to find someone that you vibe with, but there are many other practical considerations you need to consider before you commit. At a certain level this may not even be a metaphor. in certain species of bird the males grow colorful feathers, perform dances, and build fancy nests for the females to choose from: males can probably be thought of as “generators” and females as “discriminators” from an evolutionary standpoint. It's important to choose wisely—whether it's an arranged marriage, or a marriage of the heart, your spouse determines how your babies will turn out.

Creative frenemies

You have to be really good friends with somebody if you want to be able to scream in their face. Fair-weather friends, and people who don’t actually care, will simply give up and walk away. Even better than actually being friends, though, is when there’s some sort of prize on the table that nobody is willing to walk away from, no matter how brutal things get.

Bandmates with a taste of success often develop seething hatred for each other, bound together by the mutual pursuit of money, fame, and glory. Fleetwood Mac’s best album was made by two recently divorced couples bitterly finding new ways to say fuck you to each other while snorting cocaine, for example.

John Lennon and Paul McCartney too were famously at odds with each other: ultimately leading to the breakup of the Beatles. But the songs they made in their solo careers were mostly mediocre: Paul put out soulless trash, and John put out rambling nonsense. When people come together to create as equals they usually appear side by side: simultaneously acting as a generator of their own ideas, and as a merciless discriminator for their collaborators’, while each focusing on different aspects of the output. Working together, John and Paul made up for each other’s weaknesses.

The flavor of a well-chosen cherry

With GPT3 people tend to “cherry-pick” responses that make the output look a lot higher quality than it is in general. What critics of this often don't acknowledge is that cherry-picking is part of the human creative process as well: good generators work in a loose, high-variance style with many “dud” ideas among their brilliant insights. It would be embarrassing to release this unpolished output.

Discriminators must have in their heads a model of what success looks like, which can roughly be called their “taste”. They work to conform the output to that model, and there are two ways of going about this: one is simply filtering out ideas that fall too far outside of it (“cherry-picking”), and the other is actively rephrasing, rewording, reworking ideas into more effective versions of themselves. 

Discriminators are only as successful as their taste reflects reality. If they shape things into “circle-shaped pegs” in a world full of “triangle-shaped holes”, you have a problem. On the other hand, it’s easy to over-index on producing what you think the market wants:

I'm writing a song all about you
A true song as real as my tears
But you've no need to fear it
'Cause no one will hear it
Sad songs and waltzes aren't selling this year

I used to volunteer at the “Vancouver Island Music Business Conference” when I was in my early 20s. For the years that it ran, it successfully convinced some big names in the Nashville songwriting scene to come and get wasted on an island off the west coast of Canada. I will never forget the way they talked about music: it was cold, calculating, and strategic. More similar to playing a game like chess than doing any sort of spiritual or creative act: 

  • They would start with their target audience: the only people who bought CDs at the time were suburban minivan moms, so everything was made to cater to their tastes. 

  • They would periodically listen to the billboard top 100 tracks in every market and identify emerging trends. 

  • Songs needed to be catchy on the first listen, which meant that everything was in Major C, the instruments introduced themselves one at a time, and the lyrics were ultra-simple “hooks” that repeated often enough that you could sing along by the end. To find hooks, they would sit in bars and wait for drunk strangers to utter something profound.

  • Keep your personal life out of the writing, make things universal. 

  • The singer must always be morally upright and in a high-status position: “my friend” is the one who is an alcoholic. 

  • At any point that they are likely to turn the dial on the radio, throw in new surprises to keep them hooked.

This approach certainly sells records, and you have to respect craftsmanship and effective grifting where it presents itself. The problem is that it’s is too reactive: audiences don't actually know what they want until you show it to them, and any “rules” you develop can easily block you from seeing opportunities. Taylor Swift certainly doesn’t keep her personal life out of her songs. Radiohead’s “Creep” didn’t make Thom Yorke sound cool and high-status. “Bohemian Rhapsody” is not a simple song written in Major C.

Oh to be blessed with a capable enemy.

We've only really talked about collaborations so far, bitter as they may be, but another way that this dynamic plays out is in true competition: survival of the raddest and most bitchin' out in the field. 

“Scenes” are created and propped up by the existence of rivalries between 2 or more parties, who just so happen to agree about the rules of the game that they are playing against each other. Any logical flaw or lame rap lyric will immediately be pounced on by ruthless competitors: after all, reputation is at stake here.

Mozart would not have been nearly as good of a composer if it were not for both the high expectations of the court and rivalry from contemporaries like Salieri—even if Salieri is basically only remembered as a foil for Mozart. The scene that they were a part of drew out his best possible work: if it were not possible for Mozart to score favor with the royalty, earn respect from his peers and onlookers, get laid, and put that stupid Salieri in his place [authors note: eat shit Salieri], why would he bother?

I read somewhere (but am unsure how to look up if it’s true) that once Michael Jordan was dominating the NBA, he switched his focus from becoming an even better basketball player to maintaining the health of the league itself: fostering new talent to compete against. He (presumably) recognized that the higher the overall skill level of the players was, the more interesting it is for audiences in general, and the more money floats up into his pockets.

Without a good enemy, it can be lonely at the top.

Society-scale creative systems

Once your creative output is released into the world, it is subject to further systematic forces that are more powerful than any individual or small group. You come to find that you are merely one line worker in an insanely large factory.

The Feed

I make a lot of bad tweets.

I used to go back through my timeline and ruthlessly delete almost everything. Then I learned a fundamental truth about twitter: nobody really cares. Tweets that nobody likes are hardly seen by anyone. What people see on their timeline is what other people have already liked, the “better tweets” as determined by engagement and the algorithm.

The algorithm itself becomes a discriminator and a dance partner that shapes the content showing up on each particular site. What we see on Twitter vs Facebook vs Reddit vs Tik Tok vs 4chan is a result of each unique algorithm that surfaces different content, and shapes the character of the discourse. The platform is at least half of the creative system. 

Reddit surfaces annoyingly agreeable content (and creates niche echo chambers where everyone agrees that “fat people are subhuman” or whatever), 4chan surfaces shocking edgy content and reusable images that slowly infect the rest of the internet in a sort of “trickle-down meme economy”. Instagram surfaces butts, and Twitter surfaces enraging clickbait and pithy jokes.

It has always been the case that algorithms have shaped content on the web: how many websites have been molded by the google search algorithm and the SEO cottage industry? We are all annoyed by recipe sites that start with extremely long-winded personal stories, but that is a direct side effect of google’s ranking algorithm that prioritizes “original content”.

Capital D “Discrimination”

Up until this point, you may have been chafing at the term “discriminator” because of its rather negative connotations:

“Is my book editor... racist?”

The answer is, of course, absolutely. Our good recently deceased friend James Carse makes the case in his book “Finite and Infinite Games” that society itself is a machine that transforms raw materials into “goods”, and produces and discards “trash” as a necessary part of its functioning. Some of those goods are functioning adults, and some of that discarded trash is human beings: bums, alcoholics, junkies. Tucked away on reservations, halfway houses, and mental institutions. Out of sight, out of mind. 

It's unpleasant to see yourself as the creative product of a system, and it's depressing to think about how the sausage is made—I, for one, especially don’t like to think of myself as a sausage. The world it seems is a cruel place, and we are all on some level clawing to remain on a vanishingly small stage of relevancy.

We prove our value to society through our usefulness: this is a universal truth known by 3-year-olds and senile 95-year-olds alike. 

Thin skin

Neural networks don’t have feelings… this is a point that I keep coming back to. They don't feel bitter at the end of the day. They don’t give up on promising projects due to “creative differences”. I have, regretfully, proven time and time again to be too sensitive. I’m hurt easily. I push myself past my limits to avoid offending people, and I block well-intentioned people out of my creative process for fear of what they might have to say.

But how much would I want to change myself? To become desensitized and machine-like in order to produce more work efficiently? Maybe that’s what it would take to make it in “the art world” or “the entertainment industry” or whatever, but I rather like being a sensitive person. It’s what drives me to be curious about the world in the first place. And what sort of an artist isn’t sensitive? Even if GPT3 can write passable poetry, ultimately I don’t give a shit about what it has to say because it’s not reflective of any true experience. Allowing yourself to be vulnerable gives people the opportunity to see their vulnerability reflected in you, and it’s very important part of art—at least the kind of art that I like.

Nevertheless, I am afraid. And that is something I will admit that needs to change. Many years of my creative life were never shared, and there is no good record of what I was thinking about in my 20s because of that. Probably for the best I suppose. Machines are not afraid to produce terrible crap, or to try their best for fear of falling short. It does help, as a means of escaping self-judgment, to think of my work as the inevitable product of forces outside my control: to follow creative processes, step by step, with faith that everything will work out in the end.

It is not in my constitution to retreat alone to a Walden pond to work. I’m still hunting, as always, for collaborators who might be willing to put up with me. People who are different enough from me creatively to make up for my shortcomings, but similar enough that they can understand where I’m coming from; who care enough to believe in things and fight for them, but are kind enough to not become malicious. People who surprise me. People who are fun.

If you are interested in hearing more from me and you’re not already signed up for this newsletter I would appreciate it if you signed up:

You can also follow me on twitter (DMs are always open), or email me at dschorno@gmail.com