Endless Media

https://www.generalist.com/briefing/endless-media

Among trivial agonies, few are as painful as finishing a good book. It is a strange heartbreak, simultaneously too big and too small. You are losing an entire world of characters, struggle, and meaning – but one that is entirely invented, holding no real purchase over your life. 

A story can be made or ruined by its ending, a fact any author or filmmaker knows. Hemingway wrote forty-seven versions of A Farewell to Arms to “get the words right,” while Kubrick fussed with the closing of The Shining until the very last moment before ordering the destruction of remaining takes so that no interloper could rearrange his composition.

Allow me to adopt my best infomercial voice to ask: What if there was another way? What if the stories you know and love didn’t have to end at all? What if you could keep reading your favorite book or watching your favorite movie forever?

Like most infomercials, the implied product sounds too fantastic to be real – as if Borges wrote copy for Billy Mays. For the sum of human history, that has been true. It may not remain that way for much longer.

The rapid maturation of generative artificial intelligence has not so much “disrupted” media creation – as if it were an irritating bumblebee knocking against a glass pane – as unseamed it. Instantly and cheaply, it is possible to produce stunning imagery and competent writing with AI algorithms. Like the precocious child it is, it seems inevitable that existing models will improve and new skills will be added until the creation of a new Hemingway novel or Hitchcock film is no trickier than spawning an image with Midjourney or short story with ChatGPT, OpenAI’s newest creation.

What will this mean for content production? How will it change how we consume it? Who stands to win and lose? And in an age of endless media, what will narratives mean to us? 

Going endless
First, a little more definition. What do we mean when we say “endless media?”

Fundamentally, we’re talking about a shift in how media is made and consumed. Rather than being created episodically, AI will enable it to be generated on demand. If you want J.K. Rowling or George R.R. Martin to write another book, you must wait for them to do it. Depending on how interested they are in writing, that might take a while. Years. Decades.

The same is true of TV shows, albums, games, and every other form of media. The creative process depends on human capital, which requires time.

By contrast, AI doesn’t have human limitations. It can process and generate information at much faster speeds. While it’s not capable of making a prestige HBO drama yet, we’ve already seen it is capable of writing blog posts or conjuring images much faster than a human can. A single DALLE-2 picture might take weeks or months for a person to produce but is spawned by the model in a couple of seconds. 

Once media can be generated instantly, it will change how we consume it since our desires will no longer be bounded by content availability. If you want to read another Lord of the Rings book, Murakami short story, or play a Red Dead Redemption-style game, you don’t need to wait for their creators’ effort (or resurrection). You’ll simply navigate to one platform or another and produce a new piece. 

You will also tailor your media in unforeseen ways, specifying setting (enter: neo-punk Middle Earth), plot structure (enter: a love triangle with Bilbo, Saruman, and an Ent), tone (enter: dark comedy), and any number of other variables. Creating original narratives will be near-effortless, and – depending on how copyright is enforced– creating fan-fiction-style mashups may be equally simple.

This is the world of endless media I believe is coming. How close are we today?

For this movement to emerge, generative artificial intelligence must succeed in three dimensions: 

1. Quality. AI must be capable of creating media that is high enough quality to sustain attention. Eventually, it should be able to create media of equivalent or higher quality than humans.
2. Speed. How fast is necessary for an on demand model to emerge? I suspect it will vary considerably by medium. Waiting minutes for every newly generated song as you make your commute is unlikely to be attractive, but to mint a new movie, it might be ok. 
3. Cost. Media generation will need to be much more economical. Specifically, it must be cheap enough for consumers to justify the cost through direct payments or attention.  

Depending on your benchmarks, one might argue AI has either met these criteria or is miles away. Products like Midjourney make it easy to generate a pleasing image in a matter of seconds for a few cents. When I asked it to create a Turner-esque alien seascape with frigates battling at sunset, it dreamt this:

Pretty good! Strangely, I feel proud of it, as if my prompting demonstrated taste or intelligence. It will be interesting to see how attached or remote we feel from our synthetic compositions.

OpenAI’s ChatGPT function – which uses a new, advanced model dubbed “GPT-3.5” – has made it even clearer how powerful a writer AI is. I asked ChatGPT to create several short stories in the style of authors like Cormac McCarthy, Haruki Murakami, Shirley Jackson, and Chinua Achebe. ChatGPT doesn’t seem to understand style yet. It mostly anchors around the authors’ favorite words, themes, and settings rather than truly mimicking sentence construction. A Nabokov imitation features a beautiful woman but doesn’t attempt his mischievous wordplay; an impression of McCarthy features a badland setting and roguish protagonist but doesn’t dare try to match his run-on, galloping prose. 

The results are nonetheless frighteningly good. Cogent, interesting, and redolent of the writer in question. Stories arrive in seconds and are entirely free. Experiments with poetry show GPT-3.5 understands rhyming schemes and meter, too. 

If your media diet was entirely comprised of pretty good paintings and solid short stories, you could already enjoy effectively endless media. More complex products remain out of reach. We do not yet seem close to producing a drama series that could supplant Succession or a prose poem Anne Carson could have written.

Whether the rest of this essay is a pure thought experiment or practical discussion depends on how you answer this question: will we get there? Is there a technologically traversable path from AI-generated blog posts and pretty images to facsimiles of Stendhal, Scorsese, and Swift (both)? If not, then we can keep our human theory of genius – AI will handle the pesky and derivative, leaving great art and great media to the minds of humankind. If so, then our remaining concern is one of timing. Is it arriving sooner or later? Next decade or next month? (Or, terrifyingly, tomorrow?)

Even before the year is out, we may see considerable progress. If rumors are to be believed, OpenAI’s GPT-4 is expected to launch between December 2022 and February 2023. Though details are sparse, it is being spoken of as a significant leap forward. Betting against rapid progress in generative AI increasingly feels like an untenable position. 

Value shifts

In a world of endless media, where does value accrue? What shifts should we expect?

When thinking about endless media, I have traditionally expected holders’ of valuable IP to benefit. If you own the rights to the Harry Potter franchise, the ability to generate infinite media inexpensively, on demand, would be a massive boon. You would cut production costs to near zero; better yet, you’d be able to monetize customers better, increasing frequency, reliability, and, likely, LTV. Instead of shelling out $20 for a movie ticket or book every couple of years, fans might pay $20 per month to access a service that shares new content from the Potterverse or allows customers to generate it themselves.

Wouldn’t customers get bored? Some might. But the popularity of fan fiction and fan forums suggests that consumers are looking for ways to exert their obsession with a franchise outside of traditional releases. Hollywood’s most reliable performers tend to be remakes or sequels. A 2017 study found that 41 of the 50 highest-grossing films ever built on existing IP; the sheer volume of reboots or sequels rose from 16% in 1981 to 80% in 2019. Disney has been particularly attuned to these dynamics and adept at exploiting them. Through its Disney+ subscription service, it has expanded box office smashes like Star Wars and the Marvel catalog into streamable series.

The extent to which IP holders profit will likely depend on how aggressively they are permitted to protect their assets. As more money can be made monetizing fans, franchises are becoming increasingly litigious, shutting down unauthorized games, merchandise, and experiences. While there is a benefit to protecting IP holders – especially when speaking about a living artist or author – over-policing may limit further creativity.

The companies behind large language models take differing stances on IP usage that illustrate how access to source material influences the content you can make.

When I entered terms like “Harry Potter” into OpenAI a week ago, a message appeared stating the search was against the company’s content policy. When I tried again today, I met no such friction, happily generating a rolling collection of IP hits: Harry Potter playing a magical saxophone, Mickey Mouse high-fiving Pikachu, and a neopunk Luke Skywalker roasting marshmallows with his lightsaber.

Though DALLE-2 allowed these prompts, it struggled to produce high-quality images. Attempts to generate characters like Hello Kitty and Pacman created unrecognizable results. Compare them to Midjourney, and you’ll see a stark difference. Similar prompts produced better quality images: 

(I can’t figure out why, but Midjourney struggled to understand the idea of two characters high-fiving. Still, Mickachu is so cute!) 

What accounts for this disparity? Likely the extent to which the models were exposed to relevant imagery during training. To avoid copyright issues, OpenAI seems to have ensured it has a reduced understanding of what Hello Kitty or Luke Skywalker look like. Midjourney – and Stable Diffusion – seem to have fewer qualms.

Much of the most interesting endless media will rely on entirely new IP. Rather than entering a prompt like, make me a new Star Wars movie, we’ll be able to say, make me an original space fantasy. Happy accidents will abound with machine intelligence spawning stories that surprise and thrill us, forging brand-new fandoms. Especially if access to aggressively protected IP is restricted, many of these works may be of better quality. (A topic for another essay: what kinds of narratives might be smuggled into new media by the makers of AI models, if we are not careful?)

Who will own this original content? The company that built the model? The platform that helped you create it? Or you, the prompter? This, too, will be a contentious question. DALLE-2, for example, allows users to sell images, but some lawyers consider the terms contradictory.

Regardless of ownership, we should expect a flood of content that makes current volume look quaint. At least in some regards, this may be a meaningful win for users, able for the first time to generate what they want to consume at the moment they want it. Looking for a mystery novel set in 18th century Japan in the style of Orhan Pamuk to enjoy by the fire? Or how about a darkly comic British detective drama set in the near future to watch on your next plane journey? All moods, whims, and predilections may be sated as they arise. 

If thinking of what you want to generate feels like too much work, you’ll also be able to dive into increasing infinities of content spawned by other users. New platforms will help us rank these endless masterpieces and guilty pleasures, and tools will map our interests to this vast compendium. Products that aid discovery, curation, and recommendation will accumulate value as abundance makes all three more important.

Making sense
Something I have been thinking about lately: technology renders reasonable-sounding analogies senseless. The scale and speed of the industry surpass our ability to simplify and translate. Too much nuance is lost as we try to stuff a universe through a letterbox.

With this in mind, we should speak of the great unfairness of generative AI when it comes to professional artists. The industry’s most powerful models have been trained on billions of data points generated by humans. Millions of paintings, cartoons, photographs, essays, and books are ingested and absorbed, feeding the model. The unique and strange skills humans have labored to develop are suddenly swallowed whole, turned into a commodity anyone can use. Now, with access to the internet, it is possible to paint like Van Gogh or animate like Disney.

Is this fair? In some ways, it feels like a fruitless question. It is happening, and it will continue to happen. What does fairness have to do with it? The tsunami does not have a conscience. When technology picks up momentum in such moments, the best thing to do is to ensure it is developed and directed as responsibly as possible – to guide its force away from vulnerable areas. 

The answer is, of course: no. No, it is not fair. And this is when analogies show their limitations. The counterpoint is that an AI develops similarly to a human. It is not so much mimicking existing artistic works – copy-pasting pixels or shuffling text to fit new circumstances – as learning. It studies and improves, just like us! 

Except, we cannot ingest five billion images. We cannot look at a handful of imitations and magically ape their style. It takes us more than five seconds to paint a portrait of the Pope eating an apple, sitting on a bicycle. Comparing AI’s abilities to human learning expunges its wonder and trivializes its risk. The unfathomable span of a galaxy or the contained infinity of grains of sand on a beach – these are more sensible metaphors. 

Will artists adapt? Some will be boosted by this technology, using it to improve their work and save time. One optimistic framing is that artistic work will be “centaurized.” Humans and machines will combine to create something greater than either might manage alone. For a time, “centaurs” were the best-performing chess players, besting both solitary humans and unaided algorithms. But as AI has improved, human intervention has become increasingly superfluous or counterproductive. In the medium term, though, “centaurs” may thrive, making impressive paintings, films, and books.

Depending on the rate of change wrought by AI, many artists may successfully shift into adjacent roles: the writer becomes an editor, the painter becomes an art designer. Those gifted at “worldbuilding” – creating the confines in which the AI plays – could also prosper temporarily. AI doesn’t seem to lose its appetite, though. It will likely absorb these tasks, too.

In many ways, I hope I am wrong about AI’s impact on artists. (I often am, I remind myself.) While I’m excited by AI unleashing new forms of creativity, I am still…a writer. A concoction of fingers, thumbs, muscles, and grey matter that requires food to operate and money to buy food. A world in which words are machine-made would be financially unfavorable and egoistically disastrous. (If I am not a writer, what am I? Please don’t let me go the way of the carriagemaker and lamplighter.)

One possibility is that as abundance becomes superabundance, we value human work more. Just as there is a market for handcrafted goods in the age of the machine, there may be a desire for human-made art, music, literature, and film despite AI’s efficiency and economy.

Beyond the shift generative AI heralds for artists, the velocity and volume of its creative output may have broader implications for society. In “The Decentralized Country,” I argued that the internet and social media have ushered in an age of “fractal truth”: 

The internet’s reduction of publishing costs and removal of gatekeepers has allowed the number of "publishers" to skyrocket. Of course, the definition of a publisher has changed radically, shifting from referring to an organization to an individual. Now, every person that tweets or posts acts as a sovereign publisher contributing to the information diet of other internet citizens. 

Under the weight of this expansion, Truth has become fractal. Not long ago, individuals received reporting from just a few news sources. While there might have been some variation in opinion, by and large, the number of perspectives on a given topic was limited. Today, every topic is refracted through thousands if not millions of competing opinions.

In the vast, indefinite hinterland of internet-space, distinguishing between these requires time and constant cognitive energy. Because the development of citizen reporting has revealed both the blind spots and biases of traditional publishers, consumers can no longer simply default to believing one source but have to bounce between hundreds and thousands. Users encounter far more varied and fantastical positions because there is no arbiter, and existing platforms favor extreme positions. The result is the creation of fractal truth, in which every person sees just a sliver of the overarching pattern but still believes in the authority of their position. Each new opinion shared online can be forked, opening up further fractalization.

E

ndless media will have a similar impact, creating an increasingly fractal culture. Although the internet already allows subgroups and niche communities to prosper, there are cultural phenomena that cut across. For example, if you were a literate child living above the poverty line in the 1990s, it was effectively inevitable that you had some relationship to Harry Potter. Maybe you loved it, maybe you hated it, maybe you were studiously indifferent – but you knew about it. And you could relate to others on that terrain. Variations of this exist for every mega-bestseller, box office smash, platinum album, and watercooler TV show. Though it’s so ingrained we may seldom recognize it; we have a comprehensive set of shared references.

Will this survive endless media? What narratives become part of our culture if our consumption is balkanized across custom-generated content? What shared language will we lose? 

“Since when are the first line and last line of any poem where the poem begins and ends?” Seamus Heaney wrote. 

Artificial intelligence will be responsible for many of the miracles of our lifetime. If we are pointed in our development, every field may see part of itself remade, from healthcare to education to heavy industry. Huge leaps seem possible. 

Like these other industries, media will be altered beyond recognition, changing our relationship to art, information, and meaning. The poem shows no sign of ending; let the film run; and listen as the music keeps playing.