A Future Where Most Media Is Produced for an Audience of One
When we think about media, we're used to thinking about mass media: the shows everyone's watching, the op-ed everyone's talking about, the headlines we all see. The media business has always been fragmented, with local news, topic-focused publications, and niche genres. But distribution changes have made it easier for that content to get distributed to the exact people who care about it.
If there's a tiny population interested in all of rugby, haiku, classical guitar, and animal rights, that doesn't imply the existence of a single periodical covering all of these topics and reaching each interested reader. But a social media feed, RSS reader, or clever set of Google Alerts approximates the existence of one for every odd venn diagram of interests. One result of this is that the minimum number of consumers necessary to make a piece of content economically viable has gone down. Readers are easier to find and easier to track, so niche content thrives.
But there's a limit to this, defined by both labor costs and human interests: the floor for minimum interest in a topic necessary to make it worth covering has dropped, but not to zero.
But that assumes that humans are the main input. As DALL-E recently demonstrated, it's entirely possible to get humanoid outputs from non-human inputs. These are partly just responses to human-generated prompts, but they show a lot more creativity than humans are usually capable of. This is an eerie match for the prompt, for example:
If this kind of style transfer gets easier in some creative media, we could see a feedback loop:
- Art—or text, or eventually video, and possibly games—gets created based on a series of prompts.
- Over time, AI companies that create this kind of media will learn who responds well to which prompt.
- As long as people are more responsive to getting exactly what they want than to a particular quality bar, more and more of their content consumption will be bot-generated rather than human-generated.
Is that true? My go-to example for how much demand there is for niche content is Kindle's vast library of paranormal romance, gritty realistic romance, cowboy romance, helicopter pilot romance, etc. Basically any conceivable configuration of romantic archetypes has a library of Kindle romance content. I don't know how well it sells, but I'd bet that aggregate consumption of these and genres like them is higher than aggregate reading time spent with timeless classics—and, if not, it's certainly growing faster.
The asymptote we reach is that for most people, most of the time, their reading and viewing material is created just for them, based on things they've either voluntarily expressed an interest in or have subconsciously indicated they like. When this applies to entertainment, it's merely strange, but applied to news it's fairly horrifying: will we all get completely fictional narratives crafted around our existing biases?
Not exactly: they still need to be seeded with something real. And the seeding process will create a fascinating transitional period. If text style transfer gets as good as image style transfer already is, there will be two phases: first, it will be great for content creators, who will be able to spin up niche publications on any imaginable topic. Instead of a binary newsletter subscription decision, you might be prompted to suggest a few topics you'd like to follow, after which you'll get regular newsletter updates on exactly what you care about. In this model, the creators are still important, because they're seeding the style matcher. On the other hand, the job is different; it's basically a stud farm for content, where writers provide a fairly minimal input and most of the work is done by someone else.1
But as the stud farm example makes clear, much of the value will probably be captured by someone else, too. For one thing, the next decade will be a very pleasant one for GPU designers and manufacturers. But also, style transfer raises lots of intellectual property rights questions. If you ask GPT-3 or GPT-5 or whatever to write a novel in the style of Steven King, but featuring people you know and don't like very much, does King get royalties? Does he deserve them? He did provide some of the original input, but if you're more likely to read that story than you are to read every single book he's published, clearly he's not providing 100% of the value. On the other hand, the AI media company providing the story also didn't create the specific thing you asked for all by itself.
Given how fast AI moves—consider how many people who work in the field full-time and were taken aback by DALL-E—it would be a good idea to establish legal precedents, whatever they are, while this is still a science fiction hypothetical, and not in a few years (months?) when there's real money on the line. (I am biased, naturally; the companies that do this will be very fundable, and will be able to afford better lawyers than I can. Also they'll be able to crank out AI-generated legal briefs at superhuman speed. If you're in the bottom 99th percentile of any job whose output involves typing, brace yourself!)
Helpfully, there is a precedent. Some people can choose to consume high-quality media designed entirely around their preferences. These people tend to be executives with research staffers. Hedge fund and VC fund managers are a classic category: they can definitely order up a well-researched writeup on whatever topic they happen to be interested in at the moment, and can expect it to hit certain standards of accuracy and readability. But sometimes this misfires: is the job of an analyst to tell their manager what the manager wants to hear or what they need to hear? Ideally it's the latter, but you can never be sure: one feature of bonus negotiations is a debate over who gets credit for what ideas, and oddly, people have a photographic memory of their contributions and a bit of amnesia when it comes to their mistakes. The analyst's job is definitely easier if they're right, but only if they can tell their boss really wants to be convinced.
Financial returns provide some good guardrails here, as do industry norms. It's pretty typical to put together an investment memo with an explicit discussion of the upside case and downside risks, so in one sense the idea isn't complete until it's been checked against how it might go wrong. But in other domains, there are fewer guides. Someone in a position to be tasked with writing a memo on, say, how much of a risk Saddam Hussein's WMD efforts were would probably be a diligent, competent individual committed to doing a good job, but depending on the definition of their job the end result could be quite bad.
If you have an unlimited source of content that exactly matches what you're interested in, and is getting better all the time while also improving its model of what you, specifically want, will this be addictive? Game addiction is already an issue. Within games, people seem to get addicted to game mechanics instead of the graphics and story. (There’s not enough story for someone who plays thirty hours a week! And if you spend a thousand hours optimizing supply chains in Factorio or Hearts of Iron, it’s probably not because you’re craving a new animated explosion.) So customized content might end up being a compulsion even if the quality is not that good.
It's possible that in this hypothetical world, human-written content, or at least human-edited content, will still be valuable, though eventually more as a status symbol. Recorded music didn't destroy live performances, though the "live performance" of a blogger or newsletter writer is a very boring one. Part of the process of commoditizing an output is that roughly the bottom 95% of performers lose their livelihood, and the top tier do better than ever, at least for a while. But eventually, the output gets divorced from the original writerly input. If AlphaGo Zero could create novel Go-playing strategies back in 2017, then it doesn't seem inconceivable that entirely new artistic styles will be developed by more complex systems.
If this happens, the world will remain multiplayer, but be increasingly designed around a single-player experience. Video games are a particularly rich source of analogies here, because game designers are consistently tasked with making a game enjoyable to players with a wide range of skill levels, who may be playing together. For some games, this means gently introducing game mechanics and concepts, or promoting an ecosystem of online tutorials. For other games, like Elden Ring, it’s almost the opposite: the game promotes a mystique of extreme difficulty, so players feel like their status as Real Gamers is threatened if they can’t figure it out. Similarly, for media consumers part of the task will be to decide if what they're reading is simply too good to be anything other than something customized around their biases. Even if everything in it is factually true, the presentation will have a slant. And since this kind of media will wildly outcompete legacy products, the total volume of human-written news, commentary, and fiction could easily decline even as consumption of those genres goes up. Which will end up leading to a situation where the only things you can trust are old books, new data, and perhaps your peers; anything else will be trying too hard to please.
Further reading: I started thinking about this with Liberty's newsletter a few weeks ago. Tyler Cowen has less dire scenarios about email bots in mind. And for a more speculative look, Neal Stephenson's Fall; or, Dodge in Hell talks a bit about even more extreme variants on this scenario in the first third, before getting much weirder.
A Word From Our Sponsors
The world's best companies have a secret when it comes to onboarding their customers. Learn it!
Untangle the messy data your customers need to import using intuitive automation, and an elegant user experience built for simplicity.
This afternoon, I’m doing a Callin interview with William Hockey, a Plaid cofounder who acquired a bank a few years ago to rebuild its systems from the ground up, allowing it to offer much more comprehensive banking services by API. You can read more about the backstory here ($, WSJ).
If you have questions for him, please reply to this email with them so I can ask during the interview.
Amazon is allowing other retailer sites to use stored Amazon payment information, and Prime shipping, to deliver products. This points to the strategic tension in Amazon's model. The way individual business units make money is by building durable recurring revenue, whether it's contractual as in the case of AWS or more behavioral, as in instinctively going to Amazon.com to buy products. But Amazon as a company tries to operate in a decoupled way, where each business unit is treating other parts of the company as if they were external vendors; there can be some collaborative planning, but the cross-subsidies are done through capital allocation decisions, not by letting one unit ignore the costs it imposes on others. That model means that every expense for Amazon can turn into a revenue source, but it also means that some things that seem proprietary to one part of Amazon, but that are offered by another part, end up being just another part of the broader e-commerce toolkit.
Disclosure: I own shares of Amazon.
The Information has a great look at Netflix's internal debates on ad-supported subscription plans ($). The discussion had happened for years, but was apparently undecided until the earnings call. That's a pattern that shows up sometimes: if something is discussed consistently for years, it has obvious costs and obvious benefits, but their magnitude is still up for debate. It's hard to explore something like this gradually, especially since Netflix has partly defined itself through its ad-free experience for so long. If it were explored as a tentative strategy, the internal dynamic would be that pro-subscription employees would be implicitly competing with another model that happened to be run by the same company, and people would worry about career risk from both participating in the ad business and ignoring it. So if it's going to happen, the choice will be made suddenly; a coin toss is 50/50 while the coin is in the air, but never once it hits the ground.
In other streaming news, Warner Brothers Discovery is shutting down CNN+ a month after launch. Cable economics are always opaque, and may be somewhat opaque to the companies themselves: they have a good sense of how much people watch, but not a good sense of how strongly they prefer one channel to another. A single-brand streaming product can work if it's indispensable, and if its fans have plenty of spending power (in CNN's case, that's true; like most news media its audience skews old, and therefore richer). But bundles are spectacularly good when they combine background information with a few high-profile pieces of content. So it's not so much that CNN+ can't be viable on its own, just that CNN's massive viewership spikes during big news events could translate into recurring revenue for the Discovery and HBO back catalogs instead.
In Monday's piece ($) I looked at Twitter's tricky status as both a protocol and a product, and how these lead to very different business models. Twitter is tilting a little more towards the protocol approach by promoting some third party developers' apps to users. If Twitter thinks that its client and featureset can't serve all use cases, but that products built on top of the protocol could, then the right approach is to promote those as aggressively as possible—someone who uses Block Party or Buffer is a potentially-churned user who stuck around because they had access to a better version of the product. On the other hand, it could also use this kind of promotion within the context of the main Twitter product; the question to ask is whether the things Twitter promotes are closer to a replacement for the core app or an enhancement of it.
One of my favorite hacks is that the chemical mercaptan, which makes otherwise-odorless natural gas smell disgusting, warning people of leaks at home, is also a smell that attracts buzzards, which helps gas companies identify leaks in pipelines. In parallel with this, you can model the crypto mining industry as constantly searching for leaks where power is either underpriced or wasted. A recent example of this is Crusoe Energy Systems, which raised half a billion dollars to scale crypto mining associated with flared natural gas. It's a significant power source; the IEA estimates that flaring wastes enough natural gas to meet the demand of Central and South America. Gas gets flared because it's not cost-effective to build transportation infrastructure and because the externalities aren't internalized, but since crypto mining converts energy directly into money, and doesn't have to be especially concerned with intermittent availability, it's a natural fit.
- A startup that helps crypto exchanges, tradfi institutions and governments detect financial crime on blockchains is looking for someone with FedRAMP experience to help them achieve regulatory compliance. (US, remote)
- A company growing exceptionally fast in the home services market is looking for their first engineer, who will help build out the entire technical team. Netsuite experience preferred. (US, remote)
- A company building a wildly ambitious decentralized future is looking for someone to join their finance team, working on financial modeling and investor relations. (US, remote)
- A company building the next generation of wealth management tools is looking for backend and fullstack engineers. (US, remote)
- A startup which is helping local restaurants expand to new cities with a unique model is looking for a product manager to be responsible for food quality/restaurant operations. (San Francisco)
One interesting possibility here is that you might be able to get "someone who writes like X, only more so," for any value of X. You might mix and match styles, 50/50 one author and 50/50 another. Or you might end up preferring music reviews written by someone who is 300% Lester Bangs. If you've played "semantle," it offers a nice analogy . In Semantle you’re trying to guess a word based on its semantic distance from other words; in Semantle you try to draw a straight line from a less relevant to a more relevant word, to figure out what word continues in that direction. In writing style transfer, you’re also trying to go past whatever endpoint you’ve found.
As with food and drugs, the more concentrated form of something that's already appealing can be dangerously appealing. Doritos are a lot more compulsive than the corn they're made of or the cheese they taste like. ↩