Business Model Determinism

The Napoleon Dynamite Problem Problem

Ashley Bickerton - Tormented Self-Portrait (Susie at Arles)

In the mid to late 2000s there was this thing called the Napoleon Dynamite Problem. It had to do with the challenge of recommendation - accurately predicting how someone would rate a movie. Netflix even offered a million dollar prize for anyone who could substantially improve their algorithm.

You don’t hear much about this problem anymore. Did it turn out to be easier than they thought? Is there one dominant approach that everyone now knows and uses? Are there a variety of custom approaches that are jealously-guarded industrial secrets? In the case of Netflix it’s easy to see one part of the answer - they no longer care. This isn’t an important problem for them anymore. They no longer have a vast collection of movies they want to help their customers navigate through, they have a few dozen shows they’ve produced or acquired that they want their customers to pick from, the rest of the collection is ballast.

What would it be like to have direct access to a state-of-the-art recommendation engine, one whose entire goal function was trying to guess what things you would like? You can get a glimpse of this sometimes, when you stumble into a good groove on TikTok or Spotify or sometimes even YouTube. I’ve even heard people praise Instagram’s ability to serve up very appealing, well-selected ads. But there is no way to experience it directly, because there is no way for such a thing to evolve organically. The powerful content-suggestion algorithms which do exist and are hugely influential on our browsing, scrolling, surfing lives, are aimed at solving different problems - how to shift our attention in certain predetermined directions, how to extract money through certain kinds of ancillary transactions, and, most commonly, how to maximize time on device.

I want this product to exist because I think it would be amazing. I want direct access to a gold-standard recommendation algorithm that I control, one that isn’t embedded in a network of orthogonal incentives. I want to be able to ask it to suggest graphic novels, albums, sandwiches, TV shows, cocktails, couches, lightbulbs, anything. I want to be able to push and pull on it, to get it to go further afield or stay closer to home, to surprise me with unlikely suggestions or persuade me to check out established classics that I’ve overlooked. I want it to be trying to understand me, where I’m at and where I’m going, trying to discover the hidden patterns that connect the things I like, the same way I am.

A product like this would drastically improve my life. I imagine it replacing, or at least radically altering, the substantial portion of every day I spend idly browsing through various feeds, looking for distraction, entertainment, and edification. Instead of being pulled, passively, through these flows as they update, partially, slowly, clumsily, to my revealed preferences, I would be engaged in a process over which I had far more control and for which I took active responsibility.

I’m not so self-delusional as to imagine that, given such a tool, I would only read Marcus Aurelius and watch quantum mechanics lectures, I know there would still be plenty of times I just want to relax and veg out to mindless clips, times where I just want to check out what was happening in the meme-y world of the shared feed of daily public content. But I also know that, as a human, my tastes, interests and attention are shaped by what is put in front of me. I’ve seen my 3-year old granddaughter, on her way to do one thing, decide she wants to do something else, simply because the button for it appeared in her vision cone. I know I’m not so different. I want to stop being constantly bombarded by the click-me buttons of mindless clips and daily public content so that I can discover more things that I genuinely love, that will make me a better, smarter, happier person, but I want to do it with a computer, you know, on the internet.

I would happily pay a hefty monthly fee for such a thing, and I would be willing to open up my feed to it, sharing everything I read, watch, play, listen to or look at. But here’s the crux - there’s no way for anything like this to exist within the context of the standard business models for software products. Because there’s no version of this thing that is also promoting specific content, or showing me ads, or harvesting my data, or earning some share of the transactions that it helps initiate.

The process of exploring and evolving the edges of one’s taste, discovering, experiencing, responding to, and eventually forming value judgments regarding new stuff, is too delicate, too personal, too sensitive to slight fluctuations and shaped too much by recursive, self-sustaining dynamics, to co-exist alongside these other utility functions. What does it even mean to like something? Liking something isn’t just having it fit your pre-existing preferences. Liking something means allowing it to change you. This is not something any of the big tech or social media companies could offer because it would require a level of trust that they are millions of miles away from and could never, ever, earn back.

I do think a small startup could offer such a product, but I think the only way it could work is if they made their business model an explicit part of the deal. If they made a plausible, transparent, enforceable commitment to only charging for the engine itself and never making any money in any other way. Such a company would need to forswear any content development, advertising revenue, or promotional partnerships, indeed any strategic partnerships of any kind.

I’m not a naïve purist, I know art and culture is always made and consumed within a tangled web of worldly incentives, transactional relationships, and pragmatic forces. But I also believe there are some kinds of products that can’t co-exist with some kinds of business models. There are specific domains of product space that can’t be reached by companies that have the most common software business models - especially the most common current one, which is “grow your user base now and then figure out how to make money later.” I’m certain this is true for the SotA recommendation engine product I am imagining here, and I suspect it may also be true for many, possibly all, important AI-based products.

This isn’t a blanket condemnation of markets and capitalism. What I’m proposing is a way of deliberately engaging with markets in order to reach certain kinds of products that couldn’t otherwise exist, and then selling them for money.

The familiar process of “enshittification” by which all platforms rot and decay is only going to accelerate in the age of AI. Arguably, we can see it happening almost instantaneously in, for example, ChatGPT, where the RLHF fine-tuning that was a necessary step for making LLMs into a certain kind of acceptable product distorted and diminished the capabilities of the model in profound and unrecoverable ways.

Avoiding this trap will require acknowledging that your business model is part of your product, and designing both of them to work together to achieve an intentional outcome. As a game designer, I am highly sensitive to this dynamic, because, in a game, the user experience is the product. If, for example, you add gacha mechanics to your game to increase retention, you are now making a gacha game. We should not sit idly by and allow the voracious propagational inertia of gacha kudzu to ruin the future of AI-powered software before it even has a chance to exist.

I do not know, as a practical matter, what kinds of corporate mechanisms could be put to use to enable this, but I can’t imagine it’s impossible to make these kinds of commitments in your corporate structure, to make them clearly and simply, and make them in a way that is binding and persistent. And I don’t see any other way to reach a world in which the power of AI is harnessed, at scale, to enable us to find the things we want.