It’s (sort of) lost in the digital churn, but a long time ago my entry point to the group weblog Terra Nova, a set of scholars across the disciplines who had staked something in the relatively new genre of massively-multiplayer computer games that were referred to as “virtual worlds”, was an essay where I argued that the fundamental problem with the genre in its early going (games like Ultima Online, Everquest, Asheron’s Call, back before World of Warcraft existed) was its character-centric vision of persistence.
Let me unravel this a bit. Persistence was at that time one of the signally interesting features of these virtual worlds, and a major reason they drew so much academic attention. These were not merely games where you logged on and competed with someone else for a single session, but games where what you did in one play session carried over to the next, where the results of your play accumulated over time, and you logged back into a gameworld where other people were also persistent in that way. That alone was enough for a kind of economy to develop, for social relations to develop, for otherwise ephemeral discourse between players to attach to persistent identities and so on.
So my basic argument, which I remember irked the venerable sage of virtual worlds Richard Bartle enough to get him to announce it both unrealistic and unappealing, was that persistence should be in the world, not in the character. I was making the argument rather chimerically, one-half scholarly, one half-as a player whose preferences were going unmet. (I’ve argued for decades that there’s a hidden “Bartle-type” that I think of as the demiurge, the player who wants to create within a virtual world, and not merely items or possessions belong to characters, but to tell stories.)
What did I have in mind? That era of virtual worlds, and most of those that followed, vested accumulation in the character who represented the player. That meant that what changed in the world itself was mostly individual. It was in that sense almost a parody of methodological individualism, or classical liberalism: the individual’s steady accumulation of resources and capacity was the only thing pushing the evolution of the virtual world’s sociality, and for the most part, individuals accumulated in a strict ratio of time invested in play to rewards derived from that labor. No lasting better mousetraps or comparative advantages here—when players figured out the underlying code enough to spot an imbalance in outcomes to one kind of character, players shifted into that niche so hard that its competitive advantage was more or less instantly negated. Usually at that point the developers would try to “rebalance” the game by weakening that type of character—a never-ending labor.
But because the gameworlds were persistent, certain kinds of advantages accumulated regardless, most particularly anything that could function as currency or that had in-game value because of its rarity or power. Almost from the beginning that produced real-world markets—Journeyman’s Boots in Everquest, for example, let characters traverse the gameworld more rapidly, which effectively reduced labor time for task completion. The supply of J Boots was limited and effectively was “mined” at regular intervals when they became available, with the successful miners then often reselling the item for real-world money through exchanges outside the game itself. (There’s a real link between the concepts and practices that define cryptocurrency now and the virtual world economies of the early 2000s.)
From a ludological standpoint—that is to say, the experience of actually having fun—centering persistence in individual characters and expressing it primarily through the accumulation of resources and wealth that were created either through an investment of labor time or through a real-world exchange of money for virtual wealth made virtual worlds less like worlds with stories and more like neoliberal simulators attached to a social media platform. You related to the game through an individual avatar who gained levels and got more powerful gear and had money. You generally belonged to a group or guild of other players who worked together, and that group had as much dynamism and mutability as any other sociality you might be a part of in the rest of the world. But the gameworld only changed or evolved when the developers manually added new content or made manual changes to existing content. You and your allies could ride forth and battle your enemies, defeat an army of foes, and an hour later find the same army still there.
So what I suggested in this long-ago essay was to make a virtual world where persistence was in the world, not in the characters. What I had described was the following:
Much of the world’s content should be generated procedurally so that new content—new characters, new structures, new narratives—continuously enters the game and gives it that sense of being alive. As I saw it, there would still be room for handcrafted designer content to enter as well, but a living virtual world needed procedural content.
The world should be extremely large and the player count on any individual server very small so that there might be forms of play that were closer to being solo and so that there would be lots of space for players to spread out away from one another.
The world should be deformable: meaning that as gameplay proceeds, the world should permanently alter in response to gameplay. If the textile guild’s headquarters in a far northern city burns down, it stays burned down. If a small village is destroyed by an ogre invasion, it stays destroyed. If a mad wizard builds a mysterious tower with a dimensional portal to hell in it, the portal is there from that point onward until or unless someone seals it.
The world should be richly populated by autonomous-agent AI-controlled non-player characters who are continuously generated as existing NPCs age and die. The NPCs should have something like a Maslow’s hierarchy of needs (only more complex) at generation and it should be somewhat randomized, with additional procedural elements built into the characters. (Place of origin, associations with the gameworld’s societies and structures, etc.)
The real-world players of the game should not have characters in the world. Instead, they should interact with the world via indirectly influencing the actions of non-player characters. Essentially, you would play via being a kind of “guardian angel” (or devil) sitting on the shoulder of an AI and trying to push it towards actions in the world that would lead to consequences.
Basically, what I was imagining was something like The Sims + Minecraft + Elder Scrolls Online.
So the objection at the time was this is not possible with current technology. That was absolutely true. The other objection was this is not fun. That was wrong, and I think The Sims and Minecraft subsequently demonstrated that quite well. Minecraft was everything that skeptics about procedural content said could never happen and would never be fun if it did; The Sims demonstrated that there’s plenty of fun to be had in manipulating what third-person characters do and generating unpredictable and emergent stories from that.
Why am I returning to this idea, two decades later? Because it is now possible. Procedural generation has become much more sophisticated (though it can still go horribly wrong if it’s treated naively or as a cheap substitute for design work. The remaining bottleneck was on the agent-based AI side. For a world-centered game to work, the NPCs you encounter need to feel individual and they need to be responsive both to your manipulations and to each other. But this is precisely what GPT and its cousins will allow.
E.g., going back to the original essay, I imagined that a player might decide that they wanted to engineer the rise of a new religion in a major desert city. So they might flit invisibly around the desert city listening for an appropriate NPC to work through—say the impoverished priest of the cult of a Jackal Spirit. For that to happen, the NPCs need to be in motion and in conversation all the time, persistently, even when players aren’t there to influence them, much as Sims in a neighborhood are all doing their thing while you focus on trying to get two thirty-something Sims to get married or learn a new hobby. Up until now, no AI could possibly have generated syntactically and semantically coherent conversation strings from a procedural seed. But now that’s pretty imaginable. When I find my Jackal priest, now I can interact with her and even potentially alter her behavior or aspirations through conversation. “You should hire several members of the Thieves’ Guild and steal the treasury of the Bat Cult, that will let you build a new temple for the Jackal Spirit,” I say. But then I find the Jackal priest has “Personal safety” high on their randomly generated Maslow’s hierarchy. I’ll have to work around that, or find a new cat’s paw. In the meantime, maybe another player in town has decided to build up the power of the Bat Cult, so we’re on a collision course. We just don’t know it yet.
This is the first possible use of LLM-based generative AIs that I feel genuine enthusiasm for—a use that would be genuinely creative and would satisfy a new need for which we could never hire sufficient numbers of human workers. So here’s hoping someone’s already hard at work on an idea of this kind.