LLMs In Games: Stop Chasing Infinite Content—Start Building Infinite Possibilities.
The following is part 2 of an interview with lead designer Eric Lindstrom.
Eric’s reputation proceeds him, but if you need a refresher: over the past few decades Eric’s led design and innovation spanning from an MS-DOS Sherlock Holmes adventure game to a Creative Director in charge of Tomb Raider to building live service mobile games. The man has seen it all. And now he’s on the front lines trying to understand what (if anything) LLMs can do for gameplay.
Charlie Witkowski (CW): At Atomic Dirt, we’re focused on using technology to create completely new kinds of gameplay. But let’s get a baseline here. How would you define video gameplay? And then what the hell is emergent gameplay?
Eric Lindstrom (EL): If I drastically oversimplify, a game is rules and content, with players navigating them to achieve goals. This navigation can come in different forms like puzzle solving, hand-eye coordination, making strategic and tactical choices, etc. Emergent play is where the actions a player can take, and the systems the game includes, can be combined in ways not explicitly designed for. This can be as simple as in the early days of shooters where players started using rocket and grenade explosions to execute longer or higher jumps, or later in much more complex open world games with many objects and a physics system, where players can combine them in logical ways not specifically planned.
CW: So emergent play is kind of finding your own fun in a robust but hard-coded game world like GTA or Baldur’s Gate 3. But, if I’m understanding you correctly, in these instances the games are doing a great job of feeling like open worlds, but they’re really more well-crafted sets or soundstages?
EL: I think open worlds are definitely open in the sense of players having a lot of freedom as to where they go and what they do at any given moment, but yes they are created sets – the freedom is not being required to follow a predefined itinerary. The world is open in that way, but it’s still what it was designed to be. You might get encounters in a different order, and they are sometimes tweaked to reflect your play path, but it’s all designed and scripted.
CW: You’ve played some early LLM powered games before that are more “open” as far as things you can try and incorporate LLM powered proc-gen with AI dialogue, item descriptions, etc.. How’s your experience been so far? What’s working well, what do you see as a gameplay challenge?
EL: What I've seen most is what seems to many like low-hanging fruit, to have chatbot NPCs. But these generally fail now because you can trick them into responding in ways that break the experience. Even if they successfully stay within the behavioral boundaries they’ve been trained for, you can talk about things they weren't trained for and get them to misbehave, like saying they’re handing you an object that the game doesn't actually give you. A different way I've seen LLM integration is allowing players to architect their own solutions to game problems, but these often fail because LLMs have a built-in framework to provide whatever is being asked for, so they let players succeed too easily and in nonsensical ways.
CW: You’re not describing what I'd consider to be a great player experience. So what’s to be done!? Is there hope!?
EL: So much hope! I think using LLMs for NPCs was a red herring. LLMs can pass the Turing Test but only because they can sound like *a* human -- making them adhere exactly to being *a particular human* is outside its wheelhouse. But focusing on what LLMs excel at opens up a world of possibilities. For example, if you give an LLM a hundred questions to answer, a lot of its answers will be wrong, because LLMs don't truly "answer" questions. But instead ask the LLM to concisely identify the intention of each question. So given a long, rambling paragraph, the LLM would say, "They want to know the capital of Scotland." Now look down the list of a hundred inputs, and what the LLM said was the intent of each, it likely got them all right. LLMs are very good at extracting intentions.
CW: Absolutely agree. In fact, we did a deep dive previously on this blog about LLMs being good at extracting intentions. Are there any styles of games or genres where you feel this capability is or could be particularly useful?
EL: I think the key isn't content, it's complexity. I've always craved as a player to be able to do what occurs to me, not select from a short provided list, and then have the game respond and also remember it all for subsequent interactions, not just a few predetermined choices. And not just dialog trees, but everything, like how in graphic adventures where you needed to block a door open but only one designated object succeeded despite the presence of other things that could also work as door stops. When we talk about truly free action for players, and rippling consequences, so much attention is paid to an explosion of content, but I think the challenge isn't in making the dots but in connecting them all in a vast web of full interconnectivity. Much of my career I've been a narrative designer and I have zero interest in having LLMs replace writers, or any other developers, and I don't think that's what will vault gameplay forward anyway. The real challenge isn't a content explosion, it's an interconnectivity explosion, and that's something an LLM is surprisingly well-equipped to solve.
CW: That's exactly what we're tackling at Atomic Dirt—a solution not to use LLMs to generate endless content, but to weave true interactivity into gameplay, making every choice matter in a way that's actually systemic and sustainable. If that sounds intriguing and you want to see it in action, Eric and I will see you at GDC and GTC this week.