If you haven’t seen the brilliant indie comedy Computer Chess, it might be worth a watch as an eccentric view of 1980s gaming nerds, and a testament how far gaming has come in the subsequent 44 years. To be fair traditional rules based “AI” in games has been the mainstay of single-player experiences for a long time. But Google’s GameNGen scalable diffusion paper from last week, and now a new infinite Mario engine trained on a consumer GPU this week have created huge interest in ‘AI as game engine’. These projects demonstrate AI’s ability to essentially hallucinate environments and mechanics without code, learning all of the necessary complex physics and interactions in virtual space. While still facing performance limitations – MarioVGG takes six seconds to generate half a second of gameplay – these developments hint at a future where game creation could be as simple as describing your vision to an AI.
Current generation AI’s applications are still in their infancy, with long gaming development cycles, cost constraints, and high performance demands. But the long-term uses are diverse, from enhancing graphics and procedural generation to improving NPCs (non-playing characters), analysing and personalising player experiences and the increasing the speed of development (this Minecraft clone was created entirely via prompts using Vercel’s v0). Nvidia’s Avatar Cloud Engine (ACE) is set to bring a new generation of intelligent, personality-rich NPCs to games, potentially revitalising experiences and attracting new players. But the real game-changer might be when these NPCs cease to be “non-player” at all.
Robert Yang, CEO of Altera, posted on X this week, “As AI agents become an integral part of our human civilization, they must effectively collaborate with each other and the rest of us.” This collaboration is already taking shape in projects like Altera’s simulation of over 1,000 AI agents in the game Minecraft, forming their own societies complete with economies and governments.
The methodologies developed for these gaming AIs have broad potential. GameNGen’s approach could benefit fields like urban planning, scientific simulations, and social science research. Altera’s work in games could mean vast agent simulations that model complex real-world dynamics.
The diffusion models being used in many images, video and now game generation innovations could find applications in diverse fields, creating everything from protein structures to complex mathematical concepts.
Takeaways: The fusion of LLMs, diffusion models and gaming is opening up all kinds of possibilities for creators. Forward-thinking business should be taking a closer look at this evolution of agent and engine simulation and exploring how these technologies could enhance their products, services, or research. At ExoBrain we’re helping clients to make agent-based simulations a useful reality. Of course, AI games can also be fun (and historic). Why not try 80s inspired AI Dungeon, an early demonstration of GPT-2’s power, and a surprisingly immersive experience for a text adventure.