Why Local Models Make Sense for a Game That Plays Itself

<p>Most of what I write here is about work: AIREP, the gym platform, Nova. This post is about GAIA, which isn't work. GAIA is a personal world-evolution game. You don't really play it, you watch it. A terraforming AI and her council of sub-intelligences shape a world over time, and the whole thing runs on local LLMs.</p><p>I want to explain why I think that setup fits, because I believe the reasons apply beyond one hobby project.</p><p><strong>Slow software changes what the model needs to do</strong></p><p>Most talk about choosing a model comes back to latency and quality. A user is waiting, so the answer has to be fast and good. That's the right lens for a chat interface or a business tool. If someone is waiting on a quote in an ERP, a slow or vague response costs you.</p><p>A passive game turns that around. Nobody is staring at a spinner. If the world moves forward while I'm doing something else, a slower generation step doesn't matter. What matters is that the world stays coherent, that it keeps going, and that running it doesn't turn into a bill I check nervously. In my view, that makes local models a good fit, not just a compromise. They're the more natural choice.</p><p><strong>Cost shapes design, even when you don't notice</strong></p><p>When every call to a hosted model costs something, you start designing around the meter without realising it. You batch more, cut prompts down, and hesitate before letting an agent think out loud or argue with another agent. For a simulation that's meant to run in the background, that pressure pulls in the wrong direction. A world that evolves needs room to be chatty, to explore, and to make a lot of small decisions that mostly don't matter.</p><p>Running locally removes the per-call cost. That doesn't make it free, since the hardware and setup are real costs. But the cost no longer grows with every message, and I think that changes what you're willing to try. A council of sub-intelligences that discuss things before the terraformer acts is a design you can afford to try when the next call doesn't cost anything extra.</p><p><strong>Owning the whole loop</strong></p><p>I care about owning my stack, and I've written before about treating AI as infrastructure. A personal project is the easiest place to actually do that. Nothing depends on an outside API staying the same, a model being retired, or pricing changing. The world is mine, the models are on hardware I control, and the only person relying on it is me.</p><p>That matters more for something long-running than for something you use once. A world-evolution game is meant to keep going. I'd rather its existence not depend on someone else's roadmap.</p><p><strong>Why bother with a project that isn't a product</strong></p><p>I'm bootstrapped and I have real goals on the business side, so it's fair to ask why I'd spend any attention on a game. My honest answer is that I think toy projects are where you learn things you'd be too careful to try in production.</p><p>A multi-agent setup where a lead intelligence works with a council is close to the patterns I use in serious systems. In GAIA, though, if a decision goes oddly the world just gets stranger. Nobody's invoice is wrong. You can watch how agents behave over long stretches, how they drift, and what keeps them consistent, without anything important riding on it. I think you develop intuition there that's hard to get when every mistake costs a customer.</p><p>Local models also force a useful kind of honesty. They're smaller than the frontier hosted models, so you can't count on raw capability to cover a weak structure. If the world holds together, I'd argue that's because the system around the models is doing its job: the roles, the handoffs, what each agent is allowed to know. That's the part I care about most in every agent system I build.</p><p><strong>The general point</strong></p><p>My view is that the question of which model to use has a quieter twin: how much does this task care about speed, and how much does it care about running indefinitely, cheaply, and under your control? For anything interactive and customer-facing, speed and quality usually win. For anything that runs in the background and is meant to keep going, I think local deserves to be the default, not the fallback.</p><p>GAIA is where I get to test that idea with nothing at stake. I expect the lessons to carry over well beyond the game.</p>