The 60fps Problem: Where Does the Model Actually Run?

The frame budget never moves

Your game has ~16.6 milliseconds per frame at 60fps. That budget doesn't care how cool your AI is. The moment a model call pushes past it, players feel it as a hitch, and a hitch reads as broken.

This is the constraint that shapes every other decision in a live build.

Options, roughly sorted by where they run

On-device. A small model running locally means no network, no cost, and privacy. But you're trading away model capability, and you're still fighting for CPU/GPU time with everything else your game does.

In the cloud. A bigger model, but now your game depends on a network round trip and a monthly bill. Latency becomes a design problem: the agent needs to think while the player keeps playing.

What this means for the build

Our working theory for Episode 001:

• Keep the reactive stuff (flavor, small decisions) on a local model.
• Push the big thinking (planning, memory) to something smarter, but off the critical path of the frame.
• Never block the render thread on an await.

We'll find out how wrong that is on stream.

The takeaway

The interesting question in AI × gaming isn't "can the model do it?" It's "what does the game have to give up so the model can do it?" Bring your worst latency stories to the Q&A.