Treating models as processes in an AI-native operating system
If a model can be scheduled, it can be a citizen of the kernel, not a guest of it — the OS primitives that manage processes (priority, preemption, resource limits) apply almost directly to managing models.
Every AI assistant today runs as an application sitting on top of a conventional OS — it asks for resources the same way a browser tab does, and the OS has no concept of what it's actually running.
Reframing a model as a schedulable unit — with its own priority, memory ceiling, and preemption rules — starts to look less like a metaphor and more like a straightforward port of process-management primitives that already exist.
The payoff shows up at the interaction layer: instead of opening an app and finding the right button, the system takes a spoken or typed intent and routes it to whichever process (model or otherwise) can act on it, the way a shell routes a command to a binary.
Unresolved: security boundaries. A process with filesystem access is a known, fenced risk. A model with the same access, reasoning about what to touch, is a different shape of problem entirely.