Building AICore: An Adaptive Computer Control Layer in Rust

Personal computers are increasingly complex systems with multiple processors, memory types, and power states. Workloads have also become more varied, from software compilation to gaming, AI inference, and background services. This diversity leads to the question of whether a control layer could adapt resources based on the current workload.

This inquiry gave birth to AICore, an experimental AI-assisted computer control layer for Windows and Linux written primarily in Rust. AICore sits between applications and OS/hardware interfaces, not to provide performance presets but to explore a continuous control model. It collects platform telemetry, workload context, and application task information, possibly incorporating AI inference to generate a resource policy.

This policy is then validated against the machine's capabilities, executed, verified, and used to learn from the process. The control layer exists to bridge the gap between various components managing different aspects of computing resources. Applications understand their own workloads, operating systems manage processes, hardware runtimes handle specific accelerators, drivers deal with individual devices, and AI runtimes manage model execution.

However, none of these components have a comprehensive view of the user's computational intent. AICore seeks to fill this gap by connecting these signals through a separate control plane. AI inference is not treated as an authority in AICore. Instead, an AI component makes a proposal, which must then be validated by the control system.

This hierarchical structure ensures that AI is not directly controlling the machine. AICore includes three levels of application integration, ranging from no modification (L0) to native task runtime support (L2). Applications can communicate their computational intent without directly controlling hardware policy. However, AICore is an engineering prototype, not a claim that it can universally control every PC's CPU, GPU, and NPU resources.

The system is designed to handle failures gracefully, with concepts like capability validation, post-condition verification, rollback, safe mode, degraded operation, and policy expiration. A successful API call doesn't automatically mean the intended hardware state was achieved; the control loop distinguishes between requested state, execution result, and observed state.

AICore V4 is currently a private engineering prototype with a multi-crate Rust workspace covering various components like the control loop, policy system, platform adapters, AI runtime, task runtime, IPC, plugins, memory, simulation, CLI tooling, and a local dashboard. The next major phase will involve demonstrating that AICore improves performance or energy efficiency across representative systems.