This paper presents LiberaGPT as a local-first, bounded AI architecture for iPhone, positioned not as a general chatbot but as a mobile cognitive interface optimized for privacy, accountable behavior, and useful reasoning within handset constraints. Its core thesis is that modern iPhone hardware can support serious on-device inference when the system is designed around model specialization, thermal awareness, explicit memory management, and strict platform-native execution rather than cloud dependency. It describes a layered runtime in which compact quantized language models run entirely on-device through Apple’s stack, with inference adapting in real time to thermal state, memory pressure, and workload intensity. The paper argues that practical mobile AI should be built around multiple specialized small models rather than one oversized generalist, with the user selecting the model best matched to the task. A second major pillar is silicon alignment: the paper frames LiberaGPT as intentionally composed around the iPhone’s heterogeneous compute architecture, using Core ML, the Neural Engine, GPU, CPU, and unified memory as the execution substrate rather than fighting the platform with custom infrastructure. In that sense, the white paper is as much a systems architecture argument as an AI product description. The final layer is governance and trust. The paper emphasizes that user data remains local, that deletion is immediate and irreversible, that telemetry is absent, and that safety is enforced through concrete control points such as prompt-ordering discipline, local storage boundaries, and platform-native file protection. The overall positioning is: high-agency personal AI, constrained by device physics, strengthened by privacy, and made credible through operational transparency.
Stephen John Pereira (Thu,) studied this question.