LARI is designed around local execution and user control. Our research goal is to make increasingly capable AI practical on everyday desktops, laptops, and affordable computers—with reusable learning that does not require a paid cloud inference call for every use. Local control does not mean permanent disconnection: a system may access the internet to research or acquire information while its model and primary execution remain under the user’s control.

Efficiency, modest hardware needs, and lower dependence on expensive GPUs, APIs, and subscriptions are goals—not measured outcomes today. We have not established supported consumer hardware, comparative performance, memory requirements, or inference-cost savings.

We want capable AI to become accessible to low-income households and underserved communities, not limited to people who can afford premium computers, recurring subscriptions, or paid APIs. This technical ambition has a human purpose: changing who can own, control, and benefit from advanced AI. It is a long-term direction for our research, not a claim that LARI has already delivered that outcome.

LARI has an implemented shared task kernel, structured executable state, bounded procedures and coding workflows, and explicit candidate qualification and model-change controls. The current general-chat surface is limited; the architecture is not a broad native text generator or a proven general-purpose local AI. Lari Live publishes reviewed, scope-limited records and preserves negative results. It is a static observatory, not a live connection to the private runtime.