A developer has revealed that Lamplight, a self-built AI system, now completes about 95% of usage tasks locally. This figure comes from personal use rather than a large-scale test, but it reveals a clear shift: the value of local models is no longer merely that they “can run.” They are beginning to take on sustained, continuous, real-world work.
What this is
Lamplight is a custom-built local AI operating harness—a system that organizes models, tools, task scheduling, and interfaces into repeatable workflows. It consists of a MacBook Pro and a desktop computer equipped with an RTX 5090: the former coordinates tasks, while the latter handles most inference. The two devices work together through Git worktrees and an event-driven task system.
Underneath, it runs Pi-Bolt, custom extensions, local models, and Claude. In other words, this is not a standalone product sold to the mass market. It is the阶段性 result of two years of one developer's practice. The author is not currently releasing the complete system, but plans to extract reusable components gradually.
Industry view
The important change is that the yardstick for evaluating AI is shifting from the quality of a single answer to “whether it can become continuously embedded in personal work.” Local execution can reduce the exposure of sensitive data and lessen ongoing dependence on cloud services. The division of labor across multiple devices also shows that, beyond model capability, task orchestration and system design are now determining the actual user experience.
The risks are equally clear. The 95% figure comes from personal use and cannot be directly extrapolated to enterprises. Hardware expenses, system maintenance, and human review have not disappeared. The source code is not yet open, and there is no independent replication. Lamplight is more of a proof of concept than a mature product. The author still inspects all the code, confirming that local AI has not truly escaped human oversight at this stage.
Impact on regular people
For enterprise IT: Sensitive data, dedicated hardware, and local inference provide organizations with more granular deployment options. But cost calculations should include hardware, maintenance, and model replacement—not just the unit price of cloud API calls.
For individual professionals: People familiar with command-line tools and development environments will be more likely to gain access to this kind of controllable, customizable AI workflow. Nontechnical users will still struggle to reproduce it in the near term.
For the consumer market: Competition in the local AI experience may shift from chat interfaces to cross-device collaboration and long-term memory. Personal computer manufacturers will also place greater emphasis on model-running capabilities and service integration.