Local and constrained inference
Running capable models on hardware people already own — treating efficiency as a first-class research constraint rather than an optimization afterthought.
The constraint is the subject
Most progress in machine intelligence has assumed abundant computation: clusters, datacenters, and energy budgets measured in megawatts. But the majority of computation in the world happens on devices people already own — laptops, phones, machines in closets. The gap between what frontier systems require and what everyday hardware provides is treated as an engineering nuisance. We treat it as the research subject.
The interesting question is not how to make constrained hardware behave like a datacenter. It is what systems become possible when inference is designed for scarcity from the start: smaller working sets, quantization as a native representation rather than a lossy afterthought, and scheduling that treats attention and memory as competing budgets.
Questions we are tracking
Where are the real quality cliffs when models are compressed — the points where capability drops non-linearly rather than gracefully? Can retrieval, caching, and speculative execution substitute for raw parameter count? What does an operating system look like when a local model is a resident system service rather than an application?
We are less interested in benchmark scores than in the shape of the trade-off surface: what a system gives up, at exactly what cost, and what it gets back in independence, latency, and privacy.
Why it matters to the lab
Local inference changes the economics of intelligence. When inference moves to the edge, the marginal cost of a query approaches the cost of electricity, and software built on it inherits a different cost structure than software renting compute by the token.
Any product the lab builds on this research owns its full stack — from the quantization decisions to the scheduling policy. That ownership is the point.