Research note · 2026-10-06

How much can a smaller system do?

Capability is concentrating wherever compute is. That is a problem for anyone who cannot rent the largest cluster.

The question we want to answer is concrete. For a given task, what is the smallest or cheapest system that is good enough, and how do you know?

That requires measuring what the task needs instead of starting from the biggest available model. It also requires reporting cost with every result: what it takes to run, not only how well it scores.

Running where the data lives matters too. Local and on-device operation changes what is possible for privacy, for latency and for people with poor connectivity.

Smaller systems have ceilings. Part of the work is finding them and saying so plainly.

Related project: Lean