Most attention has followed the largest models, and most is moving in the opposite direction.
A model small enough to run on a phone answers in , works without a connection and sends nothing anywhere.
Those three properties matter more than for a large class of tasks.
Transcribing a doctor's notes, sorting photographs and translating a menu do not require a system.
They require a system that works in a basement with no signal.
is the technique that made this possible.
A large model generates training material for a small one, which then reproduces much of its behaviour at a fraction of the size.
The legal status of that practice is unsettled, since it uses one company's output to build a competitor.
prohibit it and are difficult to enforce across borders.
Privacy is the property that changes institutional decisions most.
A hospital that can process recordings without them leaving the building faces a different legal question from one that cannot.
Several health systems have adopted local models for exactly this reason rather than for cost.
Energy is the argument that appears last and may end up first.
A answered on a device uses a tiny fraction of the electricity of the same in a data centre.
Multiplied by the number of queries, that difference decides more about the sector's footprint than any efficiency gain inside the centre.