The original note, as written at the time
Problem solving intelligence seems to be the sweet spot between overwhelming computational ability and efficient energy conscious use of that computational ability to a given task. While overwhelming computational ability is important in its own right, it needs appendages that are able to temporarily forget the computational ability of the whole and just solve a given task with as little energy as possible without being plugged into the whole continuously.
Ongoing development
Problem solving intelligence seems to be the sweet spot between overwhelming computational ability and the efficient, energy-conscious use of that ability on a given task.
The distinction matters because we keep building the first and calling it the second. A system that can compute anything is impressive and, for most tasks, wasteful. It brings the whole apparatus to a problem that needs a small part of it.
The image I keep returning to is appendages. Overwhelming computational ability is real and worth having, but it needs parts of itself that can temporarily forget the capacity of the whole and just solve the given task with as little energy as possible, without being plugged into the whole continuously.
That is a description of local competence. A hand does not consult the brain to close around a cup. It has its own reflex, its own economy, and it works when the connection is poor or the brain is busy. The system is not weaker for it; it is faster and cheaper everywhere.
For machines this means the intelligence has to be distributable. A camera that must send every frame to a data centre is not intelligent, it is a wire. The same capability, small enough to sit in the device, forgetting the whole and solving the one thing, is the shape we actually need — and it is also the only shape that works where there is no signal, which is most of the world.
The research question is not how to make the small version as capable as the large one. It is how to decide, once, what can be safely forgotten, and then let the piece run. Getting that wrong is how you get a machine that is confident and blind.
The sweet spot is not a size. It is an arrangement: a large thing that knows what to delegate, and small things that know what they are for.