Waiting for Newton

Imagine if, after Kepler discovered that planets move in ellipses, we decided that ellipse-following was the fundamental mechanism of planetary motion. We might have spent our time asking how a planet, with access only to local information, could know how to stay on the ellipse.

Newton started from a different point: assume the planet does not need to know about the ellipse at all. Instead, begin from local dynamics (force, mass, acceleration), and derive the principles from which the ellipse emerges.

I think we have made something like the first mistake with intelligence. We observe behavior that can be described as minimizing an objective, then assume the system must somehow be implementing that optimization, and ask how biological systems might approximate it locally.

But loss minimization may be our Kepler’s laws: a powerful mathematical description of what intelligence does, yet far from the mechanism that makes it happen.