How to improve something 1,000,000 X:

Humans understand complex things via layered abstractions. Many very accomplished people become an expert in one of these layers; you're either a great driver or you know how to tweak a fuel-injection system, but usually not both.

Each layer usually has a well developed set of tools & techniques, and a set of baseline assumptions about how they must connect to other layers. Whether the car is electric or ICE, it should have a "gas" pedal on the right, and a "brake" pedal in the middle. These well-defined boundaries are convenient and help reduce learning curves.

But, if you want to make a *big* improvement relative to the state of the art, the principle of diminishing returns becomes a formidable enemy. You can often pick up a factor of two, or even ten with a bit of effort. But if a really big leap forward was that easy to find, one of the experts would have noticed it already.

An alternative is to intentionally break the rules imposed by the abstraction boundaries. To define a narrower problem, that doesn't require the full generality of each boundary crossing. It's much easier to get 10X at each of six different problems (abstraction levels) than to get 10^6 at any one of them.

The downside is this approach tends to create "brittle" systems. They work astonishingly well, but only as long as whatever constraints you had to impose to puncture the abstractions hold.

#optimization #computation #systems #abstraction #thoughtfortheday