Hierarchical Task Network Planning might finally be viable in the world of LLMs.

This was a technique I learned in my AI class in college in 2007.

You take a high-level task and break it down into smaller tasks, recursively, until they get small enough to do mechanistically.

LLMs could both help with breaking down the tasks, and with generating mechanistic code for the leaves as soon as the task is small enough to be unambiguous.

Approaches that had combinatorial mechanistic rules weren't viable in the past.

You had to go into fractal details to get to a level where the computer could handle it.

That blows up combinatorially with fractal complexity.

But if you can change the level you have to lower it to "a level where any competent LLM could robustly write mechanistic code on demand for it", then it becomes possible.

A linear reduction in required lowering leads to a super-linear reduction of effort.