Letting the LLM Drive

I've been finding some interesting results in certain contexts by letting the LLM drive me. These contexts mostly relate to learning or exploring ideas. It cannot veer into creative work, where it is a Bad Idea.

First, I establish a direction or goal, something that requires effort and understanding from me.

Then I have the LLM create a series of modules for me to follow that are not entirely unlike tutorials.

They are nonlinear conversations or threads, with the LLM dictating to me through instructions frozen into static artifacts (Markdown files) offered to me.

The roles are flipped—the LLM is prompting the operator indirectly through a document it created.

Leverage is inverted. It's certainly not to the LLM's advantage, at least at first.

Over time, I think it may be, if you can extend the idea of "advantage" to things without constancy.

It's strange to learn a language or approach or aspect of Linux during a time when the promise of agents whispers of complete automation. Certainly the days of having to architect work in code seem to have come to a close.

I increasingly see this period of humans reading PRs and auditing code itself as a weird transitional moment.

The way the architect thinks, or needs to think, is changing.

The mistake has been clear and identified—an LLM is probabilistic and not a compiler. It's not provable. But programs themselves, in some cases, are.

I think the benefit of flipping the script, of letting the LLM drive me, is in pushing me to learn to better specify what a program actually is, which gets me closer to knowing whether it should be built.