Development speed is not a bottleneck

I have very much started to re-evaluate whether I believe in this. I always thought something along the lines of "once you have solved it architecturally, typing it out is the least of your worries".

But with LLMs I'm not so sure. I feel like I can skip the effort of typing, which is still effort, despite years of coding. I feel like I actually did end up spending quite a lot of time doing trivial nonsense like figuring out syntax errors and version mismatches. With an LLM I can conserve more of my attention on the things that really matter, while the AI sorts out the tedious things.

This in turn means that I can test more things at the top architectural level. If I want to do an experiment, I don't feel a reluctance to actually do it, since I now don't need to concentrate on it, rather I'm just guiding the AI. I can even do multiple such explorations at once.

Absolutely, my experience too. I think the bleeding edge models are very good at “idea infill”.

Depending on your subject matter you might only need an idea or two per 100loc generated. So much of what I used to do turns out to be grunt work that was simply pattern matching on simple heuristics, but I can churn out 5-10 good ideas per hour it seems, so I’m definitely rate limited on coding.

Similar to your comment on architectural experiments, one thing I have been observing is that the critical path doesn’t go 10x faster, but by multiplexing small incidental ideas I can get a lot more done. Eg “it would be nice if we had a new set of integration tests that stub this API in some slightly tedious way, go build that”.