Everything about LLMs is inefficient. They have their benefits but watching them reason over things that are painfully obvious, that they've literally investigated before (before a memory compaction), never take a step back aand be like 'this is going too slow let me look for a better way', etc. is painful.

It's tuned for the kinds of tasks where "just try" doesn't get good results.

A major complaint with AI code was that AIs struggle with complex codebases, don't respect existing conventions, reinvent functionality multiple times over, etc. So, newer high end AIs are tuned with the "explore/exploit" dial turned towards "explore".

You could probably get it to do things "quick and dirty" with prompting, but that, of course, requires prompting for it.