Improving AIFPL test coverage and reflecting on architectural thinking in LLMs
Improving AIFPL test coverage
One of the more interesting problems of building a language is it has to be incredibly robust and very well tested. Designing a language for use by AIs is even harder because they won't file bug reports! A bad language design will, at best, result in the AI having to debug their way around problems. At worst, this will silently cause bad results to find their way into whatever the AI was trying to do.
My benchmark is to aim for 100% test coverage of all lines of code and branches. Even this isn't enough because we can still see "interesting" permutational problems, but at least with AIFPL we're dealing with a functional language so we can hope these problems are minimized.
What this does mean in scrutinizing code coverage reports from pytest. Today I've been finding all those places Claude didn't think to remove during rework, or just added because "I might need this later". This seems to be a problem with quite a few LLMs - they don't understand YAGNI at a fundamental level.
LLMs really don't think architecturally
I've spent several hours cleaning up the AIFPL code and finding subtle weird things that just make everything a little more tricky to work with. It's not dissimilar to reviewing the code of junior engineers who haven't yet learned to think about a bigger architectural picture.
I found a few weird almost-duplicate code paths, and in several cases it was because the LLM had debugged its way out of a problem, but hadn't stopped to look if similar code existed elsewhere.
As I'd been in somewhat of a hurry to see what I could build in the last 10 days I'd taken my eye off this. The code worked just fine, it was just a bit clunky inside.
I could conjecture several reasons why they are trained to work the way they do, but chief among them will be a lack of maintaining a vision of how the architecture is supposed to look. Humans will learn some elements of this in a codebase with which they become familiar, but LLMs don't "learn". TBH, though, give a human a few weeks away from a codebase and they'll start to forget everything too :-)