In defense of not understanding your codebase

The premise of understanding a codebase distracts from the actual issue in the debate about AI in software engineering, which is how well you have to understand the problems you're solving. I think most people who advocate for understanding the codebase actually mean to advocate for understanding the problem, they're just using the codebase as a concrete representation of it.

If ok for an individual to not understand the whole codebase, it is not ok if no one understands a whole portion of the codebase.

> LLMs... impedes the ordinary process of theory-building

As I have said on here before, I actually really enjoy LLMs for code understanding precisely because they are imperfect.

They're good enough to point you in the right direction, they might even see something you miss, but you can't trust anything they say.

This forces you to deeply understand what you're looking at, and to constantly re-evaluate your mental model against the code you're reading.

This works the same way as "writing forces clarity".

I could get to the same place on my own - I did this for many years before LLMs existed - but I feel when an LLM works well it gets you to the same place (or better) quicker.

(as idiomatic code generators I have mostly found LLMs to be poor, even Opus, but this comment isn't about new code generation)