Mastodoner @jonny has decided that understanding the badnesses inside Claude Code is more valuable than his own sanity, and if you're interested in the writing of software and/or the future of writing software, you should follow him, so that the impending sacrifice of his aforementioned sanity will not have been in vain.
1/
@jonny There's been a theme to his discoveries, well exemplified by this post:
https://neuromatch.social/@jonny/116350918703314086
The LLM can write correct and coherent _lines_ of code. It can even write functions that appear to be okay in isolation.
What the LLM doesn't seem to be able to do is create any kind of coherent or comprehensible large-scale _architecture_ for a project as a whole.
2/
@jonny And this is kind of bonkers. Because architecture has always been the hard part, the important part, of making big serious software systems.
Bad architecture is implicated in so many of the bugs that bite us, the security holes that betray us, and the technical debt that makes important projects ossify.
Architecture can be hard to see at a glance, impossible to appreciate without direct experience. One of the challenges of building complicated projects is just getting the stakeholders to take architecture seriously.
And now we have built this code-emitting machine that has no concept of architecture. Somehow, the stakeholders of the world are drooling. Because it promises to make irrelevant the things the stakeholders never understood.
/fin