So there's a common dichotomy in software engineering between the "big ball of mud" and "beautiful diamond" patterns. A big ball of mud is a software system that grows through accretion: it often has tangled flow control, data whales, lots of special-casing, lookup tables, and so on. It evolves, it isn't really designed. A beautiful diamond is a software system that is meticulously designed, consistent, oftentimes with tight interdependencies between how it works. Everything is efficient and easy to understand, but if you take one part out of it, it tends to fall apart. Common Lisp is a big ball of mud, Scheme is a beautiful diamond. C++ is a big ball of mud, K&R C is a beautiful diamond. Windows is a big ball of mud, BeOS is a beautiful diamond. Microsoft VSCode is a big ball of mud, Turbo Pascal is a beautiful diamond.

The big ball of mud structure is often taken to be an antipattern: at least, when I Googled this, Gemini told me it's a "haphazard, tangled anti-pattern" and gave me some advice about how to refactor such a software system into a beautiful diamond. But nearly all software systems that people actually use and maintain are big balls of mud. This is because the big ball of mud tolerates partial failure extremely well. It tolerates changing requirements very well. It tolerates parallel development very well. You can change part of it without throwing away the whole thing.

AI is a beautiful diamond. For the most part, if an AI model malfunctions, you have to blow it away and retrain it. Or I guess you can build layers to fix up the output case-by-case, but then you're really building a big ball of mud around the beautiful diamond, and obscuring it in the process.

That's a little worrisome when it comes to durability. Living, enduring systems usually end up becoming a patchwork of mud. Your ability to do this with LLMs are basically limited to LoRa patches and post-processing steps.