Dunno, LLMs writing code still feels like they memorized a bunch of open source code and vomited them out in worse condition.

It's not that impressive that Claude wrote a C compiler when GitHub has the code to a bunch of C compilers (some SOTA) just sitting there.

I'm using an LLM to write a compiler in my spare time (for fun) for a "new" language. It feels more like a magical search engine than coding assistant. It's great for bouncing ideas from, for searching the internet without the clutter of SEO optimized sites and ads, it's definitely been useful, just not that useful for code.

Like, I have used some generated code in a very low stakes project (my own Quickshell components) and while it kind of worked, eventually I refactored it myself into 1/3 of the lines it produced and had to squash some bugs.

It's probably good enough for the people who were gluing React components together but it still isn't on the level where I'd put any code it produces into production anywhere I care about.

That is my experience from a year ago but I no longer feel that way. I write a few instructions, guide an agent to create a plan, and rarely touch the code myself. If I don’t like something, I ask the agent to fix it.