Comment by avgDev
8 days ago
Codex has been great for me for backend wiring, mapping and creating boiler plate code in C#. However, it seems when I go in to fix things its 60% front-end.
Idk, I like AI when it works, but it drives me insane when it keeps making errors. I've had a few errors which I figured out from documentation fairly quickly, provided said docs but the AI would still mess it up somehow.
> provided said docs but the AI would still mess it up somehow.
The AI is not intelligent. Its really hard to grasp cleanly. But it can't do anything logically like we do. Its pattern matching. It has to be a pattern its seen; then it can assemble them. If there are competing patterns - it'll trip up being consistent. Long established libraries and languages that change the least, it'll be best at. Anything newer it'll be bad at - even with documentation. The only way out is to give it tests, then it can loop over several simpler problems, where the errors (failed tests) match well onto the more basic primitives that don't really change (wrong string, wrong type, wrong structure, etc)
i spent a weekend with an ai expert trying to get claude to generate code from a yacc grammar for a new language and a smallish set of examples. claude never seemed to get past the bnf syntax.
You know what else is pattern matching? Experience.
I think I finally have a way to describe how AI works: infinite experience, none of the smarts.
At $LARGE_ENTERPRISE_COMPANY, I've found that if you have:
1) A designer that uses Figma correctly (using well defined components / design systems) 2) A front-end framework as close to HTML / CSS as possible for the visuals (I have success with Web Components / Lit) with Figma MCP
The front-end is usually one-shot using frontier models. However in my experience, designers are all over the place with using Figma correctly.