Post

This is codex with gpt 5.4 (paid plan) on high reasoning: it just needed to remove the useViewer reference (it's just a react query) instead it left a useless useViewer call and removed unrelated code. You guys are joking right? This is the best LLMs have to offer?

sh03

@sh03.dev

17 Mar

Another example: told it to fix a conflicting warning in VSCode (medium reasoning) it simply replaced the entire VSCode configuration file (tbf while also fixing the issue)

sh03

@sh03.dev

Somewhere in the middle of my prompts it also created a git repo in one of my subprojects (monorepo setup) causing me to spend 10 minutes trying to understand why the git diff in VSCode and locally was different.

17/3/26, 7:45 am0 reposts0 likes

sh03

@sh03.dev

17 Mar

Another example is I asked it to update to the latest tsdown version (that generates .mjs files instead of js ). I had to spoon-feed it the other two places where it needed to update references (package.json and dockerfile).

sh03

@sh03.dev

17 Mar

I don't think it's comparable to a junior developer. When it's good it seems more knowledgeable than a senior developer without the lessons a senior developer learned (e.g. better to remove than to add). When it's bad it's worse than a junior developer: it's closer to an adversarial entity.