Neural MMO creator Joseph Suarez says OpenAI's Codex xhigh suggested a bubble sort for his high-performance C experience buffer

@jsuarez I mean your code is basically a bubble-fish so it kinda makes sense

@virgilenjoyer @jsuarez exact same thing that happened to Opus 1 month ago

@jsuarez Bubble sort is actually the best available sort algorithm if the list being sorted is relatively small. E.g. some chess engines use it to sort possible moves. So Codex might be right, depending on what you're sorting.

@jsuarez Codex got injected with retard juice a few days ago rip

@jsuarez Who knows if it was the right choice given the context, but a common theme I see is people over-indexing on Big O when inputs are bounded.

@jsuarez benchmarks don't capture everything. real-world usage can be a mess, even from top models.

@jsuarez what do the models know that we don't about bubble sort!?

@dubya_brian @jsuarez models need to do a better job investigating the potential input space to account for that. so much "wrong" optimization happening it just bloats the code and eats its context on future runs making it dumber and dumber and dumb

@jsuarez Sometimes bubble sort is the right answer. Although, that’s very rarely the case.

@jsuarez For me It started updating my Claude settings instead of its own agents.md

@jsuarez

@jsuarez Was it a find top k elements or a real bubble sort?

@jsuarez Every now and then LLMs just spit out the most egregiously inefficient code, it’s kinda funny

@jsuarez He’s just like me fr

@jsuarez What if you had no memory

@jsuarez yep, yep it did.

@jsuarez It could sense you are a bubble sort kind of person.

@jsuarez If its already almost sorted then bubble sort is pretty good

@jsuarez sounds like something woke

@jsuarez maybe codex xhigh made a bad assumption the buffer would be small.