GPT-4o is really oddly hit or miss for code. Sometimes it outperforms GPT-4 in quality by a fair amount, and other times it starts repeating itself. Duplicatin

GPT-4o is really oddly hit or miss for code. Sometimes it outperforms GPT-4 in quality by a fair amount, and other times it starts repeating itself. Duplicating function definitions, even misremembering what things are named. It seems to have to do with length. If the output exceeds a few thousand tokens, it seems to experience some pretty bad failure modes.