Lately, I've noticed issues with ChatGPT generating code that is nearly unusable, often leaving massive gaps in the code and prompting me to fill them up. Have you experienced similar outputs?
This is never reflected in the benchmarks and everyone uses them as imperial universal truths. ChatGPT once reached a certain quality but is scaling down on first requests to save compute power.
There have been several recent HN posts to the same effect. Here's a recent one I participated in: <a href="https://news.ycombinator.com/item?id=38602025">https://news.ycombinator.com/item?id=38602025</a><p>It's not limited to ChatGPT either and seems more pervasive. Not sure what's happening
Data point: it works well here (on a Plus subscription using GPT-4) with Python and JavaScript and Go.<p>I do use highly-tailored starting prompts saved in a GPTs.<p>It's definitely gotten a lot slower.