Spaces:
Running
GPT5-Nano is trash compared to GPT5-Mini: here's why
Hi @yuntian-deng , I have been contributing data to WildChat for a long time, since the ChatGPT 3.5 days, and enjoy being able to see my prompts in the dataset. Thank you for making it open. I contribute a lot of high-quality, complex programming and technology focused questions. My contributions here were largely motivated by ChatGPT's superior skills and long responses which were much better than any other available model. The large context given by this HF space was a huge attraction which brought me here. I have used this space to solve some complex programming questions no other model could solve in a way that met all my requirements.
Unfortunately the recent switch to GPT5-nano has shown me just how different these ChatGPT models are. GPT5-nano is so awful compared to GPT5-mini that its outputs for programming related questions are altogether useless. It went from being able to write an entire C++ program flawlessly to doing a "programming sketch" that is both unhelpful and extremely short. I know GPT5-mini costs about four times as much as GPT5-nano, but the difference in models is remarkable, and I would ask you to reconsider the change. I can get better responses running GPT-OSS-120b now, and contributing to WildChat has altogether lost its edge.
I saw the WildChat 4.8M dataset, with all its automatic "Discord moderation" spam. Maybe we could come up with another way to reduce the spam without falling back on garbage models like GPT5-Nano. We could do a proof-of-work computation for example to make spam more costly. I am more than willing to spend the CPU power to use a much stronger GPT model.
Thank you for providing this and for publicly releasing the data. I contribute because (1) I know I can also benefit from the data when it is made available and (2) the model is strong and context length high enough for quality prompts.