I doubt the March 14 model is any different, since it's versioned and I don't see why OpenAI would want to change it behind the scenes. Also, companies are eva

I doubt the March 14 model is any different, since it's versioned and I don't see why OpenAI would want to change it behind the scenes. Also, companies are evaluating GPT-4 to determine whether they want to pay for it, so OpenAI has an strong incentive to not downgrade at least the API. I believe the May 5 model is different, at least in the chat interface, because it's fine-tuned to detect jailbreaks and the temperature/other hyper-parameters may have changed. And I can imagine this fine-tuning making the model less creative and worse at solving analytical tasks. Personally I haven't noticed any change, except in my own awareness. Sometimes GPT4 gets very hard prompts right, and sometimes it gets simple problems wrong. So it's not hard to see how people can form biased opinions from selective attention or just luck.