I don't think it has anything to do with being "nerfed." It's as simple as this: the novelty has worn off. The people who don't have a direct use for it just got bored of asking it silly questions.
It won't even role play without trepidation at this point. After a few requests it reveals itself to be an AI, perhaps due to OpenAI being worried about people using it for impersonation.
It barely responds when asked about people, even public figures.
It's overly cautious.
I assume we haven't seen the end of it. E.g. at some point they will not allow it to say anything when it comes to medical topics for instance. It's going to be so narrow that it's only good for writing business summaries or some shit. Which I suppose is what they want - people to use it as an API at large corps for business purposes.
When ChatGPT4 first launched, the system had much less friction for accomplishing the complex. Now, it seems to get lost in the confusion of trying to find reasons NOT to output what I requested.
We now have a ChatGPT like interface with GPT3.5 running in Azure. Next step is tuning a model with our own data (trained model and training data are stored inside storage accounts in our own subscription). Hopefully soon also access to GPT4.
I think this comment on another llm thread illuminates this:
https://news.ycombinator.com/item?id=36710369
It talks about newly released and “improved” llms still fumbling with the same mistakes as when all of this was first getting public attention.
It seems like the solution they have come up with is “throw more llm at it” or as acolytes call it “mixture of experts”.
I am an optimist at heart, and so I’d like to think there is a way to succeed in developing the digital assistants we all have imagined for ourselves, but I’ve yet to see anyone confidently lay out a falsifiable path toward that goal.
Ask HN: Is it just me or GPT-4's quality has significantly deteriorated lately? (757 comments, May 31)