Response that contradicts what's written in the GPT-4 technical report, possible proof of GPT-4's performance degradation

ChatGPT
1

RDT_20230824_172717443406924837338536
RDT_20230824_1727174434069248373385361766×1093 221 KB

image taken from gpt4 technical report

note that earlier model got over-censored, so it refused to fulfill the request

RDT_20230824_1727194144174072284644226
RDT_20230824_17271941441740722846442261370×474 33.3 KB

and here’s today’s gpt4

it refuses to give me the answer

I regenerated the response over 10 times, and yet, all got refused

2

I don’t think you understand what performance degradation means, because this is not that.

1 Like
3

I know it’s about censorship
but wasnt censorship the part of openai’s training?

4

I’m entirely sure not sure what you mean when you write,

OpenAI may have simply decided this particular family of prompts was now disallowed, in which case the model’s performance would be seen to have improved as it is maintaining its alignment.

But, regardless, this is not even demonstrative of a degradation in performance so it’s certainly not proof of such.