GPT-4 selected, but receiving GPT-3.5 results

[sudden elm]

I have been consistently using GPT-4 via the OpenAI API, however, I recently encountered a bug where despite selecting GPT-4 in the list, the outputs are generated by GPT-3.5. This can be identified by the following:

1. Speed of the generated answer
2. Explicitly asking whether the model in use is GPT-3.5 or GPT-4
3. Asking for the response time, only to receive an answer indicative of GPT-4's performance

To resolve this, I have tried changing browsers, clearing cache, and sending API requests directly, but the issue persists, and I still receive responses from GPT-3.5.

Please investigate this bug and provide guidance on the appropriate steps to ensure that my requests are processed by GPT-4 instead of GPT-3.5.

Thank you.

chat_completion = openai.ChatCompletion.create(
    model="gpt-4",
    messages=[
        {"role": "user",
         "content": "Are you GPT-4?"}
    ]
)

print(chat_completion.choices[0].message.content)

Output: ```
I am an AI language model created by OpenAI, but I am GPT-3, not GPT-4. As of now, GPT-4 has not been released or announced by OpenAI.

I've been using the platform for two months now, but this problem has only just arisen.

[stone heath]

？

[somber owl]

You misunderstand how the model works. If you look at the actual returned object, you will see that it is GPT-4. If you tell it that it is based on the GPT-4 architecture, it will fix all the "indicators."

The reason it thinks it's GPT-3.5 is because it is ChatGPT and based on its data it does not know what GPT-4 is.

Just tell it that it is ChatGPT, a large language model built with GPT-4 architecture.

[sudden elm]

I agree that there may be disputes in this matter, but it is still impossible to deny the fact that something happens to the fourth model.

[somber owl]

Nothing is going wrong. That's just the way it is.

The reason that the GPT-4 model says it's GPT-4 on the web app is because it uses system messages to tell it so.

You need to use system messages if you want it to do the same.

[sudden elm]

Previously, the model could name today's date, in the chat, people are increasingly noticed by the so-called nerf 4th model, which is why it is more like 3.5

[somber owl]

There is no bug

The web app can do that because the web app has access to that data through sys messages. The API has never been able to do that. You are mistaken.

[sudden elm]

I have been using the model for two months and I noticed the difference in answers.

[somber owl]

It's likely coincidental.

My own app knows the date and time and all this because I use system messages properly.

[sudden elm]

I've always used API versions, and I think it could know the date. Well, my doubts are confirmed by the fact that using telegram bots, which are also based on the 4th version, they give answers that the neural network gave me before, I will try to explain more clearly:

[somber owl]

Okay. The outputs are not being generated by GPT-3.5.

[sudden elm]

Yes, in general, I was convinced that it is not just 3.5 model, but it seems like it nerfed

[somber owl]

The reason your outputs are different boils down to one of two reasons. Either A, your values are not the same as the telegram bots, or B, you are using a different system message. Or both.

Even something like changing the temperature by .1 can drastically change output.

[sudden elm]

Well, thanks for this discussion, I will try to find more significant evidence of my doubt about the behavior of the GPT-4 model. While they are not there, I admit that this may be some kind of deceptive effect.