When calling API with the gpt-4 model, ask " Which model of OpenAI are you? Why the reply is gpt-3?

[midnight berry]

When the program calls the gpt-4 model, ask: Which model of OpenAI are you using now? Why is the model in the return value gpt-4-0613, but the reply content is gpt-3? The same is true for testing in playgroups.

{"created":1688431923,"usage":{"prompt_tokens":52,"completion_tokens":16,"total_tokens":68},"model":"gpt-4-0613","id":"chatcmpl-***xxLh","choices":[{"index":0,"message":{"role":"assistant","content":"I'm powered by gpt-3, the latest model from OpenAI."},"finish_reason":"stop"}],"object":"chat.completion"}

Playground

[floral wolf]

Because the API model is not self aware. The training data specifies that it is ChatGPT, which it knows is a model based on the GPT-3.5 architecture.

You can fix this by including a contrary statement in the system message, such as You are ChatGPT, a large language model by OpenAI based on the GPT-4 architecture.

[midnight berry]

Why does the official website online chatgpt choose the corresponding model to ask this question, and all answers are correct?

[floral wolf]

They tell it what model it is using in the system message.

As I showed you how to do.

[midnight berry]

Why do we do this? Shouldn't the facts be respected? Isn't this way of assigning values ​​to system messages to deceive yourself?

[floral wolf]

No...and that isn't how language models work. They are not self-aware. They don't know what they are unless you tell them what they are. This is also how chatgpt can know the time and date - via system messages.

[cinder tusk]

Pretty sure GPT4 still uses the same training data as GPT3, and that being the case means that it would most likely not even know itself exists yet. And in not the larger portion of the data it's trained on is the more well known GPT3.5 and would be it's most logical answer if it's the answer most commonly found in it's data sets.