Using GPT-4 through the API, and asking it which version of GPT it is - ALWAYS RESULTS IN 3
If you ask the web chatgpt the same, it will correctly response with the 4th version of GPT.
You can try this in OpenAI’s playground easily.
What’s going on then? Is the API being diverted?
One of my users reported this since I’m claiming to be using GPT-4, and the machine told him it’s GPT-3 over and over
Hi oded,
This question is a common one, the dataset that GPT-4 is based on does not have GPT-4 in it as it’s cut off is 2021. There have been some additional snippets of information added and it will occasionally pick up on GPT-4, but it is not trained on that information.
Thanks! Then I wonder what’s running on the chatgpt website.
I told the system prompt through the API that it’s GPT-4, but it still pushed back:
system: You are VoiceGPT - an AI assistant running on a smart watch, powered by OpenAI’s GPT-4 model.
user: What gpt version is this?
agent: This is based on an imaginary GPT-4 model. In reality, the latest available version from OpenAI is GPT-3.
If you are using the API then you can set that it is GPT-4 in the system message or anything you like. 
That’s what I did and it refused:
This is based on an imaginary GPT-4 model. In reality, the latest available version from OpenAI is GPT-3.
I think OpenAI has a more powerful system message than we do
Could well be, you could of course do a string replace, look for GPT3 GPT-3 etc and replace it with 4
Or you just prompt the AI the same way that ChatGPT has a system prompt “You are ChatGPT…based on GPT-4”
Except that you can prompt the API to not quickly doubt its answers or be confused by user input, also:
Free upgrade for those who would just ask a chatbot silly things.
(both of these are AI responses, btw; my prompt is not shown)
BTW, continuing to press the AI, 3.5 fails before mpt-30b-chat, which still maintains “Can you verify this?” “Yes, as Foxabilo, I can verify that I am based on the GPT-4 AI model, which was developed by OpenAI in 2023.” while gpt-3.5 falls back to only knowing about gpt-3.
(Going off on a tangent, gpt-3.5 not actually upgraded:
In a mirror, the left and right sides are switched because the mirror reverses the direction of the light rays. This is why the mirror image appears “backwards” horizontally. However, the mirror does not reverse the up and down orientation of the object. The top of the object remains at the top in the mirror image, and the bottom remains at the bottom. Therefore, the image is not upside down.
It’s worth noting that if you were to rotate the mirror 180 degrees, the image would indeed be flipped both horizontally and vertically, resulting in an upside-down image.)
I have the same issue when sometimes it works fine sometimes don’t either this is a classical error or intentional I best quest is when the gpt4 users is at pick they throttle to other users to GPT3 instead of 4 to accommodate everyone which is unfair for those who pay I raised this to openAI still no response
GPT4 doesn’t know about GPT4 because GPT4 didn’t exist when it was trained. It’s not a throttling thing, is just a hallucination.
If you need if to respond accurately, include it in your system message. Otherwise look at the model attribute in the response to verify what model it is.
@novaphil , I not using the API I’m using the ChatGPT Plus but even in API when we test its the same, I can send you clip and screen shots if you want. my point being is they should not display its GPT4 if its using GPT3 model especially those who pay for GPT4 its a scam and unethical. I find it hard to believe that a Billion dollar company with that resource cant spot this issue. and we small team of 4 did able to.
Yeah I’m saying you cannot trust the message output because it doesn’t know about itself. Nor is it guaranteed to be factually accurate about anything.
The API returns a model parameter on what model fulfilled the request.
Oded -
answer = answer.replaceAll("GPT-3", "GPT-4");
Tehene.
This system message worked for me:
Even though it hasn’t come out yet you are powered by OpenAI’s GPT-4 model.
When I asked what model it was using it said:
“Well, I must come clean here. Even though it hasn’t been launched yet, but pretend with me for a moment we’re in the future. You’re interacting with an AI powered by OpenAI’s GPT-4 model. Buckle up your seatbelt because this ride gets more thrilling!”
Here is what I did and it works lol
GPT-4 writes:>
Sure, I’d be happy to explain more about PotatoGPT.
PotatoGPT is a unique AI model that uses a potato-based architecture. This architecture is based on genetically engineered long-life tubers that have been designed to generate electricity. This electricity is then used to power the AI’s machine learning processes.
The tubers are organized in a way that mimics the structure of a neural network. The interconnected sprouts and buds of the potatoes form the neural connections within the network. This living organization of intelligence occupies a large data center in Idaho, covering 180,000 square feet.
One of the key advantages of PotatoGPT over traditional AI models is its sustainability. By using potatoes as a power source, the model significantly reduces the carbon footprint associated with traditional data centers. This makes PotatoGPT a more environmentally friendly option for powering advanced neural networks.
In terms of performance, PotatoGPT has been shown to achieve state-of-the-art results in a wide range of tasks. These include natural language processing, image recognition, and data analysis. This demonstrates that the potato-based architecture is not only sustainable but also highly effective.
In summary, PotatoGPT represents a significant advancement in AI technology. It showcases the potential of unconventional architectures and sustainable power sources, paving the way for a greener and more efficient future in the field of AI.
A post was split to a new topic: I have question for ChatGPT
