GPT-4 is here! OpenAI's newest language model

Announcing GPT-4, a large multimodal model, with our best-ever results on capabilities and alignment: GPT-4

If you are a ChatGPT Plus subscriber, you can access GPT-4 right now! If you want access via the API, you can join the waitlist: GPT-4 API waitlist

We are also giving priority API access to those who contribute evals via: GitHub - openai/evals

Last, check out a live developer preview of GPT-4: GPT-4 Developer Livestream - YouTube

We can’t wait to see what you think and build!

Hello!

Im very excited for gpt4, i signed up with chatgpt+ and when i switch to the gpt4 model and ask it what it is it tells me its gpt3 and that gpt4 does not exist and it knows nothing past 2021

This is correct, we are aware it is doing that, see here for details on the model: OpenAI API

Nice! I’m hearing some devs already have API access.

Fingers crossed. I’ve got a good D&D demo up…

Hi, I was really hoping that GPT-4 would be compatible with the v1/completions endpoint (using text-davinci-003). I am not building a chat bot. Many other users are also not building chat bots, and there is much confusion in the community forum about this issue. Having to use v1/chat/completions and structure the inputs around 3 different roles, as if I was building a chat bot, in order to use OpenAI’s newest model, is frustrating. Can you give the community some insight as to whether GPT-4 will become available for use with the v1/completions endpoint any time soon? For people not building chat bots, the necessary changes to the code and parameters required for the v1/chat/completions endpoint take quite a bit of work and seem illogical. Obviously, we non-chat folks would much prefer not to have to make those changes. Thanks for you help.

Totally fair, I agree that it would be cool to have v1/completions, best thing to do would be to build around the existing chat completions and hopefully lean into the area where it can be more helpful to use the new format.

It’s so helpful with prompt injection, imho. Not too difficult to convert old prompts to ChatML.

I can’t wait for ChatML to be expanded upon, personally.

Hi. I was under the assumption that the new GPT-4 backed chatgpt+ would be capable of accessing links I give to it and using the information there as context.

However, I found this not to be true. For example:

HI @mo0nman

Why would you have this assumption?

got-4 is a pre-trained language model like all other OpenAI LLMs and none of these models have access to the Internet.

:slight_smile:

Hi @ruby_coder

My assumption was based on the following example shown at GPT-4

Agree @mo0nman that the example they gave in the announcement is confusing. I’m guessing they pasted the whole thing in the prompt and chose to abbreviate it that way. Confusing nonetheless.

What it actually means is that it’s capable of handling larger text unlike GPT-3. So if you copy the body of text from Wikipedia, it summarizes it well.

Ah I see. Thank you for the clarification @curt.kennedy and @smuzani. The ambiguity in the announcement is unfortunate and will hopefully be made clearer in the future.

I figured I would be able to give it scientific articles to summarize since it said it allowed for use cases such as “document search and analysis”.

after going through the description i am very much looking forward to it coming to developers to build integrated apps

That’s meant to showcase its capabilities in really long text.

GPT-4 is amazing and should have been the default mode for ChatGPT. Thanks for the good work.

Unfortunately it seems only few will be able to afford gpt4 :frowning: So, lets see what Google&co comes up with in near future.
cheers