GPT-4 (Generative Pre-trained Transformer 4) is a large language model with multimodal capabilities created by OpenAI.
architecture
OpenAI has not publicly revealed information about GPT-4's architecture, but it is rumored to be a mixture-of-experts model. Some versions of GPT-4 can input images, and some versions have at least a 32k token context window.
release
The first variant of GPT-4 to be released to the public was via Bing on February 7, 2023.
chatGPT-4 was officially released by OpenAI on March 14, 2023.
The GPT-4 base model is not publicly accessible but some researchers have been given access.
variants
base model
The GPT-4 base model (gpt-4-base on the API) is the inframodel of GPT-4, trained with pure self-supervised learning. So far, only a small number of researchers outside OpenAI have been granted access to it.
GPT-4 early
GPT-4 early is a version of GPT-4 referenced in several OpenAI and Microsoft publications. It is likely to be an earlier and separate RLHF training run than chatGPT-4. It likely to be related to the model that powers Bing, as its outputs exhibit Binglish artifacts. Evidence also suggests that this model is called Davinci 3 (DV3).
chatGPT-4
fanw-json-eval
fanw-json-eval is an undocumented model that was briefly accessible to some or all users on OpenAI's completion API, between October 23rd and 27th 2023. fanw-json-eval is likely to be a variant of ChatGPT-4, as it behaves nearly indistinguishably from ChatGPT-4 while simulating an assistant in the chatML format. However, it was accessed through the completion rather than chat API and exhibited odd default behaviors.
Deucalion
Cannot transclude hypha deucalion because it does not exist