Today, OpenAI released GPT-4, its latest text-generating model, via its API.
All OpenAI API developers “with a history of successful payments” can access GPT-4 this afternoon. The company plans to open access to new developers by the end of the month, then raise availability limits “depending on compute availability.”
“Millions of developers have requested access to the GPT-4 API since March, and the range of innovative products leveraging GPT-4 is growing every day,” OpenAI wrote in a blog post. “We envision chat-based models supporting any use case.”
GPT-4 can generate text (including code) and accept image and text inputs, improving on its predecessor, GPT-3.5, which only accepted text. It performs at “human level” on various professional and academic benchmarks. GPT-4, like OpenAI’s previous GPT models, was trained using public web pages and licensed data.
OpenAI customers cannot yet use image-understanding. OpenAI is testing it with one partner, Be My Eyes. It hasn’t said when it’ll open it to more customers.
GPT-4, like all generative AI models, has flaws. It “hallucinates” facts and reasons incorrectly, sometimes confidently. It fails to learn from its mistakes, introducing security vulnerabilities into code it generates.
OpenAI plans to allow developers to fine-tune GPT-4 and GPT-3.5 Turbo, one of its other recent but less capable text-generating models (and one of the original models powering ChatGPT), with their own data. OpenAI expects it later this year.
Generative AI competition has intensified since GPT-4’s March reveal. Anthropic increased Claude’s context window from 9,000 to 100,000 tokens. (Context window refers to the text the model considers before generating additional text, while tokens represent raw text—e.g., “fantastic” would be split into “fan,” “tas,” and “tic.”)
GPT-4 was the previous context window champion at 32,000 tokens. Models with small context windows often “forget” even recent conversations, veering off topic.
Today, OpenAI released its DALL-E 2 and Whisper APIs, which are the company’s image-generating and speech-to-text models, respectively. To “optimize [its] compute capacity,” the company will deprecate API models. OpenAI has struggled to meet demand for its generative models in recent months due to ChatGPT’s popularity.
GPT-3 and its derivatives will be replaced by “base GPT-3” models on January 4, 2024, which should be more compute efficient. Developers using the old models must manually upgrade their integrations by January 4, and those who want to continue using fine-tuned old models must fine-tune replacements atop the new base GPT-3 models.
“We will support users who previously fine-tuned models to make this transition as smooth as possible,” OpenAI wrote. We will contact developers who recently used these older models in the coming weeks and provide more information once the new completion models are ready for early testing.