“That’s something that, you realize, we will’t really comment on at the moment,” said OpenAI’s chief scientist, Ilya Sutskever, once I spoke to the GPT-4 team in a video call an hour after the announcement. “It’s pretty competitive on the market.”

Access to GPT-4 will likely be available to users who enroll to the waitlist and for subscribers of the premium paid-for ChatGPT Plus in a limited, text-only capability.

GPT-4 is a multimodal large language model, which suggests it will probably reply to each text and pictures. Give it a photograph of the contents of your fridge and ask it what you can make, and GPT-4 will attempt to provide you with recipes that use the pictured ingredients.

“The continued improvements along many dimensions are remarkable,” says Oren Etzioni on the Allen Institute for AI. “GPT-4 is now the usual by which all foundation models will likely be evaluated.”

“A very good multimodal model has been the holy grail of many big tech labs for the past couple of years,” says Thomas Wolf, cofounder of Hugging Face, the AI startup behind the open-source large language model BLOOM. “However it has remained elusive.”

In theory, combining text and pictures could allow multimodal models to grasp the world higher. “It’d have the opportunity to tackle traditional weak points of language models, like spatial reasoning,” says Wolf.

It shouldn’t be yet clear if that’s true for GPT-4. OpenAI’s recent model appears to be higher at some basic reasoning than ChatGPT, solving easy puzzles corresponding to summarizing blocks of text in words that start with the identical letter. In my demo, I used to be shown GPT-4 summarizing the announcement blurb from OpenAI’s website using words that begin with g: “GPT-4, groundbreaking generational growth, gains greater grades. Guardrails, guidance, and gains garnered. Gigantic, groundbreaking, and globally gifted.” In one other demo, GPT-4 took in a document about taxes and answered questions on it, citing reasons for its responses.