OpenAI has finally unveiled GPT-4, its next-generation large language model. Its last surprise hit, ChatGPT, was always going to be a hard act to follow, but the San Francisco–based company has made GPT-4 even bigger and better.
But OpenAI doesn’t say how big or why it’s better. GPT-4 is the company’s most mysterious release to date, marking a complete transition from a nonprofit lab to a commercial technology company.
“At this time, we cannot comment on that,” said OpenAI chief his scientist Ilya Sutskever when he spoke to the GPT-4 team in a video call an hour after the announcement. It’s pretty competitive there.”
Access to GPT-4 is available in limited text-only capacity for users who have signed up for the waitlist and subscribers to our premium paid chatGPT Plus.
GPT-4 is a large multimodal language model. This means it can handle both text and images. Give him a picture of what you have in the fridge and ask him what he can make, and GPT-4 will try to find a recipe using the ingredients in the picture.
“The continued improvement in many dimensions is remarkable,” said Oren Etzioni of the Allen Institute for Artificial Intelligence. “GPT-4 is now the standard against which all base models are evaluated.”
“A good multimodal model has been the holy grail of many large tech labs over the last few years,” says co-founder of his Hugging Face, the AI startup behind the large-scale open-source language model BLOOM. One Thomas Wolf said: “But it remains elusive.”
In theory, combining text and images could allow multimodal models to better understand the world. “We may be able to address traditional weaknesses in language models like spatial reasoning,” Wolff says.
It is not yet clear if this applies to he GPT-4 as well. OpenAI’s new model seems to outperform ChatGPT in some basic arguments, solving simple puzzles like grouping blocks of text into words that start with the same letter. My demo showed GPT-4. This is a summary of the announcement blurb from the OpenAI website with words starting with G.
“GPT-4, a breakthrough generational growth, holds higher esteem: guardrails, leadership and profit. In another demo, GPT-4 received a tax document, answered questions about it, and explained why.
Many other companies are waiting in line: “The costs to bootstrap a model of this scale is out of reach for most companies but the approach taken by OpenAI has made large language models very accessible to startups,” says Sheila Gulati, cofounder of the investment firm Tola Capital. “This will catalyze tremendous innovation on top of GPT-4.”
And yet large language models remain fundamentally flawed. GPT-4 can still generate biased, false, and hateful text; it can also still be hacked to bypass its guardrails. Though OpenAI has improved this technology, it has not fixed it by a long shot. The company claims that its safety testing has been sufficient for GPT-4 to be used in third-party apps. But it is also braced for surprises.
“Safety is not a binary thing; it is a process,” says Sutskever. “Things get complicated any time you reach a level of new capabilities. A lot of these capabilities are now quite well understood, but I’m sure that some will still be surprising.”
Even Sutskever suggests that going slower with releases might sometimes be preferable: “It would be highly desirable to end up in a world where companies come up with some kind of process that allows for slower releases of models with these completely unprecedented capabilities.”