On Tuesday, OpenAI announced the long-awaited arrival of ChatGPT-4, the latest iteration of the company’s high-powered generative AI program. ChatGPT-4 is touted as possessing the ability to provide “safer and more useful responses,” per its official release statement, as well as the ability to accept both text and image inputs to parse for text responses. It is currently only available via a premium ChatGPT Plus subscription, or by signing up for waitlist access to its API. In a preview video available on the company’s website, developers also highlight its ability to supposedly both work with upwards of 25,000 words—around eight times more than GPT-3.5’s limit.
OpenAI says this version is stronger than its predecessor in a number of ways. Based on internal evaluations after six months of fine-tuning, OpenAI promises an 82 percent reduction in the likelihood of responding to “disallowed content,” as well as a 40 percent increase in its ability to produce factually true answers when compared to GPT-3.5. In support of these improvements, OpenAI writes in its blog post that GPT-4 scores in at least the 88th percentile and above on tests including LSAT, SAT Math, SAT Evidence-Based Read and Writing Exams, and the Uniform Bar Exam. It also earned a 5 on its AP Art History and Biology tests.
Despite its currently limited public access, OpenAI announced that it has already partnered with a number of other companies to integrate ChatGPT-4 into their products, including the language learning app Duolingo, as well as Stripe, Morgan Stanley, and the independent learning website, Khan Academy.
ChatGPT and similar programs like Google Bard and Meta’s LLaMA have dominated headlines in recent months, while also igniting debates regarding algorithmic biases, artistic license, and misinformation. Seemingly undeterred by these issues, Microsoft has invested an estimated $11 billion into OpenAI, and highly publicized ChatGPT’s integration within a revamped version of the Bing search engine.
[Related: The FTC has its eyes on AI scammers.]
Even with its new features, OpenAI CEO Sam Altman has repeatedly stressed that users should temper their expectations for what GPT-4 can and can’t accomplish. “People are begging to be disappointed, and they will be,” Altman said during a recent interview. In a Twitter thread announcing GPT-4, Altman also wrote, “[I]t is still flawed, still limited, and it still seems more impressive on first use than it does after you spend more time with it.”