OpenAI Unveils Next-Generation GPT-4 Language Model

OpenAI today unveiled GPT-4, the next-generation AI technology that's set to power the viral chatbot ChatGPT and Microsoft's search engine Bing.

The AI startup touted that the latest GPT iteration boasts a number of improvements over its predecessor, GPT-3.5, chief of which is that it's "more creative and collaborative than ever before". It's also multimodal, being capable of analysing text and image inputs, or a combination of both, including images interspersed with text. Image inputs, however, are still restricted to research previews and not publicly available.

But OpenAI did say that the difference between GPT-4 and GPT-3.5, which is currently powering ChatGPT, can be subtle in "a casual conversation". It's only when tasks reach a certain complexity threshold that the new system's more nuanced capabilities shine, such as performing better exams and benchmarks like the Uniform Bar Exam, LSAT, SAT Math and SAT Evidence-Based Reading & Writing.

OpenAI CEO Sam Altman said GPT-4 is "still flawed, still limited” yet it's also "more impressive on first use than it does after you spend more time with it". The language model still suffers from some of the same problems of its predecessor, including risks of "hallucinations". "Hallucinations" refer to the system's tendency to make up information, or exhibit behaviour that's out of the ordinary. GPT-4, according to OpenAI's research, is still “not fully reliable", though the chances of it suffering from "hallucinations" is reduced relative to previous models.

OpenAI has already partnered with a number of companies to roll out GPT-4 into their respective products and services, including Duolingo, Stripe and Khan Academy, as The Verge points out. The new system is also available to the public through ChatGPT Plus, the chatbot’s subscription service which costs US$20 a month. It’s also integrated in Microsoft’s Bing chatbot, and will be released as an API for developers who want to use it for their own projects.