After months of rumors and speculation, OpenAI has announced GPT-4, a new AI language model. The company claims the model is “more creative and more collaborative than ever before” and “can solve complex problems with greater precision.” The model can parse both text and images, although it can only respond with text. OpenAI also warns that the system retains many of the problems of earlier language models, including the tendency to make up information (or “hallucinate”) and the ability to generate offensive and malicious text.
OpenAI says it is already partnering with a number of companies to integrate GPT-4 into their products, including Duolingo, Stripe, and Khan Academy. The new model is available to the general public through ChatGPT Plus, a $20 monthly OpenAI ChatGPT subscription, and is used for a chatbot in Microsoft Bing. It will also be available as a developer API (there is a waiting list that OpenAI says will start accepting users today).
The OpenAI blog post says that the difference between GPT-4 and its predecessor GPT-3.5 is “hard to catch” in casual conversation (GPT-3.5 is the model that ChatGPT runs on). OpenAI CEO Sam Altman tweeted that the GPT-4 is “still imperfect, still limited” but also “still feels more impressive the first time you use it than after you’ve spent more time with it.”
The company says the GPT-4 improvements are evident in system performance across a range of tests and benchmarks, including the Uniform Bar Exam, LSAT, SAT Math, and SAT Evidence-Based Reading & Writing exams. On the mentioned exams, GPT-4 scored 88th percentile and above, and the full list of exams and tests the AI passed can be seen here.
Speculation about GPT-4 and its capabilities has circulated in the past year, with many speculating that it would be a huge leap over previous systems. However, based on OpenAI’s announcement, the improvement is more iterative, as the company previously warned.
“People want to be disappointed and they will be,” Altman said in an interview about GPT-4 in January. “We don’t have real AGI, but that’s what’s expected of us.”
The road to GPT-4 has been a long one as OpenAI and AI language models in general slowly gained momentum over several years before moving into mass releases in recent months. The original research paper describing the GPT was published in 2018, with GPT-2 announced in 2019 and GPT-3 in 2020. These models are trained on huge text datasets, most of which come from the Internet, which is used to create statistical patterns. The patterns are then used to predict which word follows another. It is a relatively simple mechanism to describe, but the end result is flexible systems that can generate, summarize, and paraphrase texts, as well as perform other textual tasks such as translation or code generation.
Initially, OpenAI delayed the release of its GPT models for fear that they would be used for malicious purposes such as spam and misinformation. But in late 2022, the company launched ChatGPT, a conversational chatbot based on GPT-3.5 that anyone could access. The launch of ChatGPT caused a stir in the tech world, and Microsoft soon followed with their Bing chatbot (part of the Bing search engine), with Google struggling to catch up.