Here’s How OpenAI’s GPT-4 Is More Advanced Than Its Predecessor

Topline

OpenAI on Tuesday announced the launch of GPT-4, the latest version of its AI language model and a leap from the technology powering its popular ChatGPT chatbot service—here’s what you need to know:

OpenAI announced the launch of GPT-4 on Tuesday.

NurPhoto via Getty Images

Key Facts

At launch, GPT-4 will be accessible to users who sign up for OpenAI’s $20 a month subscription service, ChatGPT Plus, while free users of the popular chatbot will continue to receive responses generated by the older GPT-3.5 model.

OpenAI describes GPT-4 as more “reliable, creative, and able to handle much more nuanced instructions” compared to its predecessor.

One leap that GPT-4 makes over GPT-3.5 is its ability to parse images: In an example from OpenAI, the language model is shown a picture of cooking ingredients, asked what can be made with them, and it responds with multiple options.

GPT-4 is a step up from its predecessor when it comes to contextualizing and summarizing large bodies of text and can point out inaccuracies in a summary written by a person.

GPT-4 comfortably aces most standardized tests compared with its predecessor, scoring in the 93rd percentile for the SAT reading and writing tests and 99th percentile in the Uniform Bar Exam, up from GPT-3.5’s 87th and 10th percentile results.

While not perfect, GPT-4’s reasoning skills are also better able to parse puzzles and offer a more accurate response.

What To Watch For

GPT-4’s flagship feature—its ability to analyze and respond to image inputs—will not be available to users at launch amid concerns that it may be abused. According to the Washington Post, the feature is being held back as the company attempts to understand potential risks associated with it. OpenAI says it is working on “safeguards” to ensure that it cannot be used for facial recognition and surveillance of private individuals.

Tangent

GPT-4 is not immune from a problem affecting nearly all language learning models—hallucination. This happens when a language model generates completely false information without any warning, sometimes occurring in the middle of otherwise accurate text. OpenAI acknowledges this, noting GPT-4 “still is not fully reliable” and warns that precautions such as human review must be taken when using the language model in “high-stakes contexts.”

News Peg

Aside from ChatGPT Plus, GPT-4 has been integrated into a number of other products such as the language learning app Duolingo, educational platform Khan Academy and payments processor Stripe. The most prominent user of GPT-4 is Microsoft’s Bing search engine, which has been using a version of GPT-4 “customized for search” for several weeks. Duolingo has launched a new subscription tier called Duolingo Max which costs $29.99 a month and offers a GPT-4 powered tutor for English speaking users learning either Spanish or French.

Key Background

Interest in AI-powered chatbots has surged since OpenAI launched its ChatGPT service to the public in November last year. The technology, however, vaulted into the mainstream last month after Microsoft announced it had partnered with OpenAI to integrate its chatbot into the tech giant’s search engine, Bing. Bing’s implementation of OpenAI’s technology using real time search data drew excitement from the market with commentators suggesting that Microsoft’s AI-powered search engine might be the first serious threat to Google’s search dominance in years. Google has since rushed to announce Bard, its response to OpenAI and Microsoft’s chatbot, although the company is yet to give a release date for the wider public. Chinese search giant Baidu is expected to unveil its own AI chatbot Ernie on Thursday, although there are concerns it may be less impressive than OpenAI’s latest offering. Amid the excitement, some experts have warned that AI language models and chatbots powered by them still have major flaws, can easily present inaccurate information as facts and also be manipulated to misbehave.