The huge success of ChatGPT, which amassed 100 million users in two months and generated a buzz about generative artificial intelligence, has raised expectations. For several months, digital experts and the media speculated about the launch of GPT-4, the language processing model that was supposed to succeed GPT 3.5, on which ChatGPT’s chatbot is based. After several rumors, the artificial intelligence software was finally unveiled on Tuesday, March 14, by OpenAI, the company that has become a specialist in developing AI models capable of generating impressive texts from written questions or requests.
To tout its progress, OpenAI points out that GPT-4 scored 90% on the Bar examination (the test to become a lawyer in the United States0. GPT-3.5, on the other hand, did not score more than 10%, the company said. However, the expectations of the most enthusiastic were somewhat dampened. GPT-4 is indeed, as expected, “multimodal,” in that it is able to process images, in addition to text, in its queries.
But this feature was not yet operational in the version unveiled. And at this stage, unlike ChatGPT, GPT-4 is not accessible to the general public; for the moment it is limited to a few testers, in a paid version, with a waiting list, and to a few companies that wish to use it to develop services, through an Application Programming Interface (API). Finally, GPT-4 does not escape some of the shortcomings of chatbots, such as factual errors, although these have been somewhat reduced.
“In casual conversation, the distinction between GPT-3.5 and GPT-4 can be subtle,” OpenAI acknowledged in its presentation. “The difference becomes apparent when the complexity of the task reaches a certain threshold. GPT-4 is more reliable, creative, and capable of handling more nuanced instructions than GPT-3.5.” Queries submitted to the model by the user can now be up to 25,000 words long, allowing for the inclusion of entire documents, to obtain an analysis, comparison or summary, for example.
GPT-4 is better at answering questions that require “reasoning,” OpenAI said. For example, it can find a time slot for a meeting between three participants when given a paragraph describing their schedule. It is also said to perform better than its competitors in managing non-English languages.
We are interested in your experience using the site.
Less transparent communication policy
Reliability and security have been improved, according to OpenAI: GPT-4’s factually correct response rate is reportedly 40% higher than GPT-3.5, and the software 82% less likely to respond favorably to a prohibited query (because it is deemed dangerous or inappropriate , such as instructions for using a bomb or a racist joke).
You have 60.07% of this article left to read. The rest is for subscribers only.