New version of ChatGPT outperforms 90% of people in the exam

OpenAI, the company behind the development of ChatGPT, announced this week that the new version of its artificial intelligence, GPT-4, has arrived. The researchers emphasize the ability to conduct tests, claiming that they have made significant progress in technology.

According to the developers behind GPT-4, the new version of the robot managed to get into the top 10% in a bar exam. A significant leap forward from the previous version, GPT-3.5, which was among the worst in similar exams.

The AI ​​also scored 700 out of 800 on a standard math test. ChatGPT is not yet doing well in language and composition tests.

“It still has flaws, it’s still limited, and looks more impressive at first use than after spending more time with it,” said OpenAI CEO Sam Altman. But the developers are confident that GPT-4 will make fewer mistakes than its predecessors.

“We spent six months interactively lining up GPT-4 using lessons from ChatGPT as well as our competitor test program, and we’ve had our best results to date (although far from perfect),” the official text on the new technology says.

More progress, less transparency

OpenAI also says that GPT-4 is “hallucinating” much less than its predecessors. This happens when AI invents information and starts talking about completely false things as if it were real.

Despite being “open” even in its name, OpenAI has been somewhat closed off on the work that led to this new iteration of GPT. The company declined to comment on the size of the model used in the tutorial or details of the type.

“Due to both the competitive environment and security implications of large-scale models like the GPT-4, this report does not include further details on the architecture, hardware, training computation, dataset generation, training method, or the like,” the company said. .