ChatGPT: new version announced by OpenAI makes less mistakes – 01/30/2023 – Tech

New version of ChatGPT text generator robot should make fewer factual errors and better interpret mathematical problems. The information is contained in an announcement made this Monday (30) by the technology developer, OpenAI, in the Discord conversation application.

In the discussion on the OpenAI Discord channel, users of the text generator robot, however, complain that they are unable to experience the announced improvements.

“They perfected ChatGPT, breaking it,” said one user after failing to gain access.

THE Sheet was able to access the free version of ChatGPT at 4:10 pm this Monday, and the site informs that it is still from the last day 9.

The most recent information in the robot system is from 2021. When the report, for example, asked who the current president of Brazil is, ChatGPT replied “Jair Messias Bolsonaro”.

the access of Sheet The chat lasted only long enough for one question, which then displayed the following error message: “We are experiencing unusually high traffic. Please be patient as we work to scale our systems.”

OpenAI issued a statement at 4:26 pm with apologies for the errors on the site. At 4:48 pm, he said he had solved the problem. The update, however, was unable to access the update at 4:54 pm.

Researchers who have been testing the platform have been warning since last year that the robot makes inaccuracies when asked to do more complex searches. The algorithm has already invented scientific articles from mixtures of existing texts.

“It is pre-trained from a rich text bank of linguistic examples and, after training, it does not search any internal or external texts [para fazer checagem]”, says professor of statistics at USP Alexandre Patriota.

The text generator bot also makes basic math mistakes. On a second access, the Sheet asked the January 9 ChatGPT version how much is “(423-23) – (100+10)”. The algorithm responded, 260, while the correct result is 290.

The error is explained by the way ChatGPT works: its algorithm does not follow simple orders as in a cake recipe. What it does is deliver the most likely answer to a given sentence. This is done based on billions of texts processed during algorithm training, but there is always a possibility of error.