White Frame Corner
White Frame Corner

A Stanford University study comparing the GPT-3.5 and GPT-4 versions of OpenAI's Chatbot revealed performance variations.

credit:gettyimages

White Frame Corner
White Frame Corner

The study examined four distinct tasks: solving mathematical equations, responding to personal questions, writing computer code and visual reasoning.

credit:gettyimages

White Frame Corner
White Frame Corner

Prime number recognition by GPT-4 decreased from 97.6% in March to 2.4% in just three months.

credit:gettyimages

White Frame Corner
White Frame Corner

For the same math problem, GPT-3.5 showed improvement from 7.4% accuracy in March to 86.8% accuracy in June.

credit:gettyimages

White Frame Corner
White Frame Corner

Researchers called the performance variations "drift."

credit:gettyimages

White Frame Corner
White Frame Corner

High-profile A.I. chatbot ChatGPT performed some tasks less well in June compared to its version in March.

credit:gettyimages

White Frame Corner
White Frame Corner

Math problems and visual reasoning were included in the tasks evaluated.

credit:gettyimages

White Frame Corner
White Frame Corner

Changes in accuracy that are noticeable highlight the adaptable nature of AI technologies in tackling challenging issues.

credit:gettyimages