Why does the 'drift' problem that AI's intelligence suddenly declines occur?



It has been reported that chat AI such as OpenAI's ChatGPT can

pass the medical license exam and solve mathematics with amazing accuracy. However, it has been reported that between March and June 2023, ChatGPT's accuracy in solving simple mathematics has declined sharply. This AI decline is called “ drift ”.

What is a 'AI drift' and why is it making ChatGPT dumber? | ZDNET
https://www.zdnet.com/article/what-is-a-ai-drift-and-why-is-it-making-chatgpt-dumber/



Why ChatGPT Is Getting Dumber at Basic Math - WSJ

https://www.wsj.com/articles/chatgpt-openai-math-artificial-intelligence-8aba83f0



Study Reveals ChatGPT's Struggles with Basic Math – Cryptopolitan

https://www.cryptopolitan.com/study-reveals-chatgpts-struggles/

According to a survey conducted by a research team at Stanford University in the United States, the research team divided OpenAI's large-scale language models 'GPT-3.5' and 'GPT-4' in March and June 2023. We gave four tasks: ``mathematical problem'', ``code generation'', ``visual reasoning'', and ``sensitive question'', and analyzed the speed and accuracy of their answers. As a result of the survey, it was found that the accuracy of GPT-4 answers to simple math problems such as ``Is 17077 a prime number?'' plummeted from 97.6% to 2.4% from March to June 2023.

As a result of research that ChatGPT's intelligence is declining rapidly, the correct answer rate of simple math problems deteriorated from 98% to 2% in several months-GIGAZINE



The decline in AI intelligence shown by the research team is called 'drift'. According to researchers at Stanford University and the University of California, Berkeley, ``The drift problem in AI is a problem in which trying to improve one part of a very complex AI model degrades the performance of other parts of the model.'' .

``If you fine-tune your AI model to strengthen it in a particular direction, you run the risk of falling back in another area,'' said James Zou, a computer science researcher at Stanford University. It's very difficult to keep improving.'



Zou also said, ``We suspected that AI models such as GPT-4 would have drift problems at some point, but we were very surprised that drift problems occurred so quickly. It was reported.

Overseas media Cryptopolitan speculates about the drift problem in AI, ``This problem may be intertwined with a rapidly growing trend called prompt engineering .'' According to Cryptopolitan, prompt engineering is the concept of a user creating a prompt to elicit a specific response from an AI, but 'the reduction in mathematical ability in GPT-4 was an inadvertent move taken to counter prompt engineering. may be the result of



OpenAI said, “When we release a new AI model, our top priority is to make the new model smarter overall. We are conducting various surveys and studies to confirm that it is connected to , but our evaluation method is not perfect, so we are constantly improving it. '

Mr. Zou said about the drift problem in AI, ``The important thing is not to abandon technology as intelligence declines, but to monitor AI more closely than ever before.'' The research team has revealed that AI models such as ChatGPT will continue to be systematically tested by asking thousands of questions and analyzing changes in performance over time.

in Software, Posted by log1r_ut