If you were afraid that artificial intelligence would destroy humankind, it seems that you don’t have to worry about that anymore. Instead of getting smarter, ChatGPT may be going in the opposite direction.
Well, the newest research study from Stanford and Berkeley universities suggests that it’s getting dumber. The worst part of this story is the fact that even scientists themselves don’t know why this is happening.
The paper examines changes to OpenAI’s ChatGPT tool during a certain period of time. The slowing of progress has been spotted earlier, but what is most worrying is the performance of the ChatGPT 3.5 and ChatGPT 4 versions, which began to “vary a lot”.
The paper states that the problem was not only observed in general performance but also in advanced skills. Such as recognizing and understanding images, text, and programming code. The more testing was done in these fields, the less the ChatGPT performed. It seems that ChatGPT is getting dumber
In order to make the research representative, scientists from the universities gave ChatGPT a variety of tasks. For example: solving mathematical problems, asking sensitive questions, generating code, and visual recognition. In these tasks, even ChatGPT 4, i.e. the latest paid version, showed results below expectations.
In this case, from the 97.6 percent accuracy in recognizing prime numbers that were achieved in March, the performance fell to an incredible 2.4% in June. It has also been proven that ChatGPT produced more coding errors each new month compared to the previous one. Can we say that because in this case ChatGPT is getting dumber? I think so.
Meanwhile, the OpenAI developer forum is discussing this issue. The degradation of the ChatGPT 4 version is obvious, subscribers pointed out. Peter Welinder, one of the main product managers of the OpenAI company, stated that they “didn’t make ChatGPT dumber”. Every new version is smarter, he added. Well, after the publication of this paper from two prestigious universities, maybe he changes his opinion.
Of course, this situation raises many new questions. How reliable the model of artificial intelligence in facilitating our daily tasks is, and to what extent can we trust that it will do the task correctly?
On the other hand, it is likely that other research will appear soon that may suggest the opposite, but the fact is that not everything is going according to the OpenAi company’s plan. We also need to see how other similar AI models will perform, especially Google Bard.