Chat GPT is better than undergraduates, according to researchers at UCLA. They found that GPT-3 correctly solved 80% of reasoning problems compared to 60% of humans.
ChatGPT can solve problems at a level that is equal to or exceeds that of an undergraduate student, according to a new study.
The researchers found that the GPT-3 model that powers the chatbotperformed as well as US college undergraduates when asked to solve reasoning problems that appear on intelligence tests or exams, such as the US college admissions test, the SAT.
See also: ChatGPT now comes to Android
Psychologists at the University of California, Los Angeles tested GPT-3's ability to predict the next image in a complex series of shapes, after converting the images into a text format that the model could process.
The same problems were posed to 40 UCLA undergraduates, and the researchers found that GPT-3 correctly solved 80% of the problems, well above the average score of just under 60% for the humans who participated.
The researchers also pushed the model to solve some SAT “analogy” questions — choosing pairs of words that are related in some way — that they believe have not been published online and therefore could not have appeared in the vast amount of data it was trained on. The UCLA team found that the AI outperformed the average human score.
In another test, the model fared less well. The researchers asked it and student volunteers to match a prose passage with a different short story that conveyed the same meaning. In this test, GPT-3 fared less well than the students, although GPT-4 — the improved version of GPT-3 — fared better than its predecessor, according to the research, which was published in the journal Nature Human Behavior.
The study found that GPT-3 exhibited an “unexpectedly robust” ability to detect patterns and infer, “matching or even surpassing human abilities in most settings.”
See also: AI will replace politicians!
The study's lead author, Taylor Webb, said the model driving ChatGPT was not up to the standards of artificial intelligence or human intelligence.
He said he struggled with social interactions, mathematical reasoning, and solving problems that required an understanding of physical space, such as figuring out which tools were best for transferring sweets from one bowl to another. However, technology was evolving very quickly.
The UCLA researchers added that without access to the inner workings of GPT-3, which was developed by the San Francisco-based company OpenAI , they could not determine how the model's reasoning abilities work and whether it thinks like a human or whether it exhibits a new form of artificial intelligence.
“GPT-3 can in some ways think like a human,” said Keith Holyoak, a professor of psychology at UCLA. “But on the other hand, humans didn’t learn by looking at the internet all the time, so the training method is completely different. We’d like to know if it’s actually doing it the way humans do or if it’s something completely new – a true artificial intelligence – which would be amazing.”.
Source: www.theguardian.com
