%20(1).png)
.png)
Scientists discovered that GPT-3, a popular AI-powered tool, can reason as well as college undergraduate students.
The artificial intelligence large language model (LLM) was requested to answer reasoning tasks that were characteristic of IQ tests and standardised examinations such as the SAT, which are used to make admissions choices by colleges and institutions in the United States and other countries.
GPT-3 was asked to anticipate the next form after a sophisticated arrangement of shapes by researchers from the University of California - Los Angeles (UCLA) in the United States. They then prompted the AI to answer SAT analogy questions, assuring that the AI had never seen these questions previously.
They also assigned the identical challenges to 40 UCLA undergraduate students.
GPT-3 was shown to answer 80 percent of the tasks correctly in the form prediction test, compared to humans' average score of little under 60 percent and their greatest scores.
"Surprisingly, not only did GPT-3 perform similarly to humans, but it also made similar errors," said UCLA psychology professor Hongjing Lu, senior author of the study published in Nature Human Behaviour.
When it came to answering SAT analogies, the AI tool outperformed the average human score. Analogical reasoning is the process of addressing new issues by comparing them to old ones and extending the answers to the new ones.
The questions prompted participants to choose pairs of terms that had similar types of associations. For example, the solution to the problem "'Love' is to 'Hate' as 'Rich' is to which word?" is "poor."
However, when it came to answering analogies based on short tales, the AI performed worse than students. These issues required reading one piece and then finding another tale that expressed the same concept.
"Language learning models are just trying to do word prediction, so we're surprised they can do reasoning," said Lu. "In the last two years, the technology has advanced significantly from its previous incarnations." The researchers stated they were unsure how GPT-3's reasoning abilities worked since they did not have access to its inner workings, which were guarded by its inventor, OpenAI, and if LLMs are genuinely beginning to "think" like humans or are doing something altogether different that only mimics human cognition.
They stated that they want to investigate this.
"GPT-3 may be thinking like a human." However, because individuals did not learn by consuming the entire internet, the training process is quite different.
"We'd like to know if it's really doing it the way people do, or if it's something brand new - a true artificial intelligence - which would be amazing in its own right," said UCLA psychology professor Keith Holyoak, one of the study's co-authors.
