during Modern journalistic appearanceThe CEO of Openai Germans said that he noticed that “intelligence” of artificial intelligence is improving quickly over the past few years.
“It is almost similar to me – this is not a scientific accurate, as this is just an answer or a spiritual answer – every year we convey one normative deviation to intelligence intelligence.”
Altman is not the first to use the intelligence rate, which is an appreciation of a person’s intelligence, as a criterion for the progress of artificial intelligence. Amnesty International On social media, I intelligence tests were given and the results were classified.
But many experts say the intelligence rate is a bad measure of the capabilities of the model – misleading.
“It may be very tempting to use the same measures that we use for human beings to describe the capabilities or progress, but this is similar to comparing apples with orange,” said Sandra Washer, a researcher who studies technology and organization in Oxford.
In his comments in The Presser, Altman equated intelligence rate with intelligence. However, intelligence tests are relative – not objective – to be sure Types of intelligence. there some consensus Intelligence is a reasonable test of logic and abstract thinking. But it does not measure practical Intelligence – know how to make things work – at best a snapshot.
“The intelligence rate is a tool for measuring human capabilities – a few – based on what scientists believe that human intelligence appears to be,” Wachter pointed out. “But you cannot use the same procedure to describe the capabilities of artificial intelligence.
To excel in the intelligence test, its origins Some historians It follows it to the improvement of birth control, and a widely credible scientific theory that people can be improved through selective education, must have the homogeneous test a A strong work memory and knowledge of Western cultural standards. This calls for the opportunity, of course, which is the reason One of the psychologist has invited intelligence tests The “ideological mechanical models” for intelligence.
This model may be good in the intelligence test indicates more test defects than the model, according to OS Keyes, a doctoral candidate at the University of Washington studies artificial intelligence.
“These tests are very easy if you have an infinite amount of memory and patience,” Keys said. “Intelligence tests are a very limited way to measure perception, sensation and intelligence, which is something we have known since the pre -invention of the digital computer itself.”
Artificial intelligence may have an unfair advantage in intelligence tests, bearing in mind that models have huge amounts of memory and internal knowledge at their disposal. Often, models are trained in public web data, and the web is full of questions taken from intelligence tests.
“The tests tend to repeat very similar patterns – a very guaranteed way to raise your intelligence rate is to practice intelligence tests, which is what each (model) did mainly. “. . “When I learn something, I don’t get tubes in my mind very clearly 1 million times, unlike artificial intelligence, and I cannot treat it without noise or signal loss, too.”.
Ultimately, intelligence tests-biased as they are-for human beings, added Cook-intended as a way to assess the abilities to solve public problems. It is not suitable for technology that is close to solving problems completely different from people.
“The crow may be able to use a tool to recover a box, but this does not mean that it can be recorded at Harvard University,” Cook said. “When I solve the problem of mathematics, my mind also competes for its ability to read the words on the page properly, or not think about the shopping that I need on its way home, or if the atmosphere is very cold in the room now. In other words, human brains correspond to A lot of things when you solve a problem – any problem at all, intelligence tests or otherwise – and they do this with much lower help (artificial intelligence.) “
All this indicates the need for better Amnesty International tests.
“In the history of the account, we did not compare computing capabilities with human capabilities specifically because the nature of the calculation means that the systems have always been able to complete the tasks that already go beyond human ability,” said Khalib. “This idea, which we compare directly to the performance of systems against human capabilities, is a very modern phenomenon that is very competing, and what surrounds the controversy in the standards that are constantly expanding-and move-which are created to assess artificial intelligence systems.”
