In a recent press conference, Openai CEO Sam Altman said that AI’s “IQ” has been rapidly improving in the past few years.
“Very rough, this is not scientifically accurate, this is just an atmosphere and mental answer. It moves one standard deviation of IQ every year,” says Altman.
Altman is not the first IQ using IQ as an AI progress benchmark. Social media AI influencers have given a model IQ test and ranked results.
However, many experts say that IQ is a poor model of model abilities and is misleading.
Sandra Wachter, a researcher who is studying Oxford’s technology and regulations, said, “It’s the same thing as using the same way to use it for humans to describe abilities and progress. Is like comparing apple and orange.
In his comment in the Preser, Altman identifies IQ with intelligence. However, the IQ test is not a relative, not a specific type of intelligence measurement. There are several consensus that IQ is a reasonable test between logic and abstract inference. But it’s not a practical intelligence, but a snapshot at best.
“IQ is a tool for measuring human abilities based on what scientists believe that human intelligence looks like.” But use the same measured value to explain the AI function. The car is faster than a human, but this is not the same as the human intelligence. This is much more complicated.
Test takers, which are not trusted scientific theory that can improve people through selective propagation, are so powerful that some historians are superior to eugenology. You need to have work memory and knowledge. This is, of course, a biased opportunity. For this reason, a psychologist calls the IQ test a “mechanical model that is easy to rot in ideology.”
According to OS Keyes, a doctorate candidate studying the ethical AI of the University of Washington, the possibility that the model may work in the IQ test shows more about test defects than the model performance. 。
“The (these) tests are very easy if they have infinite memories and patience,” said the keys. “The IQ test is a very limited method for measuring cognitive, sensation, and intelligence that measure what you know before the invention of the digital computer itself.”
AI may have unfair advantages in the IQ test, given that the model can freely use the model with a large amount of memory and internal knowledge. In many cases, the model is trained in public web data, and there are many examples of questions taken from the IQ test.
“Tests tend to repeat very similar patterns. A fairly complete way to raise the IQ is to practice IQ tests. This is basically what (model) did. ” “When you learn something, unlike AI, you will not be completely clearly piped into the brain one million times, and you can not process without noise or signal loss.”
Eventually, the IQ test is intended to be designed for humans, adds cooking, and evaluates general problem -solving skills. They are inappropriate for the technology that approaches the problem in a very different way than people.
“The crow may be able to recover snacks from the box using a tool, but that doesn’t mean you can register for Harvard,” said Cook. “When solving mathematics problems, my brain is also fighting with the ability to read the words on the page correctly. In other words, the human brain is fighting more when solving the problem. All problems, IQ tests, etc., and they do it with much less (than AI).
Heidy Khlaaf, AI Scientist at AI Now Institute, told TechCrunch, all indicating the need for a better AI test.
“In the history of calculation, the nature of the calculation is that the system is not accurately compared to the human ability because the system has always been able to complete the task beyond human abilities,” says Khlaaf. I mentioned it. “This idea, which directly compares system performance with human abilities, is a recent phenomenon that has been very contested, surrounding the benchmark controversy created to evaluate the AI system.”