Why IQ is a poor test for AI

Time The presence of a recent pressOpenAI chief executive officer Sam Altman has said that he has made AI’s “IQ” rapidly improved for the past several years.

“Very fairly, it seems to me – it’s not scientifically accurate, it is just a joke or spiritual answer – every year we removed a standard deviation of IQ,” said Ultman.

The IQ user is the first non -Ultman to estimate a person’s intelligence as a criterion for AI progress. AI influencer Models on social media have given the IQ test and the results have placed the results.

However, many experts say that IQ is a weak measure of a model’s capacity – and a confusing one.

Sandra Watchter, studying the technology and control of Oxford, told TechCrank, “We can be very attractive to use the same measures to describe power or progress for humans, but it is like comparing apples with orange,”.

In his comment in The Pressure, Ultman made IQ equally equal to IQ. Yet IQ tests are relative – not objective – the arrangement Confirmed The kind of intelligence. There is Something Sens is minimal A reasonable test of IQ argument and abstract reasoning. However it does not measure it Practical Wisdom – Knowing how things work to work – and this is the best snapshot.

Watchter mentioned, “IQ is a tool for measuring human power – a competitive one is no less – scientists believe that human intelligence shows,” mentioned the watchter. “But you cannot use the same system to describe AI capabilities. A car is faster than humans and submarines are better in diving. However, this does not mean that cars or submarines surpass human intelligence. You are equivalent to an aspect of performance with human intelligence, which is more complicated ””

To acquire skills in the IQ exam, whose source Some ies tihasik Trace back to Eugenics, the extensively disrespectful scientific theory that can be developed by electoral breeds, must have a test -taking The memory of strong work and knowledge of Western cultural rulesThe It certainly invites the opportunity to prejudice, that is why A psychologist says the IQ exam The “ideally polluted mechanical model” of intelligence.

Any model that can do better in the IQ exams further indicates the test defects than the model’s performance, according to the Doctorate candidate of Washington University OS Kes, studying moral AI.

“[These] If you have infinitely in infinite amounts of memory and patience, the tests are very easy to play, “Kes said.” IQ tests are a very limited way to measure knowledge, sensitivity and intellect, which we already know before the invention of digital computer. “

IQ tests are probably an unjust benefit, as well as considering that models have a lot of memory and internal knowledge at their disposal. Often, models are trained on public web data and are full of questions taken from the web IQ test.

“Tests repeat very similar patterns – a great foolish way to increase your IQ is to practice taking IQ tests, which basically each [model] Did, “Mike Cook, a specialist in AI, Kings College London’s research fell No. “

In the end, IQ tests-such as biased-they were designed for humans, cooking was added-as a way to evaluate the problem of solving problems. They are inappropriate for a technology that goes to solve problems in a very different way than humans.

Cook said, “A crow may be able to use a tool to restore a treat from a box, but that does not mean that it can be admitted to Harvard,” said Cook. “When I solve a mathematics problem, my brain is fighting with the ability to read properly on the surface, not to worry about what I need to do on the way home, or even cold in the house. In other words, the human brain fights a lot more when a problem is solved – a problem, testing IQ or otherwise – and they do it with much less support [than AI.]”

All of these points Need AI Exam to betterAI is now the head of the institute AI scientist Heidi Khlaf TechCrunch.

“In the history of counting, we did not compare people with computer skills because the nature of the calculation means systems have always been able to complete work beyond human skills,” said open. “This is the idea that we compare the performance of the systems directly with human expertise are a recent event that is very rivalry and surrounds the controversy of growing and ongoing-lanth in AI systems evaluating.”