During Recent appearance in printOPENAI CEO of Sam Altman said he noticed that “IQ” AI quickly improved in the last few years.
“Very roughly, it seems like – this is not scientific, this is just a vibe or a spiritual response – every year we start a standard deviation of the IQ,” said Altman.
Altman is not the first to use IQ, an assessment of a person’s intelligence, as a measure to progress AI. AI influencers On social media, they gave models IQ tests and ranked results.
But many experts say IQ is a bad measure of model ability – and misconception.
“It can be very tempting to use the same measures that we use for people to describe opportunities or progress, but it’s like comparison of apples with orange,” Sandra Wachter, a researcher who studies technology and regulation in Oxford, told Techcrunch.
In his comments on The Presser, Altman equalized IQ with intelligence. Still IQ tests are relative – not objective – measures sure types of intelligence. There some consensus That IQ is a reasonable test of logic and abstract explanation. But that is not measured practical Intelligence – knowing how to function things – and that is at best the recording.
“IQ is a tool for measuring human abilities – controversial no less – based on what scientists believe that human intelligence looks,” Wachter noted. “But you can’t use the same measure to describe AI capabilities. The car is faster than people, and the submarine is better in diving. But that doesn’t mean cars or submarines outweigh human intelligence. .
To execute on the IQ test, whose origin is Some historians Following back to eugenics, a widespread scientific theory that people can improve selective breeding, the test test must have strong work memory and knowledge of Western cultural norms. This calls for a bias opportunity, of course, why One psychologist called IQ tests “Ideological corrupt mechanical models” intelligence.
In order for the model to manage well on the IQ test, it indicates more about the disadvantages of the test than the performance of the model, according to the Keyes OS, a doctorate of a candidate at the University of Washington, studying ethical AI.
“[These] The tests are quite easy to play if you have a practically infinite amount of memory and patience, “Keyes said.” IQ tests are a very limited way of measuring cognition, feeling and intelligence, something we knew from before the invention of the digital computer itself. “
AI probably has an unjust advantage in IQ tests, considering that models have huge amounts of memory and internalized knowledge. Often, models are trained for public information, and the web is full of examples of questions downloaded from IQ tests.
“The tests are repeated by very similar patterns – a rather flawless way of lifting your IQ is to exercise IQ tests, which is basically what every every one [model] has done, “said Mike Cook, a scientific associate at King’s College London, specializes in AI.” When I learn something, I don’t go into my brain with perfect clarity of a million times, unlike AI, and I can’t process it without any loss or a signal loss. “
Ultimately, IQ tests-the trade-friendly are-dose for humans, Cook is added to a way to evaluate the general abilities of solving problems. They are inappropriate for technology that approaches solving problems in a completely different way than people do.
“The crow could use a recovery tool from the box, but that doesn’t mean it can be enrolled on Harvard,” Cook said. “When I solve the problem of mathematics, my brain also struggles with my ability to read the words on the page properly, not to think about shopping I have to do on the way home or this is too cold in the room right now. In other words, the human brain struggles with a lot More things when they solve the problem – any problem at all, IQ tests or otherwise – and make it with much less help [than AI.]”
All of this indicates need for better AI testsHeidy Khlaaf, AI chief scientist at the AI Now Institute, told Techcrunch.
“In the history of counting, we have not compared computer skills to that man precisely because the nature of calculation means that systems have always been able to complete the tasks that are already beyond human abilities,” Khaaf said. “This idea to directly compare system performance with human abilities is a recent phenomenon that is very controversial and what surrounds the controversy of constant growing-and-referential values that are created to evaluate the AI system.”