New AI model scores at ‘human level’ on intelligence tests

At the moment, AI models are very specialized. Chatbots are very good at having ‘conversations’, other models are very good at recognizing tumors on radiographs. But the AIs above are not really ‘intelligent’. Chatbots predict which words should be most coherent and logical behind each other, without giving them meaning. The AI ​​radiologist looks for spots in the same way we solve ‘spot the seven differences’ puzzles. It looks for patterns, but gives no meaning to them.

But OpenAI’s new O3 model would be different and could pass for an AGI, ie Artificial General Intelligence. This is an AI model that can handle various tasks, learns autonomously and can reason at the level of a human.

To measure how ‘human smart’ an AI model is, it is subjected to the so-called ARC-AGI benchmark. That stands for Abstract and Reasoning Corpus for Artificial General Intelligence. Simply put: the higher the score, the ‘more human’ the reasoning ability of the AI ​​model is.

Until now, the best AI models scored 55 percent on that test. O3 would achieve 85 percent, which is the score of an average person. For those interested: the test includes logic puzzles and problems that are often found in IQ tests.

However: at this point we have to take Altman and OpenAI at their word. Apart from a handful of researchers, no one has seen O3 in action. The model is currently being tested for safety. It is not known when the model will be available.