Building Intelligent Machines Helps Us Learn How Our Brain Works

March 19, 2024

Designing machines to think like humans provides insight into intelligence itself

By George Musser

he dream of artificial intelligence has never been just to make a grandmaster-beating chess engine or a chatbot that tries to break up a marriage. It has been to hold a mirror to our own intelligence, that we might understand ourselves better. Researchers seek not simply artificial intelligence but artificial general intelligence, or AGI—a system with humanlike adaptability and creativity.

Large language models have acquired more problem-solving ability than most researchers expected they ever would. But they still make silly mistakes and lack the capacity for open-ended learning: once they are trained on books, blogs, and other material, their store of knowledge is frozen. They fail what Ben Goertzel of the AI company SingularityNET calls the “robot college student test”: you can’t put them through college (or indeed even nursery school).

The one piece of AGI these systems have unequivocally solved is language. They possess what experts call formal competence: they can parse any sentence you give them, even if it’s fragmented or slangy, and respond in what might be termed Wikipedia Standard English. But they fail at the rest of thinking—everything that helps us deal with daily life. “We shouldn’t expect them to be able to think,” says neuroscientist Nancy Kanwisher of the Massachusetts Institute of Technology. “They’re language processors.” They skillfully manipulate words but have no access to reality other than through the text they have absorbed.

In a way, large language models mimic only the brain’s language abilities, without the capacity for perception, memory, navigation, social judgments, and so forth. Our gray matter performs a bewildering mashup of overlapping functions, some widely distributed across the brain, others more localized. People who have suffered a stroke in one of their language areas may be unable to speak but may still be able to add numbers, compose symphonies, play chess and communicate by gestures as well as they could before. AI developers are incorporating such modularity into their systems in the hope of making them smarter...

Read the full article on the Scientific American website using the link below.

Associated CBMM Pages: