Your phone can already hold every serious chess game that grandmasters have ever played. That is fewer than five million games, once you throw out the endless online blitz. Add the whole library of openings, and the phone still has room to spare.
Vladimir Spiridonov, a Doctor of Psychological Sciences who heads the Cognitive Research Laboratory at the Institute of Social Sciences (ION) at RANEPA, likes to start from this fact. It sounds like a victory for the machine. He sees a puzzle.
Chess computers do play better than people now. But the grandmasters they beat have neither unlimited memory nor unlimited speed, and they still play "very decently", as Spiridonov puts it with some irony. How? With the one equipment a person has: a mind that is bounded by definition.
The stakes reach well beyond the chessboard. If Spiridonov is right, the things we usually count as defects, a short memory, a small working space, the failure to grasp a problem at once, are where discovery comes from. And the tools that remove those defects may remove the discovery too.
His bet is that our limits sit at the base of our biggest advances and creative wins. His reasoning is simple. Imagine material that offers you no resistance. You can chew it, rearrange it, plough it over in any pattern you like, and nothing new happens. Something new has a chance only when you hit the edge: you use every method you have, every piece of knowledge you can reach, and it is not enough.
When the material puts up no resistance, there will be no creativity. You will just chew it, shuffle it, plough it over any way you like. Everything most interesting begins when you can't cope.
Когда материал не оказывает сопротивления, никакого творчества не будет. Ты просто будешь его пережевывать, перекладывать, перепахивать как угодно. Все самое интересное наступает тогда, когда ты не справляешься.
— Vladimir Spiridonov, Think Like a Scientist, 71:08
Now put a large language model next to this picture. Spiridonov is careful here. He says "as far as I understand", and he admits he has only modest experience, mostly from tinkering with DeepSeek. His reading is that these systems are poorly built for this exact spot. The line between what they know and what they don't barely exists for them. Even when people make special efforts, nothing changes: a system that does not understand what it does not understand keeps not understanding it.
He sees a second difference in how they learn. When a GPT model passed a medical exam, people applauded, and the result is impressive. But the machine got there by taking in a boundless pile of textbooks, which no person can do. Would you let such a system into a patient's room, he asks, and how would it examine anyone? Something has been swapped in the discussion, he says: knowing the text is not the same as doing the work.
Here the story takes a turn, and Spiridonov grants something to the other side. In the conversation the host wondered whether insight could ever live inside a language model, which learns from words that people have already written. Spiridonov points to where insight was first found. Not in people, but in chimpanzees. Wolfgang Köhler watched them on Tenerife, more than a hundred years ago. A chimp would fumble about, and then the behaviour changed sharply, as if someone had turned a lever. Köhler could prove something had happened because the animal carried its trick over to a similar, different task. The cleverer chimps did this well.
Spiridonov finds nothing like this in the language models. Their basic principles are associative, and associationists were exactly the people Köhler spent his career arguing against. The Gestalt psychologists, Köhler's camp, defeated the associationists in theory in the early twentieth century, he says, and now the Gestaltists are gone while the associative models triumph. "Science goes in circles", he remarks. Where the road leads from here, he adds, is unclear.
Studies of machine creativity fit this picture in an odd way. In Science Advances in 2024, a team led by Anil Doshi found that short stories written with help from generative AI were rated as more creative and better written, most of all for less creative writers. Yet the AI-assisted stories were more like each other than stories written by people alone. Each writer gained, and the group lost some novelty. A separate paper in Scientific Reports in 2024, led by Kent Hubert, compared GPT-4 with 151 people on creativity tests. GPT-4 scored higher on originality and elaboration once the number of answers was matched. But people used far more words that nobody else in the sample used: about 70 percent of their words occurred once, against 48 percent for the model.
Does that prove Spiridonov right? No, and he never claims it does. The numbers show only that a machine can score well on a test while its output drifts toward sameness. The harder question he leaves open: can a system with no sense of its own edge ever want to push past it?
A grandmaster sitting over a position he cannot solve is, on this view, in the best possible place.
