Cornell mathematician Strogatz divided the role of computers in mathematics into three stages: servant, scout, and oracle. AI now only gives answers but no reasons. Mathematics may have to switch to medical verification, at the cost of losing absolute certainty. (Background supplement: OpenAI threw out 722 mathematics papers at one time, and some scholars’ years of hard work returned to zero overnight) Mathematicians say that AI is becoming an “oracle,” which means that it gives seemingly correct answers, but cannot explain the reasons, and humans can neither understand nor catch up. Steven Strogatz, an applied mathematician at Cornell University, was one of the authors of the 1998 paper on small-world networks. He and his colleague Alex Townsend co-wrote the new book "Big Math: The Hidden Codes That Run Our World". In an exclusive interview with Ars Technica, he divided the role of computers in mathematics into three stages. The background is that last month OpenAI released 722 mathematical manuscripts produced by AI, which is called "mathocalypse" (mathematical doomsday) in the circle. What is the difference between "Oracle" and the previous use of computers to do mathematics? The first stage is "servant". The computer calculates according to the instructions given by the person. He said that it is like a microscope or telescope, but it only extends the human senses. The second phase was the emergence of the "Scout" in the 1950s. Enrico Fermi and others used computers to simulate the vibration of chemical bonds between atoms. They originally expected that the energy would gradually spread out and become disordered. As a result, 98% of the energy returned to the original vibration mode. This accident later developed into the soliton theory. He said that when entering the "jungle" of nonlinear systems, human intuition will fail, and the computer can only be sent to explore the path and report back. But at that time, the computer is teaching people, and humans can still catch up and find theoretical explanations for the results. The third stage is the current "oracle": the computer gives results that humans cannot understand and cannot explain, but the results seem to be correct. Strogatz gave two examples. Google Translate does a good job of translating, but it doesn’t discover new linguistic principles, and even if it does, it can’t explain them. AlphaFold allows developers to win the 2024 Nobel Prize in Chemistry. Humans can solve a thousand times more protein structures than before. It does not rely on learning new physics or chemistry, but just identifying patterns. It is also impossible to say how to do it. He said the situation will only get worse in the future... If you don’t understand the answer, how do you know it’s correct? Strogatz's answer is to borrow from medicine. The launch of new drugs requires a large number of controlled experiments and randomized trials to prove their safety. Doctors do not need to know how aspirin works to use it with confidence. "If AI is an oracle, you can still check it by how often it gets it right," he said. The trade-off is that math will become more like science, medicine and engineering. Pure mathematics has always prided itself on absolute proofs: the proofs in "Elements", written around 300 BC, are still valid today, but none of them are still accepted in contemporary science and medicine. If we use scientific methods to verify instead, the purest pure mathematicians would be sacrificing the certainty they cherish most. He also left some room: AI may become a very good teacher, and maybe humans will understand it in 10 years. But he believes that some systems will really not understand because of human limitations (not fast enough, not smart enough, not memory enough), even with AI assistance. What is left for mathematicians? Let’s talk about losses first. He said he half-jokingly went through the five stages of grief, but the 25-year-old student who dreams of becoming a mathematician now doesn't find it funny. Pure mathematicians such as Terence Tao have been emphasizing "understanding": the process of seriously trying to understand something will force out various solutions and simplified versions, which means that you learn more along the way than the answer itself. And where is the hope? He said that AI is currently not good at asking good questions, just like doctoral students who just entered graduate school, because universities are taught to give answers, not to ask questions. But he thinks it's something that can be trained. Even if you can't make big discoveries, you can still enjoy small discoveries, just like solving Sudoku. The fun lies in thinking itself. It's just that it's hard for someone who wants to be the first to make a big discovery. As for the overall situation, he says: "It's very confusing, that's the only honest assessment." His new book ends on a Faustian note: You may be redeemed by continuing to pursue something bigger than yourself, but it's still a dangerous game. Related reports OpenAI overturns the exponential time hypothesis and solves the problem of Millennium Yang-Mills: Three AI scientific breakthroughs shocked the academic world. Mathematicians are unhappy with AI companies! 25 people signed the Fields Medal to criticize: It can only solve problems, not mathematics. OpenAI said that the new model it is training has solved more than 100 mathematical problems, and it has simultaneously established a scholar advisory group. AI is not smarter than mathematicians, but it remembers more than people. OpenAI’s reasoning model cracked 80-year-old mathematical puzzles, and three top mathematicians endorsed it. Can you still trust AI proofs if you can’t understand them? Mathematician: It’s like a new drug. Look at the correct answer rate. This article was first published in BlockTempo, the most influential blockchain news media.