Dmitry Rybin, a graduate of the Higher School of Economics and a mathematician, has made an unexpected breakthrough in network flow theory. Using the GPT-5.6 Pro language model, he managed to find a counterexample to the Dinic–Garg–Goemans conjecture, which had remained unproven for nearly 30 years. This discovery could radically change the approach to solving complex combinatorial problems.
The conjecture, formulated in 1999, concerned fundamental properties of network flows and was long considered one of the key unsolved problems in this field. Rybin did not give the model any special hints or guiding algorithms. Instead, he used just four prompts with general formulations: "make a breakthrough," "refute the general case," "continue exploring." According to the mathematician, he viewed this chat as a "pure meme"—an experiment that unexpectedly yielded a real scientific result.
GPT-5.6 Pro constructed the refutation in a few hours, which is impressive in itself given the complexity of the problem. However, it is worth noting that the result has not yet undergone independent verification by the scientific community. If the counterexample is confirmed, it will become a powerful argument in favor of using large language models in fundamental mathematics.
My analysis of the situation: This case demonstrates that modern AI models are capable not just of processing data, but also of generating fundamentally new ideas in areas requiring deep abstract thinking. However, we are still far from full automation of mathematical discoveries—the role of the human expert, who asks the right questions and interprets the results, remains critically important. If independent verification confirms Rybin's finding, we will witness a new stage in the symbiosis of humans and artificial intelligence.