Agricultural Bank Chairman Gu Shu: Big models currently face three major types of risks On June 18th, at the "Sixth Plenary Session of the 2026 Lujiazui Forum: Empowering High Quality Financial Development with Technological Innovation", Gu Shu, Chairman and Executive Director of Agricultural Bank of China, stated that in the process of applying big models, the main risks currently faced by AI big models include model black boxes, model illusions, model autonomous thinking, and uncertainty brought about by decision-making. In his view, this uncertainty risk can be further divided into three categories: the first is the difficulty of model interpretability caused by the massive number of parameters. Nowadays, the parameter scale of mainstream large models has reached the level of billions or even trillions. The matrix operation and generation of nonlinear superposition of massive parameters result in the decision-making mechanism and output results of the model being opaque and difficult to explain. The second challenge is the accuracy brought by probability generation. The decision-making process of a large model is not like the linear thinking process of humans when making decisions, but rather based on the probabilistic rules of word tokens calculated from massive training data. Essentially, it is probabilistic statistics, not logical factual deduction. When there is insufficient data and factual basis, it is easy to create self consistent illusions. Thirdly, the model is now capable of autonomous thinking and decision-making. With the evolution of large models and the deep application of intelligent agents, it has broken through the relatively rigid paradigm of traditional software input-output and can think independently. This has to some extent amplified the risks of uncontrollable processes and unknown results. (Securities Times)