On the 7th, it was learned from Tsinghua University that Professor Fang Lu's research group in the Department of Electronics and Professor Dai Qionghai's research group in the Department of Automation have made significant progress in the field of intelligent optical chips. They pioneered a fully forward intelligent optical computing training architecture and developed the "Taiji-II" optical chip, which achieved in-situ optical training of large-scale neural networks and explored a new path for optical training of artificial intelligence (AI) large models. The relevant results were published online in the latest issue of the international academic journal Nature. The rapid development and widespread application of AI big models have made computing power a key strategic resource. Intelligent optical computing, with its advantages of high computing power and low energy consumption, has shown great potential in the post Moore era. Training and reasoning are the two cornerstones of the core capabilities of AI models. Previously, the emergence of the intelligent optical chip "Taiji" brought dawn to the "inference" of large-scale complex tasks, but failed to unleash the "training power" of optical computing. The training of existing optical neural networks heavily relies on GPU offline modeling and requires highly matched forward backward propagation models. This imposes strict requirements on the precise alignment of optical computing systems, making gradient calculation difficult and training scale small, which restricts the advantages of optical computing. Unlike existing training paradigms, we have abandoned backpropagation and taken a different approach by constructing a symmetric propagation model for optical neural networks, which can achieve efficient and high-precision optical training using only the forward propagation of the optical system Fang Lu told a reporter from Science and Technology Daily. It is reported that the emergence of "Taiji II" has filled the gap in the core field of intelligent optical computing for large-scale neural network training. In addition to accelerating AI model training, it has also demonstrated outstanding performance in high-performance intelligent imaging and efficient resolution of topological photon systems, opening up new paths for efficient and precise training of artificial intelligence large models, general artificial intelligence, and complex intelligent systems. (New Society)