As the father of deep learning, Jeffrey Hinton's every move has attracted worldwide attention. Recently, he resigned from Google and publicly expressed concerns about AI, which sparked a thousand waves of discussion among AI researchers and the public for a month. In the interview at that time, his main argument was that AI could achieve greater intelligence than humans by continuously spreading copies and sharing information through self replication. As one of the most active brains in the AI research community, he always has a new perspective on how machines understand the world. During the month of calling for danger, Hinton constructed a complete new theory for understanding neural networks to explain his previous arguments about threats in detail. Recently, Hinton attended the 2023 Zhiyuan Conference, known as the "AI Spring Festival Gala" in China, and delivered a keynote speech titled "Two paths to Intelligence" online as the closing speech of the conference. He discussed a new type of hardware based finite computing to limit the potential threat of AI infinite replication, but overall he still believes that AI's capabilities will soon surpass those of humans, posing a greater threat. In this speech, Hinton proposed a new possibility for artificial intelligence: finite computing. A more pure natural machine. Unlike our current situation where hardware and software can be separated, in finite computing, hardware itself is the software that operates. It requires building hardware using our learning of neurons, and using voltage to control hardware learning, just like a human brain. This type of limited computing can bring lower energy consumption and simpler hardware production, but there is currently no good learning algorithm to achieve effects like deep learning, and its limitations also make it difficult to expand. Afterwards, Hinton discussed the issue of how intelligent groups, such as humans and AI, share knowledge. There are two ways to share knowledge, one is through biological distillation, which has low bandwidth but also low energy consumption. Another type is digital copying, which requires high bandwidth and energy consumption. Deep learning currently uses distillation to learn from us (in the text), but it can itself share knowledge using digital copies. Therefore, although AI currently has low learning efficiency, once it obtains its own learning methods for understanding the world through multimodal means, and through digital copying, it can quickly share knowledge. This will ultimately make them much smarter than us. Sinton continued to infer that under the principle that if one wants to complete a task, one must gain power, it is difficult to imagine an artificial intelligence that is smarter than us not going to do everything possible to gain power. And this is the root cause of the threat. In nature, there are no examples of low intelligent creatures controlling high intelligent creatures, and often high intelligent creatures are not very friendly to low intelligent creatures. (New News Agency)