Future Trends and Prospects of Artificial Intelligence
-
"Even someone like me, who works in artificial intelligence, was utterly amazed by the performance of ChatGPT." This sentiment was expressed by Harry Shum, a dual-appointment professor at the Institute for Advanced Study at Tsinghua University.
As an expert in computer vision and graphics, Shum has long been engaged in research areas such as human-computer interaction, statistical learning, pattern recognition, and robotics. In recent years, he has frequently debated with peers about whether artificial general intelligence (AGI) can be achieved, when it might be realized, and what standards define 'intelligence' in AI.
What struck him was that "before we could even settle these debates, ChatGPT emerged out of nowhere."
On July 23, at the Basic Science and Artificial Intelligence Forum held at the National Center for Science Communication, renowned scholars in the AI field gathered to discuss the challenges and issues facing large-scale models and AGI. Topics included the boundaries of AI models, datasets and training sets, computer vision, and knowledge graphs.
"Not only did people like us fail to anticipate this, even Bill Gates was caught off guard. Last June, Gates didn’t believe this could be achieved. It wasn’t until August when he saw the model—which correctly answered 59 out of 60 questions—that he became convinced," Shum said.
Even world-class scientists and decades-long experts in the computer industry couldn’t help but marvel at the changes brought about by AI products like ChatGPT from late last year to early this year.
"Mathematics and physics are crucial to information science, while fundamental disciplines also need to effectively utilize new-generation technologies like AI to advance their own development," said Shing-Tung Yau, a Fields Medalist and Chair of the International Congress of Basic Science, in a video address played at the forum.
In his view, information science can generate important and meaningful mathematical problems. These problems are "being vigorously studied by mathematicians," and AI is influencing the development of mathematics itself. He hopes young scientists will fundamentally understand AI and play a significant role in its widespread application.
Large AI models require vast amounts of data and computational resources to build. Zhou Ming, Vice Chair of the China Computer Federation (CCF), pondered how these models could be practically "implemented" in future societal life.
During the training of large models, at what point does intelligence truly emerge? What is the mechanism behind this emergence? Shum raised one question after another during the discussion, ultimately concluding that the key issue is the lack of proper mathematical tools.
"This is precisely what we’re discussing today—the relationship between basic science and AI," he said.
As Shum put it, behind many scientific and technological advancements, there are powerful mathematical tools and principles at work, and the field of AI is no exception.
Adding to this topic, Jianwei Zhang, a professor at the University of Hamburg and Director of the Institute for Multimodal Intelligent Systems, noted: "We don’t just need mathematical models; we also need physical models, biological models, and brain science models."
Zhang’s research focuses on perception learning and planning in intelligent systems, multi-sensor information processing and fusion, intelligent robotics, and human-computer interaction. He mentioned that while robotics has made significant progress in processing single-modal information, it still lags far behind humans in handling multi-modal information, especially in dynamic environments.
"I believe that a method combining physics, physiology, models, and big data to drive AI is the only path forward for achieving intelligent robots," Zhang said.
Zhang observed that enthusiasm for robotics is high in China, with a strong industrial foundation and supportive environment. However, he emphasized the importance of advancing robotics research within an ethical framework, avoiding excessive planning of AI’s creativity to preserve human innovation—a topic he considers "highly worthy of attention and exploration."
Regarding model training, Yang Ge, a founding member of Elon Musk’s AI company xAI, argued that as models scale, richer and higher-quality datasets are needed. These should shift from being opinion-heavy to emphasizing mathematical and scientific content, with a focus on logic and reasoning.
In Yang Ge's view, the cognitive structure of AI is completely different from that of humans. The only similarity is that large AI models are trained on human data, which to some extent "may feel like we are talking to ourselves." However, artificial intelligence remains distinct from humans, and whether AI can integrate into society like humans is still "hard to say" at this stage.
"ChatGPT is not a person, and we do not treat it as one. AI communicates with humans in its own unique way," Yang Ge said.