Andrew Chi-Chih Yao, Turing Award winner, an academician with the Chinese Academy of Sciences, dean of the Institute for Interdisciplinary Information Sciences at Tsinghua University

From a theoretical perspective, I would say that AI for science is the most interesting, important and promising direction for AI research over the next three to five years. AI has demonstrated tremendous power across many domains, but AI algorithms still operate within the boundaries set by the laws of physics and mathematics. Many exciting avenues of exploration lie ahead, including emerging fields such as quantum AI, reliable large-scale AI systems, AI safety and AI for AI. AI and quantum technologies are two powerful tools that continue to expand the frontiers of human knowledge. AI can also accelerate the development of quantum computing. Error correction has been one of the greatest challenges in building quantum computers for the past 40 years, but AI has recently begun to provide promising solutions. Quantum AI is beginning to emerge. I believe we will see tremendous progress over the next five to 10 years. It represents a frontier beyond present-day AI and has the potential to become even more powerful.

Su Hao, inaugural dean of the Institute of General Physical Intelligence at Fudan University

Hallucinations in large language models stem, to a large extent, from their lack of a physical "body". To overcome these hallucinations, we must move beyond the boundaries of the digital world, engage in firsthand experience, and submit ourselves to the judgment of the physical world by making predictions, taking action and learning from reality's feedback. This age-old process is called experimentation, and it is also the key to physical intelligence. Physical intelligence is not a substitute for humans, but a collaborator. It fills the gaps where human capabilities fall short, allowing people to focus on judgment and creativity. The mission of physical intelligence is to return humanity to humans. The breakthrough in physical intelligence lies not in model architecture, but in the aggregation of knowledge. The industry's focus will shift from spectacular demonstrations to reliable system operation. Generalizability is the destination, but reliability is the starting point.

Richard S. Sutton, Turing Award winner, principal founder of the field of Reinforcement Learning, professor of computing science at the University of Alberta

The "Bitter Lesson" is based on the historical observation that AI researchers have often tried to build knowledge directly into their agents. This approach always helps in the short term and is personally satisfying to researchers. But in the long run, it often reaches a plateau and can even inhibit further progress. Breakthroughs eventually come from the opposite approach: scaling computation through search and learning. The lesson, therefore, is that from the very beginning we should focus on scaling computation through search and learning. We should seek the fundamental principles of intelligence rather than become distracted by the specific characteristics of the human mind. We should allow the agent to discover these principles on its own.

Zhu Songchun, founding director of the Beijing Institute for General Artificial Intelligence

It is not true that simply having enough high-quality data will lead to artificial general intelligence. The first principle of building AGI is that it must possess a subjective mind, or humanlike "heart". It should have a value system that encompasses the self, others and the collective. It should also possess a cognitive framework that enables it to learn, communicate and collaborate with people. I hope we do not remain followers. A fundamental prerequisite for scientific and technological innovation is independent thinking and cultural confidence. Traditional Chinese philosophy and culture are, in fact, highly advanced. We should draw more on this Chinese wisdom in technological innovation, and I believe we will be the first to achieve AGI.