**Abstract: Key Perspectives on the Development of Artificial Intelligence by Dr. Ya-Qin Zhang** On June 6th, Dr. Ya-Qin Zhang, a member of the Chinese Academy of Engineering, Distinguished Professor at Tsinghua University, and Dean of the Tsinghua University Institute for AI Industries (AIR), delivered a speech titled "Some Perspectives on the Development of Artificial Intelligence" at the "Taihu Dialogue·Artificial Intelligence+" event, which also marked the establishment of the Tsinghua Wuxi Institute for AI Industry Innovation Center. In his speech, Dr. Zhang outlined several critical directions and viewpoints on the evolution of AI, particularly focusing on large AI models, the integration of AI into various sectors, and the governance of AI technology. Here are the main points from his address: ### Key Directions for AI Large Models Dr. Zhang identified five primary development directions for AI large models, which are foundational for the digital transformation (Digitalization 3.0): 1. **Multimodal Intelligence**: This involves the development of models that can analyze and generate content across multiple forms of data, including text, images, videos, LiDAR point clouds, 3D structures, 4D spatiotemporal data, and even biological information. The goal is to achieve comprehensive and deep intelligence in perception, decision-making, and content generation. 2. **Autonomous Intelligence**: The creation of intelligent agents (Agents) that can autonomously plan tasks, write code, use tools, and optimize processes. These agents will be capable of self-iteration, self-upgrade, and self-optimization, leading to highly autonomous intelligence. 3. **Edge Intelligence**: Deploying large AI models on edge devices such as AI PCs, AI smartphones, and AI TVs to achieve efficient, low-power, low-cost, and low-latency processing and response, thereby enabling edge intelligence. 4. **Physical (Embodied) Intelligence**: Applying large models to physical infrastructure such as autonomous vehicles, robots, drones, factories, transportation systems, communication networks, power grids, and power stations to enhance automation and intelligence levels, leading to embodied intelligence. 5. **Biological Intelligence**: Integrating large models with the human brain, living organisms, and biological entities to achieve biological intelligence. This will eventually lead to the convergence of information intelligence, physical intelligence, and biological intelligence. ### Five Perspectives on AI Development Dr. Zhang emphasized the following points regarding the future trajectory of AI: 1. **Large Models and Generative AI as Mainstream Technologies**: Over the next decade, large models (e.g., GPT-3, BERT) and generative AI will become the dominant technologies and industry trends. 2. **Ecosystem of Large Models**: The foundation of AI technology will be large models, which will coexist with vertical industry models and edge models, forming a new industrial ecosystem. This ecosystem is expected to be 100 times larger than the PC era and 10 times larger than the mobile internet era. Open-source and commercial models will provide developers with flexible choices. 3. **Core Elements of Large Models**: The two most critical elements of large models are token-based representation and scaling laws. Token-based methods unify the encoding of text and other data types, enabling the model to handle diverse inputs. Scaling laws demonstrate that increasing model parameters and data size significantly improves performance. 4. **Need for New Algorithm Systems**: Current algorithms are less efficient and more energy-intensive compared to the human brain. Dr. Zhang called for the development of new algorithm systems, including world models, DNA memory, intelligent agents, reinforcement learning (RL), probabilistic systems, and deterministic systems, to achieve a 100-fold efficiency improvement. He predicted that the current mainstream AI frameworks, such as Transformer, Diffusion, and AR, might be replaced by new technologies within five years. 5. **Path to General AI (AGI)**: Dr. Zhang forecasted that AGI could be achieved within 15-20 years, with the system passing a "new Turing test" in three stages: - **0-5 Years**: Information Intelligence, where AI will excel in language, image, sound, and video understanding and generation. - **0-10 Years**: Physical Intelligence, where AI will demonstrate robust understanding and manipulation capabilities in physical environments. - **0-20 Years**: Biological Intelligence, focusing on applications in the human body, brain-computer interfaces, biological entities, pharmaceuticals, and life sciences. ### Five Perspectives on Autonomous Driving Dr. Zhang highlighted the significance of autonomous driving as a critical application of physical (embodied) intelligence: 1. **Autonomous Driving as a Key Application**: Autonomous driving is expected to be the most important application of physical intelligence in the next five years, potentially becoming the first embodied AI system to pass the "new Turing test." 2. **Enhanced Safety and Human-like Experience**: Autonomous driving systems will need to be at least 10 times safer than human drivers and provide a more natural driving experience by mimicking and learning from human driving styles. 3. **Role of Large Models and Generative AI**: These technologies will enhance the generalization capabilities of Level 4 (L4) autonomous driving systems by generating high-quality corner case data, addressing long-tail problems, and improving common-sense reasoning. 4. **Multimodal Fusion and End-to-End Training**: Autonomous driving will integrate data from multiple sensors (visual, LiDAR, etc.), adopt end-to-end training, and utilize both cloud-based large models and vehicle-based real-time precise models to ensure comprehensive perception, timely response, and accurate decision-making. 5. **Vehicle-Centric Intelligence with V2X Integration**: Future autonomous vehicles will primarily rely on vehicle-centric intelligence, complemented by vehicle-to-everything (V2X) integration to ensure safety redundancy and enhance traffic efficiency. ### Five Suggestions for AI Governance Dr. Zhang also provided recommendations for the responsible governance of AI: 1. **Grading and Classification System**: Implement a grading system for large models, with specific constraints and evaluation metrics tailored to different application scenarios, especially high-risk areas like healthcare and autonomous driving. 2. **ID Entity Mapping**: Establish a system for mapping AI-generated content and entities to responsible entities, ensuring accountability and the reliability of information. 3. **Increased Investment in Risk Research**: Allocate at least 10% of AI investments to research the risks associated with large models, integrating development and governance, and aligning technology with policy and ethics. 4. **Red Lines and Boundaries**: Set clear boundaries for AI development to prevent catastrophic risks, such as prohibiting AI systems from autonomously replicating or improving themselves and seeking to increase their power and influence. 5. **International Communication and Collaboration**: Strengthen international cooperation in AI development, including joint research projects, international laboratories, and standardized governance frameworks to ensure global technological safety and stability. ### Conclusion As the fourth industrial revolution progresses, AI is becoming a pivotal force in societal and industrial advancement. The development of large AI models is crucial, with significant impacts on the intelligence levels of various industries. By focusing on multimodal, autonomous, edge, physical, and biological intelligence, and through the implementation of a robust governance framework, AI is poised to achieve new heights. In the realm of autonomous driving, the integration of large models and generative AI, combined with V2X technology, will enhance safety and efficiency, marking a significant milestone in AI's evolution.