2 1 Shi Ji Jing Ji Bao Dao·2025-12-30 10:25Core Insights - The rapid development of AI large models is becoming a key factor in the new round of technological competition, with a belief that the number of foundational large models will converge to a single-digit figure, while numerous specialized models and applications will emerge across various industries [2] - Physical AI is highlighted as a significant area of focus, accelerating advancements in fields like embodied intelligence and autonomous driving, which are expected to profoundly change societal operations [2][3] - The transition from generative models to world models and visual language models (VLA) represents a paradigm shift in AI, moving from mere prediction to simulation and physical alignment [3][4] Industry Trends - The emergence of Sora has sparked discussions about world models, indicating a shift in AI capabilities from being mere predictors to becoming simulators [3] - The divergence in world model approaches has led to the classification of models into "generative" and "representational" camps, with each having distinct applications and strengths [4][5] - The integration of VLA and world models is seen as a trend, with VLA focusing on sequence modeling for robot control and world models emphasizing internal environmental modeling for efficient learning [5] Challenges and Solutions - Three major challenges remain for world models: understanding causality, building effective simulators, and addressing data scarcity issues [6] - The competition for high-quality synthetic data is crucial for the next phase of AI development, particularly in data-driven AI applications like autonomous driving [6] - The timeline for the realization of world models is projected to span from 2024-2025 for visual simulation to 2028-2030 for general embodied intelligence [6] Technological Evolution - The network architecture is evolving from "cloud-native" to "AI-native," necessitating a focus on performance and collaboration between computing and networking [7] - ZTE has been progressively advancing its hardware and software integration from 2G to 5G, now incorporating large models into its development paradigm [8] - The integration of AI into core business processes is expected to transform industries, with a shift from content generation to autonomous action [9] Implementation and Applications - ZTE's "Co-Sight Intelligent Agent Factory" aims to enhance reasoning capabilities and ensure decision-making reliability through advanced verification mechanisms [11][12] - The successful application of AI requires a combination of robust infrastructure, effective methodologies, and deep industry engagement [17] - Industries such as education, healthcare, software development, and smart manufacturing are identified as likely candidates for early AI value realization due to their structured data environments and feedback mechanisms [14][13] Future Directions - The hybrid approach of "cloud-edge collaboration" is recommended for integrating general foundational models with industry-specific enhancements [15] - The need for specialized models in non-natural language data scenarios is emphasized, particularly in high-stakes environments like finance [16] - The overarching narrative of AI is shifting towards practical applications in various sectors, moving away from mere technological showcases to tangible value creation [18]