AI下半场将走向何方？

Core Insights - The current stage of AI is characterized by rapid evolution, with a focus on the integration of large models, embodied intelligence, and scientific intelligence, forming a "knowledge flywheel" that could potentially surpass human learning capabilities in certain dimensions [1][2] - Despite advancements, AI development faces structural challenges such as computational power bottlenecks, data scarcity, and outdated evaluation metrics [1][2] Group 1: Challenges in AI Development - Data scarcity is a significant bottleneck for large models, limiting their growth despite attempts to expand multi-modal inputs and synthetic data generation [2][3] - The efficiency of AI systems is declining even as their intelligence levels improve, highlighting a need for a focus on the effective generation of tokens per unit energy consumption [2][3] - Current evaluation systems for AI models are prone to optimization for specific tasks, necessitating a shift towards dynamic, task-oriented assessments [2][3] Group 2: Originality and Causal Modeling - AI's limitations in natural sciences and mathematical modeling stem from its reliance on correlation rather than causal modeling, which is essential for scientific inquiry [3][5] - While some large models can understand causal language structures, their true comprehension of underlying causal logic remains uncertain [3][4] - The development of multi-modal large models is a new trend, raising questions about the need to move beyond token prediction to new paradigms like "world models" [3][4] Group 3: Future Directions and Breakthroughs - To achieve large-scale AI applications, overcoming energy efficiency bottlenecks is crucial, requiring real-time perception of physical environments and deep integration with sensors and actuators [8][9] - Two paths to enhance AI system performance include improving intelligence levels while maintaining energy efficiency and optimizing hardware-software collaboration [8][9] - The exponential growth in computational power requirements for large models poses a significant challenge, with training costs reaching approximately $10 billion and requiring 200,000 GPUs [8][9] - The potential of optical computing to enhance energy efficiency and communication bandwidth in distributed model training is highlighted, suggesting a shift towards low-precision model optimization [8][9] Group 4: New Paradigms in AI - A new paradigm proposed involves experience-driven AI, utilizing a large number of robots for intelligent collaboration in the physical world, which could surpass traditional large model training methods [9][10] - Future breakthroughs in AI will depend on advancements in both theoretical frameworks and system architectures [10]