Core Viewpoint - The evolution of AI applications can be categorized into four stages: Perception AI, Generative AI, Agent AI, and Physical AI [9]. Group 1: Stages of AI Evolution - The first stage, Perception AI, includes traditional technologies such as natural language processing, speech noise reduction, and image recognition, which have been commercialized in many terminal devices for years [13][14]. - The second stage, Generative AI, emerged with the rise of ChatGPT, focusing on pre-training with large datasets and completing specific tasks under human supervision, including text-to-image generation and chatbots [14][19]. - The third stage, Agent AI, allows for autonomous actions, predictions, intent understanding, and task orchestration with minimal human intervention [18][19]. - The fourth stage, Physical AI, is still in the research phase, where AI can understand the physical world and respond according to real physical laws [21][22]. Group 2: Current Industry Trends - The industry is currently transitioning from Generative AI to Agent AI, with a focus on enhancing terminal capabilities from single text modalities to multi-modal interactions [4][19]. - The deployment of large models on terminal devices faces challenges such as memory limitations, bandwidth constraints, and power consumption [6][30][34]. Group 3: Advantages and Challenges of Edge AI - The primary advantage of running large models on terminal devices is personalization, as data generation occurs close to the source, enhancing privacy and security [31]. - Edge AI also offers the benefits of being free and not requiring internet connectivity [32]. - Challenges include memory limitations that restrict model size, bandwidth limitations affecting inference speed, and the need for efficient power management in high-integration devices [34][35][36]. Group 4: Technological Innovations - Qualcomm has developed several technological innovations to address these challenges, including quantization and compression techniques to reduce memory usage, parallel decoding to enhance token generation speed, and advanced NPU architectures for improved performance [37][39][40]. - The parallel decoding technique allows for the generation of multiple tokens simultaneously, improving efficiency and user experience [41][42]. Group 5: Future of AI Experience - The future AI experience is expected to evolve towards a hybrid AI model, where efficient models run on the edge provide personalized services, while larger models in the cloud offer more powerful capabilities [55][57]. - Qualcomm aims to ensure seamless collaboration between edge and cloud environments through low-latency, high-speed, and secure connectivity [58].