CICC: Internet AI Layout Starts Catch-up Mode, Focuses on Agent in Key Application Directions.

CICC released a research report stating that the large AI models have already learned enough from existing data, and the remaining new training data may be insufficient, leading to a slowdown in pre-training of AI large models. As training methodologies become more mature, the gap in computational power narrows, and talent mobility increases, leading companies with strong financial resources and capabilities are expected to narrow the gap with leaders and even lead them. Looking ahead to 2025, AI-related companies are expected to focus on Agent as the key application direction. Key points from CICC: Slowdown in pre-training of AI large models Pre-training is the process through which large models learn general features and acquire the basis of intelligence. However, since the second half of 2024, there have been signs of a slowdown in pre-training of large models, mainly because these models have already learned enough from existing data, and the remaining new training data may be insufficient. The highly anticipated release of OpenAI's next-generation flagship model, GPT-5, has been postponed, reportedly due to lower-than-expected training effectiveness, high training costs, and lack of significant capability improvement. Competition in large AI models may gradually favor followers instead of leaders If the development speed of industry leaders slows down, followers may benefit. In the wave of large AI models, companies such as OpenAI, Claude, etc., are considered industry leaders, while Alphabet, Meta, Amazon, etc., are often considered followers. In China, companies like Moon's Dark Side, Minimax, etc., are considered industry leaders, with leaders seen as followers. In the comparison between the development of AI in China and the United States, American AI companies are seen as model leaders, while Chinese companies are seen as followers. Agent leading AI application directions, strong competition among Chinese internet companies Currently, AI application directions include dialogue systems, programming assistants, office efficiency tools, education, entertainment, empowering traditional businesses, etc. Looking ahead to 2025, AI-related companies will focus on Agent as the key application direction, such as Apple's Apple Intelligence assistant, Alphabet's Gemini 2.0 Project Astra, Salesforce's AgentForce for enterprise users, etc. The success of the supply-side dividend depends on whether the product can meet user needs, control costs, and sustain a business model. Compared to being more chasing in terms of models, Chinese internet companies are more competitive in AI application directions compared to their American counterparts. Risks: Risks of AI productization and commercialization falling short of expectations; external environmental risks.