发展大模型要摒弃短视冒进

Core Viewpoint - The article emphasizes that the evaluation of large models in the AI era should not be based solely on usage rates and traffic, but rather on the depth of technological accumulation and ecological collaboration [1][2]. Group 1: Transition in Evaluation Criteria - The shift from "traffic supremacy" to "technology victory" reflects a fundamental change in the logic of competition in the AI era [1]. - In the internet age, product competition was driven by quick user acquisition and traffic accumulation, but large model competition is fundamentally different, focusing on hard technology metrics and model performance [1][2]. Group 2: Ecosystem Development - DeepSeek has established a vast ecological network by integrating with various cloud service providers, search platforms, smart terminals, and industry applications, which enhances its strategic value beyond mere usage rates [2]. - The open API and training framework of DeepSeek allow developers to quickly build vertical applications, fostering a win-win situation and leading to a surge of AI applications across various industries [2]. Group 3: Future Innovations - The upcoming DeepSeek-R2 is expected to feature a collaborative optimization design of "model + hardware," indicating a shift towards parallel development of software and hardware for large models [2]. - Breakthroughs in hardware, such as Huawei's Ascend 384 super nodes and Shanghai AILab system platform, are laying a solid foundation for the next generation of AI systems [2]. Group 4: Long-term Focus - The article warns against the "traffic anxiety" and "hit mentality" that can hinder innovation in the high-cost, rapidly evolving AI landscape, advocating for a focus on deep technological development and ecosystem building [3].