现有路径不通？OpenAI、亚马逊考虑改变大模型训练方式

Core Insights - The article discusses a potential shift in AI model training paradigms, moving away from the traditional "pre-training followed by fine-tuning" approach to a more task-specific training model that incorporates relevant data earlier in the process [1][3]. Group 1: Training Paradigm Shift - Leading researchers from OpenAI, Thinking Machines Lab, and Amazon are questioning the efficiency of the current training paradigm, advocating for a model that introduces task-specific data earlier to address inefficiencies and issues like the "brittle brain problem" [1][3]. - David Luan from Amazon argues that the current method of broad pre-training followed by fine-tuning is not always logical, suggesting that if the model's final use is known, task-relevant data should be included during pre-training [1][3]. - This shift could lead to a new era in AI development, moving from a "one-size-fits-all" model to specialized models built on different datasets, requiring stricter data selection from the outset [1][4]. Group 2: Emergence of Specialized Models - OpenAI is already showing signs of this specialization by routing ChatGPT queries to different models and developing dedicated models like GPT-5-Codex, reflecting the diverse needs of consumers [2][4]. - The proposed changes will necessitate early decisions on data inclusion, directly impacting the model's capabilities, such as prioritizing coding data over creative writing data for a programming assistant [4]. Group 3: Hardware Innovations and Capital Investment - Concurrently, hardware innovations are accelerating, with companies like Neurophos raising $110 million to develop photonic chips aimed at enhancing AI computational efficiency [5][6]. - OpenAI is also enhancing its infrastructure, with its custom inference chips in the final stages of production and significant progress on its $500 billion Stargate infrastructure project [6]. Group 4: Industry Dynamics and Mergers - The AI sector is witnessing active mergers and funding activities, such as the merger of Lightning AI and Voltage Park, valued at over $2.5 billion, and Yelp's acquisition of AI startup Hatch for $300 million [7]. - Major players like Apple are negotiating with Google to leverage cloud infrastructure for an updated Siri, while Nvidia's CEO is reportedly seeking to re-establish a foothold in China [7].