商汤科技大模型论坛上，阶跃星辰、第四范式和智谱等公司的相关负责人齐聚一堂；无问芯穹智能算力 生态论坛上，阶跃星辰、银河通用等企业的相关负责人同样在列。各方均围绕"模型之问"这一核心议 题，共同探讨大模型行业的协同创新之道。 在近日举行的2025世界人工智能大会上，记者发现大模型行业出现了一个有趣的现象：同行非但没有成 为冤家，反而相互站台、彼此助力。 如何打通"算力-数据-模型-应用"的最后一公里？这是摆在所有大模型公司面前的共同考题。 近半年来，大模型的发展逐步从OpenAI开创的预训练为主、监督学习为辅范式，转向显著提升推理能 力的强化学习范式。能否降低推理成本，成为决定大模型应用渗透率的关键因素。 正如阶跃星辰创始人、首席执行官姜大昕所言，"多开好省"是大模型应用落地的四大"黄金法 则"，"多"即多模态，"开"代表开源，"好"指的是模型的性能好，"省"则强调节省成本。 尽管各家选择的模式和路径不同，但在大模型时代，只有让每家公司充分发挥自身专长，形成协同合 力，才能真正打通从技术创新到产业应用的最后一公里，加速推动行业的繁荣发展。 模型厂商与芯片厂商通过联合创新，实现大模型和算力的双向价值最大化，被认为 ...

Core Insights - The 2025 World Artificial Intelligence Conference (WAIC) will take place from July 26 to 28 in Shanghai, focusing on three critical questions in AI: the mathematical question, the scientific question, and the model question, which aim to explore the essence of AI technology and its applications [3][4][5] Group 1: Event Overview - WAIC is a significant global event in the AI sector, promoting technological breakthroughs, industry integration, and deep dialogues on global governance [3] - The event will feature a forum titled "Boundless Love, Shaping the Future," hosted by SenseTime, focusing on the "model question" and its implications for AI technology [3][4] Group 2: Model Question Focus - The "model question" series aims to create a global platform for top researchers and technical experts to discuss the intrinsic issues of AI models, particularly the relationship between model generalization and underlying architecture [4] - The event will explore the integration of Transformer and non-Transformer architectures, addressing challenges such as semantic mismatches in multi-modal intelligence and optimizing performance-cost curves [5] Group 3: Global Collaboration and Innovation - The conference will gather leaders from academia and industry to discuss the future trends and development paths of large model technologies, focusing on obstacles to achieving higher-level intelligence [6] - Experts will engage in discussions on innovative solutions for model architecture and computational optimization, aiming to bridge the gap in multi-modal semantics and performance boundaries [6]