听雨 发自 凹非寺 量子位 | 公众号 QbitAI 中国大模型顶流罕见同台！这一次是在中关村论坛上。 聚焦最热龙虾话题，月之暗面 杨植麟 、小米MiMo大模型负责人 罗福莉 、智谱AI CEO 张鹏 、无问芯穹联创兼CEO 夏立雪 ，以及香港大 学助理教授 黄超 齐聚一堂，共论agent的下一代演进。 张鹏 ：OpenClaw是 脚手架 ；模型从简单对话到真正能干活，背后消耗的token是 十倍甚至百倍 ，所以新模型涨价是回归正常商业价值。 夏立雪 ：从一月底开始，token调用量每两周翻一倍，到现在已增长十倍。 上次见到这个速度还是3G ，手机流量快速普及的时候；真正的 AGI时代到来，连基础设施本身也会是智能体。 黄超 ：未来很多软件可能 不再以人类为中心 ，因为人类需要GUI，但很多系统可能会越来越偏向agent-native，也就是原生面向agent使 用。 以下附上对话实录，为提升可读性，量子位在不改变原意的前提下做了适当调整。 这也是刚发布小米新模型的罗福莉，在公开论坛上的首次露面。 杨植麟还犀利提问智谱张鹏：智谱新模型怎么涨价了？ 各位大咖的观点相当硬核，信息密度极高： 罗福莉 ：中国大模型团 ...

Summary of Key Points from the Conference Call Industry Overview - The focus of the AI industry in 2025 is shifting towards the application layer, with significant changes expected in the latter half of the year, particularly in pre-training and post-training models [2][5][20]. Core Insights and Arguments - **AI Model Development**: The emphasis is moving from pre-training to post-training, with companies like OpenAI and Google leading the charge. Pre-training is expected to regain importance by the end of 2026, impacting computational power needs significantly [3][5][20]. - **Computational Power Demand**: Although no significant changes in computational power are anticipated this year, the overall demand is more optimistic than market expectations, particularly for the ASIC industry. Long-term demand will continue to grow due to increasing data and parameter volumes [3][4][6][32]. - **Dual Architecture Models**: The trend is towards dual architecture models (e.g., combining Transformer and GNN) to enhance model capabilities, which may become a consensus among major model manufacturers by the end of the year [9][10]. - **Synthetic Data Utilization**: The value of synthetic data is becoming more apparent, with a focus on increasing new data and improving the efficiency of existing data usage [12]. - **Reinforcement Learning**: It plays a crucial role in post-training, enhancing specific domain capabilities through repeated practice, although it is seen as less effective for overall model performance compared to pre-training [17][18][19]. - **Commercialization of AI**: The commercialization process is centered around "agents," with major manufacturers competing to enhance model capabilities and improve user experiences through engineering [8][20][22]. Additional Important Insights - **Challenges for Intelligent Agents**: Current intelligent agents face issues with task execution accuracy, which is critical for building reliable general AI systems [22][23]. - **China's Competitive Edge**: Chinese firms show relative advantages in engineering innovation, allowing them to respond quickly to market demands and develop competitive products [24]. - **Common Agent Platform (CAP)**: CAP provides shared tools and data for developers, lowering development barriers and promoting the penetration of agent technology [26][27]. - **Model Control Platform (MCP)**: MCP simplifies the agent development process, enabling broader participation in agent research and indirectly promoting technological advancement [28]. - **Key Companies to Watch**: OpenAI, Anthropic, and Google are pivotal in understanding future computational power demands and AI commercialization trends [36][37]. Market Dynamics - **Microsoft's Position**: Microsoft has seen a decline in its AI capabilities, affecting market perceptions of its computational power needs. The company is shifting focus from pre-training to inference, aligning with its commercial needs [34][35]. - **Overall Computational Demand**: The overall computational demand in 2025 is expected to be slightly better than market predictions, with a focus on enhancing model capabilities and meeting user expectations [38]. - **Investment Directions**: Investors should closely monitor developments from AAA-rated companies, as significant changes are anticipated in the second and third quarters of 2025 [40]. This summary encapsulates the key points discussed in the conference call, highlighting the evolving landscape of the AI industry and the strategic focus of major players.