Core Insights - The article emphasizes a paradigm shift in the tech industry, where traditional roles such as front-end, back-end, and full-stack developers are being replaced by AI Agent engineers, marking a significant transformation akin to an industrial revolution [1][2] - Demis Hassabis, CEO of Google DeepMind, predicts that the scale of this change will be ten times that of the industrial revolution, with a speed that is also ten times faster [1] - AI is evolving from a mere enabling tool to a transformative force in business processes and organizational paradigms, as concluded by the Singularity Intelligence Research Institute after surveying over 100 companies [1] Event Overview - The "Global Machine Learning Technology Conference" has been upgraded to the "Singularity Intelligence Technology Conference" to reflect the rapid advancements in AI technology [2] - The 2026 Singularity Intelligence Technology Conference will take place in Shanghai on April 17-18, 2026, featuring over 50 leading technology figures and more than 1,000 elite attendees from various industries [3] Conference Themes - The conference will focus on the core logic of scaling AI from mere technological breakthroughs to practical applications, specifically how the Agent paradigm can drive business growth and ensure a positive ROI from computational investments [5] - Twelve key topics have been established for the conference, including: - Evolution of large language model technology - Multimodal and world models - AI computing platforms and performance optimization - AI-native software development and ambient programming - Intelligent agent systems and engineering - AI-native application innovation and development practices - Agent-enabled DevOps - Large model system architecture [5][6] Expert Contributions - The conference will feature expert speakers with deep expertise and practical experience in AI, ensuring that discussions are grounded in real-world applications and engineering truths [7][8] - Notable speakers include: - Duan Nan, Vice President of JD Group, with extensive experience in multimodal foundational models [11] - Li Yongbin, Head of Dialogue Intelligence & Code Intelligence at Alibaba, focusing on large model technologies [14] - Wang Shengjie, Head of AI Products at Tencent Cloud, with a background in software architecture and AI development efficiency [19] - He Wanqing, Vice President of Qingcheng Jizhi, specializing in HPC and AI application performance optimization [20] Call to Action - The conference invites participants who are leading teams in AI-native software development, multimodal world models, embodied intelligence, or AI infrastructure performance optimization to attend [62] - The event aims to foster collaboration among long-term thinkers in the AI era, creating a platform for sharing verifiable and reusable engineering experiences [65]

吴翼： RL 是机器学习这个大概念下一类比较特殊的问题。 曲凯： 今天我们请来了国内强化学习 (RL) 领域的专家吴翼，吴翼目前是清华大学交叉信息研究院 助理教授，他曾经在 OpenAI 工作过，算是国内最早研究强化学习的人之一，我们今天就争取一 起把 RL 这个话题给大家聊透。 首先吴翼能不能简单解释一下，到底什么是 RL？ 传统机器学习的本质是记住大量标注过正确答案的数据对。 举个例子，如果你想让机器学习能分辨一张图片是猫还是狗，就要先收集 10000 张猫的照片和 10000 张狗的照片，并且给每一张都做好标注，让模型背下来。 上一波人工智能四小龙的浪潮其实都以这套框架为基础，主要应用就是人脸识别、指纹识别、图 像识别等分类问题。 这类问题有两个特点，一是单一步骤，比如只要完成图片分辨就结束了；二是有明确的标准答 案。 但 RL 很不一样。 RL 最早是用来打游戏的，而游戏的特点和分类问题有两大区别。 第一，游戏过程中有非常多的动作和决策。比如我们玩一个打乒乓球的游戏，发球、接球、回 球，每一个动作都是非标的，而且不同的选择会直接影响最终的结果。 第二，赢得一场游戏的方式可能有上万种，并没有唯一的标准答 ...