【聊了什么 The What】
本期节目邀请到前 Meta 人工智能研究院高级经理田渊栋，与哥伦比亚大学商学院助理教授彭天翼，共同探讨 2026 年 AI 的核心命题：个性化。从冯·诺依曼架构的‘存算分离’到大模型的‘存算一体’，我们深入拆解了 AI 如何通过显性与隐性记忆构建个人数字孪生。我们还特别针对学术界与工业界的‘地位倒挂’、端侧 AI 的隐私平衡以及 AI for Science 等热门话题进行了全方位碰撞。当下的 AI 究竟是通才工具，还是具备独特审美与思维方式的个人助理？在 AI 时代，人类如何通过提出‘宏大愿景’来定义自我价值？
In this episode, we are joined by Yuandong Tian, former Senior Manager at Meta AI Research, and Tianyi Peng, Assistant Professor at Columbia Business School, for a wide-ranging conversation on one of the defining questions of AI in 2026: personalization.
From the “separation of memory and computation” in the von Neumann architecture to the emerging “fusion of memory and computation” in large models, we unpack how AI may build a personal digital twin through both explicit and implicit forms of memory. We also dive into a range of major topics, including the role reversal between academia and industry, the privacy tradeoffs of on-device AI, and the promise of AI for Science. Is today’s AI simply a general-purpose tool, or can it become a personal assistant with its own distinctive aesthetic and mode of thinking? In the age of AI, how can humans define their value through the ability to articulate bold visions?
【时间轴 The When 】
00:00 - 老赵介绍嘉宾田渊栋与彭天翼
09:50 - 大模型中的存算一体与冯诺依曼架构
20:40 - AI的记忆困境：显式记忆与参数化记忆的路径之争
39:38 - 语言是智能的上限吗？处理非语言信息的挑战
48:51 - 实验趣闻：为什么模拟出的AI学生往往比真人更“聪明”
51:51 - 执行与反馈：验证者定律及强化学习中的奖励机制
72:40 - 学术界与工业界在AI浪潮中的角色转变
96:32 - 端侧AI（On-device Agent）的前景与隐私保护
00:00 – Zhao introduces guests Yuandong Tian and Tianyi Peng
09:50 – The fusion of memory and computation in large models, and the von Neumann architecture
20:40 – AI’s memory dilemma: explicit memory vs. parametric memory
39:38 – Is language the ceiling of intelligence? The challenge of processing non-linguistic information
48:51 – A research anecdote: why simulated AI students often seem “smarter” than real humans
51:51 – Execution and feedback: the verifier’s law and reward mechanisms in reinforcement learning
72:40 – The shifting roles of academia and industry in the AI wave
96:32 – The future of on-device AI agents and privacy protection
【我们是谁 The Who】
选修课 Universus 是一档分享和探讨泛文化话题的中文播客，脱胎于2013年创办的纽约文化沙龙。 我们力求引发对学科方法和视角的认知和思考，而不仅仅是知识的传递。我们在组织活动的过程中积累了一些经验和人脉，尝试打造成年人的博雅教育，并在播客中拓展这样的探索。我们希望审视和反抗无意识的价值规训与随波逐流。我们希望冷酷地剖析自己的偏见，热忱地拥抱多元视角。我们希望去工具化，真诚地感知、理解、创造自己的生活和复杂的世界。我们追随原初的好奇而非代餐式的猎奇。话题可以轻松，方法力求严肃。在《选修课》里，我们互相映照，唤醒失落的维度。
选修课 Universus is a Chinese-language podcast created by the same people who founded the New York Chinese Cultural Salon (纽约文化沙龙）in 2013. The past decade of organizing events and talks put us in a position of creating a sort of liberal arts education experience for those who have left school. We strive to go beyond merely transmitting knowledge to reflecting on complex topics using cross-disciplinary perspectives. We strive to resist conformity, to dissect our own biases, and to embrace diverse schools of thought. Our goal is to help professionals resist becoming just a tool of production, and to understand the world around us in earnest. We let our genuine curiosity be our guide. With a light-hearted tone, we approach each topic seriously. We hope that “选修课 Universus” creates a space to explore lost dimensions.
嘉宾：
田渊栋 博士：前Meta人工智能研究院研究员及高级经理
彭天翼 博士：哥伦比亚大学商学院决策风险与运筹系助理教授
Dr. Yuandong Tian: Former Research Scientist and Senior Manager at Meta AI Research
Dr. Tianyi Peng: Assistant Professor of Decision, Risk, and Operations at Columbia Business School
【买咖啡 Please Support Us】
如果喜欢这期节目并愿意想要给我们买杯咖啡：
- 打赏平台：www.patreon.com/universuspod
- 商务合作邮箱：[email protected]
If you like our show and want to support us, please consider the following:
- 打赏平台: www.patreon.com/universuspod
- Business Inquiries Email: [email protected]
