World Models Through Yann LeCun’s Lens, with Xuefeng from Ami AI Lab ｜ 法国前沿AI Lab研究员雪峰从Yann Lecun的观点聊世界模型

​AI会说话之后，下一步是学会理解世界。

​过去几年，大语言模型让机器越来越擅长理解和生成语言，但如果人工智能真的要走进现实世界，它还需要学会另一件事：理解物体、空间、时间与因果，并对接下来可能发生什么做出预测。

​这也是“世界模型”正在尝试回答的问题。

​本次活动，我们邀请现任 AMI Labs 技术团队成员胡雪峰，与 Joey 展开一场关于世界模型的对谈。从计算机视觉、视频基础模型到今天的世界模型研究，雪峰长期关注一个核心问题：机器如何从感知信息出发，逐步形成对真实世界的理解。

​对谈中，我们会聊到世界模型到底是什么，它和今天的大语言模型有什么区别，为什么越来越多人认为“理解现实世界”会成为下一阶段人工智能的重要方向。也会从雪峰在字节跳动 Seed、TikTok Research、Meta、Amazon 等团队的研究经历出发，聊聊视频、多模态、基础模型这些技术是如何一步步走到今天的。

​我们也想把问题再往现实里拉近一点。

​如果机器开始理解空间、时间与因果，它会怎样影响机器人、具身智能、自动驾驶以及未来的智能决策？今天的 AI 距离真正“看懂世界”，究竟还有多远？

​这会是一场技术对谈，也是一场面向普通人的 AI 讨论。

​无需提前理解世界模型，我们希望从最简单的问题开始：

​AI到底什么时候，才能真正看懂我们生活的这个世界？

​Now that AI can speak, the next step is learning how to understand the world.

​Over the past few years, large language models have made machines increasingly capable of understanding and generating language. But if AI is going to truly operate in the physical world, it needs to learn something else: how to understand objects, space, time and causality, and how to predict what might happen next.

​This is exactly the question that “world models” are trying to answer.

​For this event, we are inviting Xuefeng Hu, currently a member of the technical team at AMI Labs, to join Joey for a conversation on world models. From computer vision and video foundation models to today’s research on world models, Xuefeng has long focused on one central question: how can machines move from perception toward a deeper understanding of the real world?

​During the conversation, we will explore what world models actually are, how they differ from today’s large language models, and why more and more researchers see “understanding the physical world” as an important next step for AI.

​Drawing on Xuefeng’s research experience at ByteDance Seed, TikTok Research, Meta and Amazon, we will also look at how developments in video, multimodal systems and foundation models have gradually led us to where we are today.

​We also want to bring the discussion closer to the real world.

​If machines begin to understand space, time and causality, how might that reshape robotics, embodied AI, autonomous driving and future intelligent decision making? And how far are today’s AI systems from truly “understanding” the world around us?

​This will be a technical conversation, but also an AI discussion designed for a broader audience.

​No prior knowledge of world models is required. We want to begin with the simplest question:

​When will AI truly be able to understand the world we live in?