这期播客节目是 World Labs 的联合创始人李飞飞和贾斯汀·约翰逊与 A16Z 普通合伙人卡萨多的对话。他们深入探讨了人工智能的演变，特别是聚焦于在空间领域理解和生成新数据的转变，这最终促成了他们公司 World Labs 的成立。 李飞飞分享了她的经历，从她对物理学的背景开始，这激发了她对智能作为一种基本谜题的兴趣。她进入这个领域时，正值相对平静的时期，并见证了机器学习的兴起。一个关键的转折点是意识到数据（经常被忽视）可以驱动人工智能的泛化能力。这促成了 ImageNet 项目，该项目展示了大规模、数据驱动模型的力量，并重振了计算机视觉。 贾斯汀·约翰逊讨论了他在 2010 年代初进入人工智能领域的经历，这得益于深度学习的出现。他强调了计算能力的重要性，强调了计算能力的指数级增长。他还指出了两个关键的算法时代：有监督学习的时代，以 ImageNet 为例，数据由人类精心标记；以及随后的时代，算法从未标记或隐式标记的数据中学习。 对话过渡到生成式人工智能，近年来生成图像和其他内容的能力已经爆发。他们承认，虽然潜在的数学概念早已存在，但计算和数据的进步使得取得了显著的进展。贾斯汀讨论了他关于实时风格迁移的工作，这为生成式应用奠定了基础。李飞飞提到了贾斯汀关于从场景图生成图像的项目，作为早期的生成式人工智能。 讨论的核心集中在 World Labs 及其对空间智能的关注上。李飞飞解释说，空间智能是指机器在 3D 空间和时间中感知、推理和行动的能力。其目标是使机器超越简单的模式识别，并使其能够以有意义的方式理解和与物理世界交互。 演讲者澄清说，空间智能既包含物理世界，也包含世界的抽象概念。贾斯汀将空间智能与基于语言的方法进行对比，强调了底层表示的根本区别：语言模型在 1D 的 token 序列上运行，而空间智能模型则以世界的 3D 本质为中心。他和李飞飞强调，语言模型依赖于人类已经构建了实体与生成的上下文之间的关系。 讨论扩展到空间智能和 2D 视频之间的区别。即使人类通过 2D 图像感知世界，但包含 3D 表示可以更好地处理涉及对象操作、相机移动和其他空间交互的任务。他们认为，这种原生的 3D 表示对于涉及混合虚拟和物理环境的应用至关重要。 他们讨论了空间智能的潜在用例，包括世界生成（为游戏、教育等创建交互式 3D 虚拟世界）、增强现实和机器人技术。他们设想了这样一个未来：AR 设备可以实时理解周围环境，并提供上下文相关的的信息和帮助。他们强调，空间智能将成为 AR/VR/MixR 的“操作系统”，连接数字世界和物理世界。他们认为它可以减少对物理屏幕的需求，甚至使机器人能够在物理世界中执行任务。 演讲者承认其中涉及的技术挑战，并强调了建立一个跨学科团队的重要性，该团队需具备机器学习、计算机图形学、系统工程和 3D 理解等各个领域的专业知识。

This podcast episode features a conversation between Fei-Fei Li and Justin Johnson, co-founders of World Labs, and A16Z General Partner, Cassado. They delve into the evolution of AI, particularly focusing on the shift towards understanding and generating new data in the spatial realm, leading to the foundation of their company, World Labs. Fei-Fei Li shares her journey, starting from her physics background that sparked her interest in intelligence as a fundamental mystery. She entered the field during a period of relative quiet, witnessing the rise of machine learning. A crucial turning point was the realization that data, often overlooked, could drive generalization in AI. This led to the ImageNet project, which demonstrated the power of large-scale, data-driven models and revitalized computer vision. Justin Johnson discusses his entry into AI in the early 2010s, fueled by the emergence of deep learning. He emphasizes the importance of computational power, highlighting the exponential increase in computing capabilities. He also points out two key algorithmic epochs: the era of supervised learning, exemplified by ImageNet, where data was meticulously labeled by humans, and the subsequent era where algorithms learned from unlabeled or implicitly labeled data. The conversation transitions to generative AI, where the ability to generate images and other content has exploded in recent years. They acknowledge that while the underlying mathematical concepts were present earlier, the advancements in compute and data have enabled remarkable progress. Justin discusses his work on real-time style transfer, which laid the groundwork for generative applications. Fei-Fei mentions Justin's project on generating images from scene graphs as early generative AI. The core of the discussion centers on World Labs and its focus on spatial intelligence. Fei-Fei explains that spatial intelligence is about machines' ability to perceive, reason, and act in 3D space and time. The goal is to take machines beyond simple pattern recognition and enable them to understand and interact with the physical world in a meaningful way. The speaker clarifies that spatial intelligence encompasses both the physical world and abstract notions of world. Justin contrasts spatial intelligence with language-based approaches, highlighting the fundamental difference in underlying representations: language models operate on 1D sequences of tokens, while spatial intelligence models are centered around the 3D nature of the world. He and Fei-Fei emphasized that language models rely on humans already structuring the relationship between entities and the generated context. The discussion extends to the distinction between spatial intelligence and 2D video. Even though humans perceive the world through 2D images, incorporating a 3D representation allows for better handling of tasks involving object manipulation, camera movement, and other spatial interactions. They argue that this native 3D representation is crucial for applications that involve blending virtual and physical environments. They discuss potential use cases for spatial intelligence, including world generation (creating interactive 3D virtual worlds for gaming, education, etc.), augmented reality, and robotics. They envision a future where AR devices can understand the surrounding environment in real-time and provide context-aware information and assistance. They emphasize that spatial intelligence will be the "operating system" for AR/VR/MixR, bridging the digital and physical worlds. They posit it can deprecate the need for physical screens and even enable robots to perform tasks in the physical world. Acknowledging the technical challenges involved, the speaker stresses the importance of building a multidisciplinary team with expertise in various fields, including machine learning, computer graphics, systems engineering, and 3D understanding.