Jitendra Malik 探讨莫拉维克悖论、智能进化，以及为何当今 AI（如 GPT-4）尽管成就斐然，仍远不及人类智能。
Jitendra Malik discusses the Moravec paradox, the evolution of intelligence, and why today's AI, despite impressive feats like GPT-4, still falls short of human-like intelligence.
要点 · TL;DR
AI 必须复制人类智能的感觉运动基础，而不仅仅是语言。 AI must replicate sensory-motor foundations of human intelligence, not just language.
快速运动适应让机器人无需手动工程就能学习多样地形。 Rapid motor adaptation lets robots learn diverse terrains without manual engineering.
模拟通过摩尔定律和迭代改进推动机器人学进步。 Simulation drives robotics progress via Moore's Law and iterative improvement.
核心观点 · Key points
人类智能从感知运动技能进化而来，而不仅仅是语言；AI 必须复制这一基础。 Human intelligence evolved from sensory-motor skills, not just language; AI must replicate this foundation.
快速运动适应使机器人通过学习而非手动工程处理多样地形。 Rapid motor adaptation enables robots to handle diverse terrains via learning, not manual engineering.
模拟是机器人学进步的关键，借助摩尔定律并实现迭代改进。 Simulation is key for robotics progress, riding Moore's Law and enabling iterative improvement.
核心计算机视觉问题（识别、重建、重组）基本解决，但视觉语言和单图像 3D 仍有待解决。 Core computer vision problems (recognition, reconstruction, reorganization) are largely solved, but vision-language and 3D from single images remain open.
AI 研究受益于小科学（探索）和大科学（利用），两者互补。 AI research benefits from both small science (exploration) and big science (exploitation), complementing each other.
反共识 · Contrarian takes
无根基的语言模型缺失了 20%植根于感知运动经验的词汇。 Language models without grounding miss the 20% of vocabulary rooted in sensory-motor experience.
大脑在手之后进化；物理操作驱动了智能，而非相反。 The brain evolved after the hand; physical manipulation drove intelligence, not vice versa.
对抗样本揭示人类视觉使用自上而下的处理，不同于简单分类器。 Adversarial examples reveal that human vision uses top-down processing, unlike simple classifiers.
视觉中的基准测试最初遭到抵制，但后来成为科学进步的关键。 Benchmarking in vision was initially resisted but became essential for scientific progress.
AI 将增强而非取代人类工作，正如历史上计算机化所展现的。 AI will augment rather than replace human jobs, as seen historically with computerization.
本期章节 · Chapters（共 26）
Jitendra Malik 介绍Introduction of Jitendra Malik
当前 AI 与人类智能的局限Limitations of Current AI Compared to Human Intelligence
大脑发展紧随手部发展Brain Development Followed Hand Development
具身智能与语言基础Embodied intelligence and language grounding
五岁前的挑战Challenges before age five
研究议程与技能习得Research Agenda and Skill Acquisition
快速运动适应概念Rapid Motor Adaptation Concept
机器人适应性与学习 vs 工程robot adaptability and learning vs engineering
机器人学中的仿真Simulation in Robotics
盲视 vs 视觉行走Blind vs. Vision-Based Locomotion
四足 vs 双足行走Quadrupedal vs. Bipedal Locomotion
机器人学与运动Robotics and Locomotion
计算机视觉现状State of Computer Vision
对抗样本Adversarial Examples
计算机视觉基准测试动机Motivation for Benchmarking in Computer Vision
ImageNet 上的深度学习突破Deep Learning Breakthrough on ImageNet
AlexNet 与 GPU 因素AlexNet and the GPU factor
未来 AI 应用：医疗与养老Future AI applications: medicine and elder care
AI 与就业替代：历史视角AI and job displacement: a historical perspective
给 AI 博士生的建议Advice for PhD students in AI
印度早期生活与教育Early Life and Education in India
斯坦福博士：从逻辑到视觉PhD at Stanford and Switch from Logic to Vision
转至伯克利与建设 AI 项目Move to Berkeley and Building AI Program