杨立昆以《2001 太空漫游》中的 HAL 9000 为例，探讨 AI 价值对齐问题，指出目标函数设计不当会导致灾难，并将 AI 目标函数设计类比于人类法律体系。
Yann LeCun discusses AI value alignment, using HAL 9000 from 2001: A Space Odyssey as an example of misaligned objectives, and compares designing AI objective functions to human legal systems.
要点 · TL;DR
AI 对齐是关于价值错位，而不仅仅是能力。 AI alignment is about value misalignment, not just capability.
自监督学习像婴儿一样通过观察构建世界模型。 Self-supervised learning builds world models like babies learn.
当前 AI 缺乏常识，需要超越规模的新思路。 Current AI lacks common sense; new ideas beyond scaling are needed.
核心观点 · Key points
智能与学习不可分割；学习是智能的自动化。 Intelligence is inseparable from learning; learning automates intelligence.
自监督学习是构建世界模型的关键，如同婴儿通过观察学习。 Self-supervised learning is key to building world models, like babies learn by observation.
基于模型的强化学习结合预测性世界模型能实现高效学习。 Model-based reinforcement learning with predictive world models enables efficient learning.
在物理世界中的具身化对于真正的语言理解和常识是必要的。 Grounding in the physical world is necessary for true language understanding and common sense.
当前 AI 缺乏常识；解决这一问题需要超越 Scaling 的新思路。 Current AI lacks common sense; solving this requires new ideas beyond scaling.
反共识 · Contrarian takes
人类智能并非通用，而是高度特化的，这与普遍认知相反。 Human intelligence is not general; it is highly specialized, contrary to common belief.
深度学习尽管违反教科书规则（如过参数化、非凸性）却依然有效。 Deep learning works despite violating textbook rules like overparameterization and non-convexity.
主动学习并非变革性方法；自监督学习更为根本。 Active learning is not transformative; self-supervised learning is more fundamental.
具身并非智能的必要条件，但世界中的具身化是必要的。 Embodiment is not necessary for intelligence, but grounding in the world is.
情感对智能至关重要；它们源于对未来满足感的预测。 Emotions are essential for intelligence; they arise from predicting future contentment.
本期章节 · Chapters（共 17）
引言与哈尔 9000 价值对齐问题Introduction and Hal 9000 Value Misalignment
对齐与法律代码作为目标函数Alignment and Legal Code as Objective Functions
深度学习中的惊人想法Surprising ideas in deep learning
知识表示与推理Knowledge Representation and Reasoning
因果关系与常识Causality and Common Sense
90 年代神经网络为何失宠Why Neural Nets Lost Interest in the 1990s
早期软件工具与 LispEarly Software Tools and Lisp
卷积网络专利Patent on Convolutional Networks
专利与卷积网络商业化On patents and the commercialization of ConvNets
基准测试与想法验证On benchmarks and testing ideas
人类智能是专长而非通用Human intelligence is specialized, not general
NLP 与视觉中的自监督学习Self-supervised learning in NLP vs vision
当前 AI 局限与世界模型需求Limitations of Current AI and Need for World Models
少样本与迁移学习基准Benchmark for Few-Shot Learning and Transfer Learning
深度学习与自动驾驶Deep Learning and Autonomous Driving
当前自动驾驶方法Current autonomous driving approaches
索菲亚机器人批判与 AI 炒作Critique of Sophia the Robot and Hype in AI