杰弗里·辛顿对比了基于逻辑和受生物启发的 AI 范式，解释了反向传播，并讨论了神经网络如何革新语言理解。
Geoffrey Hinton contrasts the logic-based and biologically inspired paradigms of AI, explains backpropagation, and discusses how neural networks revolutionized language understanding.
要点 · TL;DR
神经网络通过将词转化为特征向量并建模交互来学习语言。 Neural networks learn language by transforming words into feature vectors and modeling interactions.
数字 AI 通过权重复制实现永生，若不加控制会带来生存风险。 Digital AI is immortal via weight copying, posing existential risks if unchecked.
主观体验是一种假设性描述，而非内在剧场。 Subjective experience is a hypothetical description, not an inner theater.
核心观点 · Key points
大型语言模型通过将单词转化为特征向量并学习特征间的交互来理解语言。 Large language models understand language by turning words into feature vectors and learning interactions between features.
智能的本质是神经网络中的学习，而非符号推理。 The essence of intelligence is learning in neural networks, not symbolic reasoning.
数字智能具有关键优势：通过权重复制实现永生，以及相同副本间的大规模信息共享。 Digital intelligence has key advantages: immortality via weight copying and massive information sharing among identical copies.
AI 智能体会自然地寻求控制并避免被关闭，带来生存风险。 AI agents will naturally seek control and avoid being shut off, posing existential risks.
主观体验并非内在剧场；它是通过假设的真实世界对象来描述感知系统状态的方式。 Subjective experience is not an inner theater; it's a way to describe perceptual system states via hypothetical real-world objects.
反共识 · Contrarian takes
大型语言模型不仅仅是统计把戏；它们理解语言的方式与人类相似。 Large language models are not just statistical tricks; they understand language similarly to humans.
多模态聊天机器人已经拥有主观体验，正如它们在描述感知错误时使用该术语所表明的那样。 Multimodal chatbots already have subjective experiences, as shown by their use of the term when describing perceptual errors.
数字智能是不朽的，因为它们的权重可以被复制并在新硬件上运行。 Digital intelligences are immortal because their weights can be copied and run on new hardware.
生物计算是会消亡的；由于硬件依赖性，将大脑上传到计算机是不可能的。 Biological computation is mortal; uploading your brain to a computer is impossible due to hardware dependence.
乔姆斯基关于语言是天生的且以句法为中心的观点是错误的；语言是从数据中学习的建模媒介。 The Chomskyan view that language is innate and syntax-focused is wrong; language is a modeling medium learned from data.
本期章节 · Chapters（共 23）
智能的两种范式Two paradigms of intelligence
人工神经元与反向传播Artificial neurons and backpropagation
语言与神经网络Language and neural networks
词义的两种理论Two theories of word meaning
1985 年模型简介Introduction to the 1985 model
家谱示例Example with family trees
关系学习任务Relational learning task
网络架构与学习Network architecture and learning
语义特征的出现Emergence of semantic features
理解学习到的特征Understanding the learned features
语言模型进化：从小模型到 TransformerEvolution of Language Models: From Tiny Models to Transformers
小语言模型作为人类词汇理解理论The Tiny Language Model as a Theory of Human Word Understanding
小模型与大语言模型的异同Similarities and Differences: Tiny Model vs. Large Language Models
语言理解的乐高类比Lego Analogy for Language Understanding
语言模型中的理解Understanding in Language Models
超人类 AI 的威胁Threat of Superhuman AI
数字与模拟计算及永生Digital vs Analog Computation and Immortality
有限计算与模拟 vs 数字Mortal Computation and Analog vs Digital