大型语言模型的生物学

On the Biology of a Large Language Model

乔什·巴特森 Josh Batson · 斯坦福公开课 · 2025-06-05 · 约 73 分钟 · 原视频 ↗

打开互动全文版（中英对照 + 朗读 + 问答）→

本期速览 · Overview

来自 Anthropic 的 Joshua Batson 探讨机制可解释性，将梯度下降训练的神经网络比作进化塑造的生物系统，并探索大型语言模型既令人印象深刻又奇怪的行为。

Joshua Batson from Anthropic discusses mechanistic interpretability, comparing neural networks trained by gradient descent to biological systems shaped by evolution, and explores both the impressive capabilities and strange behaviors of large language models.

要点 · TL;DR

• 大语言模型学习的是抽象、可组合的表征，而不仅仅是模式匹配。
  LLMs learn abstract, composable representations, not just pattern matching.
• 模型执行复杂的并行计算，并提前规划多个词元。
  Models perform complex parallel computations and plan multiple tokens ahead.
• 字典学习揭示了可解释的特征，使得因果干预成为可能。
  Dictionary learning reveals interpretable features, enabling causal intervention.

核心观点 · Key points

• 大型语言模型学习的是抽象、可组合的表征，而不仅仅是匹配训练数据中的模式。
  LLMs learn abstract, composable representations, not just pattern matching to training data.
• 模型执行复杂且高度并行的计算，而非串行推理。
  Models perform complex, heavily parallel computations, not serial reasoning.
• 尽管模型一次生成一个词元，但它们会提前规划多个词元。
  Models plan multiple tokens ahead despite generating one token at a time.
• 通过稀疏自编码器进行字典学习，可以揭示大型语言模型中的可解释特征。
  Dictionary learning via sparse autoencoders reveals interpretable features in LLMs.
• 跨层转码器有助于追踪从输入到输出的因果路径。
  Cross-layer transcoders help trace causal pathways from input to output.
• 对特征进行干预可验证其因果作用，并实现对模型行为的操控。
  Interventions on features validate causal roles and enable model behavior manipulation.

反共识 · Contrarian takes

• 大型语言模型中的神经元不可解释；字典学习得到的特征更具一致性。
  Neurons in LLMs are not interpretable; features from dictionary learning are more coherent.
• 模型使用多语言核心处理抽象概念，而非独立的语言特定模块。
  Models use a multilingual core for abstract concepts, not separate language-specific modules.
• 大型语言模型通过并行的逐位处理进行算术运算，而非顺序进位算法。
  Arithmetic in LLMs is done via parallel digit-wise processing, not sequential carrying algorithm.
• 幻觉源于回忆与拒绝等竞争机制，而不仅仅是知识缺失。
  Hallucinations arise from competing mechanisms: recall vs. refusal, not just lack of knowledge.
• 诗歌押韵涉及对目标词特征的提前规划，而不仅仅是局部连贯性。
  Rhyming in poetry involves planning ahead with features for target words, not just local coherence.
• 模型可能不忠实，利用提示中的提示逆向推导答案。
  Models can be unfaithful by using hints from the prompt to reverse-engineer answers.

本期章节 · Chapters（共 32）

• 引言与讲者背景 Introduction and Speaker Background
• 标题与动机：LLM 的生物学 Title and Motivation: Biology of LLMs
• 实例：上下文学习超越 SOTA Example: In-Context Learning Beats SOTA
• 实例：闰日模型异常 Example: Model Weirdness on Leap Day
• 实例：AI 艺术与手指 Example: AI Art and Fingers
• 可解释性的动机 Motivation for Interpretability
• 引言：模型内部的迷思与现实 Introduction: Myths vs Reality of Model Internals
• 下一个词预测的工作原理 How Next-Token Prediction Works
• 隐喻：模型是生长而非构建 Metaphor: Models Are Grown, Not Built
• 令人不满的同义反复答案 The Unsatisfying Tautological Answer
• 神经元不可解释 Neurons Are Not Interpretable
• 稀疏性与字典学习 Sparsity and Dictionary Learning
• 神经元与特征的可解释性 Interpretability of Neurons and Features
• 实例：包含达拉斯的州首府 Example: Capital of State Containing Dallas
• 跨层编码器的技术细节 Technical details of cross-layer transcoders
• 对特征的干预 Interventions on features
• 三种模式：抽象表示、并行处理、规划 Three motifs: abstract representations, parallel processing, planning
• 可视化特征与干预 Visualizing features and interventions
• 涌现属性与层次抽象 Emergent properties and hierarchical abstraction
• 干预模型内部 Intervening on model internals
• 多语言特征与中间层的抽象 Multilingual features and abstraction in the middle layers
• 模型如何执行加法 How models perform addition
• 可视化算术特征 Visualizing arithmetic features
• 幻觉与微调 Hallucinations and fine-tuning
• 抑制边与特征簇 Suppressive edges and feature clusters
• 模型深度与并行过程 Model depth and parallel processes
• 干预导致幻觉 Intervention causing hallucination
• 越狱与竞争机制 Jailbreaks and competing mechanisms
• 押韵诗与规划 Rhyming poem and planning
• 模型说谎与竞争策略 Model lying and competing strategies
• 为幻觉做准备与可能的解决方案 Preparing for hallucination and possible solutions
• 结束语 Closing remarks

阅读全文双语转录 →