心 智 七 篇 · Seven Mental Models
Wiki · Knowledge Atlas

知识图谱

每读一篇论文、博客、设计文档,都被消化成结构化条目;概念、实体、源头相互索引,自动生长。

134 概念 47 实体 82 源头

01 展陈 — 181

概念

A2A 协议

Agent2Agent Protocol

跨框架 Agent 通信的开放标准

  • concepts/agent-card
  • concepts/agent-interoperability
  • concepts/agentic-systems

Google/Linux Foundation · 2025

概念

ACI — Agent 计算机接口

Agent-Computer Interface

为模型设计接口,如同为人类设计 HCI

  • concepts/tool-design
  • concepts/augmented-llm
  • concepts/agent-native-software

Anthropic Building Effective Agents; Karpathy YC 2025

概念

动作蜕变关系

Action Metamorphic Relations

语义等价变换 · 终态判等 · 可测量的鲁棒性

  • concepts/reliability-surface
  • concepts/guardrails

ReliabilityBench · arXiv 2601.06112

概念

激活干预

Activation Intervention

修改中间层激活 · 验证表征的因果性角色

  • concepts/probing-classifiers
  • concepts/mechanistic-interpretability
  • concepts/linear-representation-hypothesis

Li et al. 2022 · Nanda et al. 2023 · OthelloGPT

概念

Agent Card

Agent Capability Discovery

Agent 的数字名片 · /.well-known/agent-card

  • concepts/a2a-protocol
  • concepts/agent-interoperability

Google A2A Protocol · 2025

概念

Agent 互操作

Agent Interoperability

从孤岛到网络 · 跨框架 Agent 协作

  • concepts/a2a-protocol
  • concepts/agent-card
  • concepts/orchestrator-workers

Google A2A Protocol · ASPLOS 2026

概念

Agent 原生软件

Agent-Native Software

以 Agent 为一等用户类型设计软件

  • concepts/aci
  • concepts/tool-design
  • concepts/agent-skills

CLI-Anything · Anthropic Agent Skills

概念

Agent OS

Agent Operating System

原生理解 Agent 语义的操作系统抽象

  • concepts/agent-resource-control
  • concepts/agent-sandboxing
  • concepts/harness-engineering

AgenticOS Workshop ASPLOS 2026 · AIOS COLM 2025

概念

Agent 可靠性评估

Agent Reliability Evaluation

pass@1 ≠ 可靠性 · 四指标框架

  • concepts/reliability-decay
  • concepts/long-running-agents
  • concepts/evaluator-optimizer

Beyond pass@1 · arXiv 2603.29231

概念

Agent 资源控制

Agent Resource Control

扩展 cgroup 到 Token · API 配额 · 工具权限

  • concepts/agent-os
  • concepts/context-management
  • concepts/harness-engineering

AgentCgroup · AgenticOS Workshop ASPLOS 2026

概念

Agent 沙箱

Agent Sandboxing

动态代码 · 语义攻击面 · 多层纵深防御

  • concepts/guardrails
  • concepts/agent-os
  • concepts/harness-engineering

AgenticOS Workshop ASPLOS 2026 · Anthropic Claude Code

概念

Agent 调度

Agent Scheduling

多 Agent 共享 LLM 资源的系统层仲裁

  • concepts/agent-os
  • concepts/llm-os-analogy
  • concepts/context-management

AIOS · arXiv 2403.16971

概念

Agent Skills

Agent Skills Standard

领域知识的可发现打包单元 · SKILL.md 为核心

  • concepts/aci
  • concepts/harness-engineering
  • concepts/agent-os

Anthropic Agent Skills 开放标准 · 2025

概念

Agentic 系统

Agentic Systems

Workflows vs Agents · 从单次调用到自主决策

  • concepts/augmented-llm
  • concepts/harness-engineering
  • concepts/orchestrator-workers

Anthropic Building Effective Agents · 2025

概念

AI 三个夏天

Three AI Summers

Kautz 历史周期论 · 神经与符号的共同演化

  • concepts/neurosymbolic-ai
  • concepts/bitter-lesson
  • concepts/physical-symbol-system

Kautz AAAI Engelmore Lecture · AI Magazine 43(1) 2022

实体

AIOS

AIOS — LLM Agent Operating System

Rutgers 2024 · 六大内核模块 · LLM-OS类比最完整实现 · COLM 2025

  • concepts/llm-os-analogy
  • concepts/harness-engineering
  • entities/memgpt

Mei et al. arXiv:2403.16971 COLM 2025

概念

Aitia — 原因与解释

Aitia (Aristotle)

希腊哲学 · 原因即解释 · 四因说的基础

  • concepts/four-causes
  • concepts/teleological-explanation

Stanford Encyclopedia · Aristotle on Causality

实体

Allen Newell

Allen Newell

1927–1992 · 图灵奖1975 · PSSH · GPS · SOAR统一认知理论

  • concepts/physical-symbol-system
  • concepts/means-ends-analysis
  • concepts/problem-space

Newell & Simon 1976 Comm ACM · Newell 1990 Unified Theories

实体

Andrej Karpathy

Andrej Karpathy

Tesla AI总监 · OpenAI研究员 · Eureka Labs · LLM-OS架构思想 · Software 3.0

  • concepts/llm-os-analogy
  • concepts/software-3-0
  • concepts/system-1-vs-system-2

Karpathy 2023 Intro to LLMs · 2025 YC AI Startup School

实体

Answer.AI

Answer.AI

Jeremy Howard创立 · llms.txt标准 · fast.ai生态延伸 · 开放实用AI工具

  • concepts/llms-txt
  • entities/jeremy-howard

Answer.AI / AnswerDotAI/llms-txt

实体

Anthropic

Anthropic

Claude系列 · 可解释性研究领先 · Agent SDK · MCP · Managed Agents

  • entities/claude-code
  • entities/claude-agent-sdk
  • concepts/mechanistic-interpretability

Anthropic Official Docs 2024-2026

实体

亚里士多德

Aristotle

公元前384-322 · 四因说 · 目的论解释 · 知识即因果知识 · 西方科学方法论奠基

  • concepts/four-causes
  • concepts/teleological-explanation
  • entities/herbert-simon

SEP — Aristotle on Causality (2006/2023)

实体

Artur d'Avila Garcez

Artur d'Avila Garcez

City University London · 神经符号计算 · Kautz分类法 · XAI忠实性批判 · 第三浪潮

  • concepts/neurosymbolic-ai
  • concepts/neurosymbolic-ai-taxonomy
  • concepts/knowledge-extraction-fidelity

Garcez & Lamb 2023 Neurosymbolic AI The 3rd Wave arXiv:2012.05876

实体

ASPLOS

ASPLOS

计算机系统顶会 · AgenticOS Workshop 2026 · Agent基础设施进入系统研究议程

  • concepts/agent-os
  • concepts/agent-sandboxing
  • entities/anthropic

AgenticOS Workshop ASPLOS 2026

概念

增强型 LLM

Augmented LLM

检索 + 工具 + 记忆 · Agentic 系统的基础构建块

  • concepts/context-engineering
  • concepts/agentic-systems
  • concepts/tool-design

Anthropic Building Effective Agents · 2025

概念

自主度滑块

Autonomy Slider

从 Tab 补全到全 Repo Agent · 部分自主产品设计

  • concepts/agentic-systems
  • concepts/harness-engineering
  • concepts/generation-verification-loop

Karpathy YC Software is Changing Again · 2025

实体

华严经

Avatamsaka Sutra

法界缘起 · 因陀罗网 · 须弥芥子 · 一即一切 · 分形哲学先驱

  • concepts/indra-net
  • concepts/sumeru-mustard-seed
  • concepts/fractal-architecture

华严经 / Cook 1973 Jewel Net of Indra

概念

后门准则

Backdoor Criterion

图形化去混淆判据 · Pearl 后门调整公式

  • concepts/causal-dag
  • concepts/do-calculus
  • concepts/frontdoor-criterion

Pearl Causality 2009 · Pearl 2010 Intro to Causal Inference

概念

启发式规则集合

Bag of Heuristics

局部 · 位置特异 · 无统一算法的世界模型

  • concepts/othello-world-model-hypothesis
  • concepts/mechanistic-interpretability
  • concepts/probing-classifiers

jylin04 MATS 6.0 · OthelloGPT 机制可解释性分析

概念

贝叶斯归纳

Bayesian Induction

先验 → 后验更新 · 归纳的概率化方案

  • concepts/induction-problem
  • concepts/uniformity-principle
  • concepts/scaling-laws

Bayes 1764 · Laplace 1814 · SEP Problem of Induction

实体

Benoit Mandelbrot

Benoit Mandelbrot

分形几何学创始人 · fractal造词1975 · 海岸线论文1967 · 云不是球体

  • concepts/statistical-self-similarity
  • concepts/richardson-effect
  • concepts/fractal-dimension

Mandelbrot 1967 Science · 1982 The Fractal Geometry of Nature

概念

苦涩的教训

The Bitter Lesson

通用计算终将胜出 · 不要内建你发现了什么

  • concepts/harness-engineering
  • concepts/meta-harness
  • concepts/scaling-laws

Rich Sutton 2019 · Anthropic Managed Agents

概念

因果语境主义

Causal Contextualism

因果判断是否依赖评估语境?三种立场

  • concepts/counterfactual-theory
  • concepts/causal-models
  • concepts/causal-relata

SEP Counterfactual Theories of Causation · Schaffer 2005

概念

因果有向无环图

Causal DAG

链 / 叉 / 对撞 · d-分离 · 去混淆的桥

  • concepts/structural-causal-model
  • concepts/do-calculus

参 Pearl (2009); Book of Why (2018)

概念

因果内积

Causal Inner Product

因果可分离的概念在此内积下正交

  • concepts/linear-representation-hypothesis
  • concepts/probing-classifiers
  • concepts/mechanistic-interpretability

Park, Choe, Veitch 2023 · arXiv 2311.03658

概念

因果干预

Causal Intervention

do 算子 · 切断原有输入 · 图上的微型手术

  • concepts/do-calculus
  • concepts/causal-dag
  • concepts/structural-causal-model

Pearl Causality 2009 · Pearl 2010 Intro

概念

因果模型

Causal Models

结构方程 · 干预语义 · 模块性

  • concepts/structural-causal-model
  • concepts/causal-dag
  • concepts/interventionist-theory

SEP Metaphysics of Causation · Pearl 2000

概念

因果关系项

Causal Relata

原因和结果各是什么类型的实体?

  • concepts/causal-models
  • concepts/counterfactual-theory

SEP Metaphysics of Causation

概念

因果性(休谟)

Causation (Hume)

必然联系不可观察 · 习惯是因果推理的真正基础

  • concepts/constant-conjunction
  • concepts/necessary-connection
  • concepts/regularity-theory

David Hume Enquiry 1748 · SEP David Hume

概念

Agent 混沌工程

Chaos Engineering for Agents

故障注入 · 可靠性退化测量 · 生产压力模拟

  • concepts/guardrails
  • concepts/reliability-surface
  • concepts/error-cascade

ReliabilityBench · arXiv 2601.06112

实体

Chroma

Chroma

开源向量数据库 · Context Rot系统研究 · 18模型测试 · RAG vs长上下文实证

  • concepts/context-rot
  • concepts/context-management
  • concepts/virtual-context-management

Chroma Research Context Rot 2025

实体

Claude Agent SDK

Claude Agent SDK

「给agent一台计算机」· 隐式循环架构 · Context Compaction · 长时运行需要外部化追踪

  • concepts/harness-engineering
  • concepts/implicit-loop-architecture
  • concepts/context-management

Anthropic Building Agents Claude Agent SDK 2024

实体

Claude Code

Claude Code

Anthropic官方编码agent CLI · 三级工具审批 · deny-first规则 · 五级作用域 · 六种权限模式

  • concepts/claude-code-permission-system
  • concepts/permission-modes
  • concepts/agent-sandboxing

Anthropic Claude Code Permissions 2024

概念

Claude Code 权限系统

Claude Code Permission System

三级工具分类 · allow/ask/deny 规则 · 纵深防御

  • concepts/permission-rules-hierarchy
  • concepts/permission-modes
  • concepts/agent-sandboxing

Anthropic Claude Code Permissions · 2026

实体

CLI-Anything

CLI-Anything

HKUDS 2026 · 29K stars · 任意软件→agent原生CLI · 7阶段流水线 · CLI-Hub

  • concepts/tool-design
  • concepts/agent-native-software
  • concepts/repl-for-agents

HKUDS/CLI-Anything github.com 2026

概念

CLI vs GUI 自动化

CLI vs GUI Automation

Agent 驱动软件的两条根本路线

  • concepts/agent-native-software
  • concepts/aci

CLI-Anything · Anthropic computer-use

概念

海岸线悖论

Coastline Paradox

测量单位越小,长度越大 · 分形维数的直觉入口

  • concepts/fractal-dimension
  • concepts/statistical-self-similarity

Mandelbrot 1967 · Wikipedia Coastline Paradox

实体

Codex

Codex (OpenAI)

OpenAI编码agent · codex-1(o3 RL微调) · 6小时+运行 · 零人工百万行代码 · Harness Engineering验证

  • concepts/harness-engineering
  • concepts/implicit-loop-architecture
  • entities/claude-agent-sdk

OpenAI Codex 2025 · OpenAI Harness Engineering

概念

组合性

Compositionality

复杂表达式的意义 = 部分意义 × 组合方式

  • concepts/systematicity
  • concepts/language-of-thought
  • concepts/neurosymbolic-ai

Frege 1892 · Fodor-Pylyshyn 1988 · arXiv 2407.13419

概念

恒常连结

Constant Conjunction

休谟因果性的经验基础 · 我们实际观察到的全部

  • concepts/causation-hume
  • concepts/necessary-connection
  • concepts/custom-and-habit

David Hume Enquiry 1748

概念

约束解码

Constrained Decoding

将非法 token 概率强制置零 · 结构化输出的底层机制

  • concepts/structured-outputs
  • concepts/context-free-grammar-llm

RANLP 2025 · OpenAI Structured Outputs

概念

Context 压缩

Context Compression

三种策略的质量 / 压缩率 / 可读性权衡

  • concepts/context-management
  • concepts/long-running-agents
  • concepts/context-rot

Factory Evaluating Context Compression · 36000+ 真实 sessions

概念

Context Engineering

Context Engineering

找到最小的高信号 token 集合 · 策展不是堆砌

  • concepts/context-management
  • concepts/context-rot
  • concepts/context-compression

Anthropic Effective Context Engineering · Manus Context Engineering

概念

上下文无关文法(LLM)

Context-Free Grammar for LLM

JSON Schema → CFG → 动态 token 掩码

  • concepts/constrained-decoding
  • concepts/structured-outputs

OpenAI Structured Outputs · RANLP 2025

概念

Context 管理

Context Management

顺行性遗忘 · 工作记忆必须被显式编程

  • concepts/context-engineering
  • concepts/long-running-agents
  • concepts/context-compression

Anthropic Effective Harnesses · Claude Agent SDK

概念

Context Rot

Context Rot

随 token 数增长的系统性性能退化

  • concepts/context-engineering
  • concepts/context-management

Chroma Context Rot Research · 2025

概念

反事实理论

Counterfactual Theory of Causation

如果 c 没有发生,e 就不会发生 · 可能世界语义

  • concepts/possible-world-semantics
  • concepts/preemption
  • concepts/causal-models

David Lewis 1973 · SEP Counterfactual Causation

概念

习惯

Custom and Habit

联想原则的一般名称 · 因果推理的真正基础

  • concepts/causation-hume
  • concepts/constant-conjunction
  • concepts/induction-problem

David Hume Enquiry 1748 EHU 5.1.5

实体

Cycle.js

Cycle.js

André Staltz · sources→sinks纯函数 · 分形架构典型实现 · 副作用推到driver层

  • concepts/fractal-architecture
  • concepts/implicit-loop-architecture
  • concepts/harness-engineering

Anton Telesh Fractal Architecture 2016

实体

David Hume

David Hume

1711-1776 · 苏格兰哲学家 · 恒常连结 · 归纳问题 · 习惯 · 康德的触发者

  • concepts/causation-hume
  • concepts/induction-problem
  • concepts/regularity-theory

SEP David Hume · SEP Metaphysics of Causation

实体

David Lewis

David Lewis

1941-2001 · 反事实因果理论 · 可能世界语义学 · 影响理论2000 · 类型因果=个例概括

  • concepts/counterfactual-theory
  • concepts/possible-world-semantics
  • concepts/preemption

SEP Counterfactual Theories of Causation · Lewis 1973 · Lewis 2000

概念

差异制造

Difference-Making

多种因果理论的共同内核

  • concepts/counterfactual-theory
  • concepts/regularity-theory
  • concepts/interventionist-theory

SEP Metaphysics of Causation · Strevens 2004

概念

do 演算

do-Calculus

Pearl 的三条规则 —— 把因果查询还原为可估计的观测表达式

  • concepts/causal-dag
  • concepts/structural-causal-model

Pearl (1995, 2009); Shpitser & Pearl (2006)

概念

经验主义

Empiricism

一切知识最终来源于经验 · 复本原则 · 定义方法

  • concepts/causation-hume
  • concepts/induction-problem
  • concepts/falsificationism

David Hume Enquiry · SEP David Hume

概念

误差级联

Error Cascade

前步小错在后续步骤被放大 · 不只是概率相乘

  • concepts/long-running-agents
  • concepts/harness-engineering
  • concepts/reliability-decay

SWE-EVO · arXiv multi-step evaluation

概念

评估器-优化器

Evaluator-Optimizer

生成者与评估者分离 · GAN 式质量反馈回路

  • concepts/harness-engineering
  • concepts/feature-tracking
  • concepts/agentic-systems

Anthropic Building Effective Agents

实体

Factory.ai

Factory.ai

AI编码agent · 锚定式迭代摘要 · 36K生产消息压缩评估 · artifact tracking弱点揭示

  • concepts/context-compression
  • concepts/harness-engineering
  • entities/anthropic

Factory Evaluating Context Compression 2025

概念

证伪主义

Falsificationism

大胆猜想 · 试图证伪 · 科学不需要归纳

  • concepts/induction-problem
  • concepts/bayesian-induction
  • concepts/empiricism

Karl Popper 1935 · SEP Induction Problem

概念

特性追踪

Feature Tracking

外部化的 machine-readable 任务清单 · 防止 one-shotting 和虚假胜利

  • concepts/long-running-agents
  • concepts/harness-engineering
  • concepts/context-management

Anthropic Effective Harnesses Long-Running Agents

概念

分叉-探索-提交

Fork-Explore-Commit

OS 级探索原语 · git 语义下沉到执行层 · 并行路径择优

  • concepts/agent-os
  • concepts/evaluator-optimizer
  • concepts/error-cascade

AgenticOS Workshop ASPLOS 2026

概念

四因说

Four Causes

质料因 · 形式因 · 动力因 · 目的因 · 因果多元主义起点

  • concepts/aitia
  • concepts/causal-models
  • concepts/hierarchical-systems

SEP Aristotle on Causality · Pearl 2010

概念

分形架构

Fractal Architecture

自相似性 · 四条规则 · 任何子树是有效的树

  • concepts/compositionality
  • concepts/implicit-loop-architecture
  • concepts/agent-skills

Anton Telesh 2016 · Cycle.js

概念

分形维数

Fractal Dimension

量化空间填充程度 · 非整数维度 · Richardson 公式的指数

  • concepts/coastline-paradox
  • concepts/fractal-architecture

Wikipedia Coastline Paradox · Mandelbrot 1967

概念

前门准则

Frontdoor Criterion

不可测混淆变量时通过中介变量识别因果效应

  • concepts/backdoor-criterion
  • concepts/causal-dag
  • concepts/do-calculus

Pearl 2010 · Tian & Pearl 2002

概念

生成-验证循环

Generation-Verification Loop

AI 生成 · 人类验证 · 优化速度决定效率 · 人是瓶颈

  • concepts/evaluator-optimizer
  • concepts/autonomy-slider
  • concepts/harness-engineering

Karpathy 2025 YC Talk

实体

Gerard de Melo

Gerard de Melo

计算语言学 · 轨迹偏差揭示者 · RANLP 2025 · Draft-Conditioned解法 · 指令模型vs基础模型分歧

  • concepts/trajectory-bias
  • concepts/structured-outputs

Schall & de Melo RANLP 2025 · arXiv:2603.03305

实体

Google

Google

A2A协议原始开发者 · ADK · Gemini · 捐赠Linux Foundation中立化 · 完整agent技术栈

  • concepts/a2a-protocol
  • entities/linux-foundation
  • concepts/agent-interoperability

Google A2A Protocol 2024

概念

Grue 问题

Grue Problem

新归纳之谜 · 同一数据支持矛盾结论 · 谓词可投射性

  • concepts/induction-problem
  • concepts/empiricism
  • concepts/bayesian-induction

Nelson Goodman 1955 · SEP Induction Problem

概念

护栏

Guardrails

分层防御 · 运行时安全监控 · 乐观执行模式

  • concepts/harness-engineering
  • concepts/aci
  • concepts/chaos-engineering-for-agents

OpenAI Practical Guide Building Agents

概念

Harness 工程

Harness Engineering

包裹 LLM 的控制系统 · 不是限制而是让能力可靠发挥

  • concepts/agentic-systems
  • concepts/feature-tracking
  • concepts/context-management

Anthropic Building Effective Agents · OpenAI Harness Engineering

实体

Henry Kautz

Henry Kautz

Rochester大学 · 六类神经符号分类法 · 第三个AI之夏 · 不会有第三个寒冬 · Allen Newell Award

  • concepts/neurosymbolic-ai-taxonomy
  • concepts/neurosymbolic-ai
  • entities/artur-garcez

Kautz 2022 AI Magazine The Third AI Summer

实体

Herbert A. Simon

Herbert A. Simon

诺贝尔奖1978 · 图灵奖1975 · 有限理性 · 复杂性架构 · 近可分解性 · PSSH共同作者

  • concepts/hierarchical-systems
  • concepts/near-decomposability
  • concepts/physical-symbol-system

Simon 1962 Architecture of Complexity · Newell & Simon 1976

概念

启发式搜索

Heuristic Search

物理符号系统的求解方式 · 搜索量不等于智能量

  • concepts/physical-symbol-system
  • concepts/means-ends-analysis

Newell & Simon 1976 CACM

概念

层级系统

Hierarchical Systems

交互强度驱动的嵌套结构 · 钟表匠寓言 · 演化的必然结果

  • concepts/near-decomposability
  • concepts/harness-engineering
  • concepts/context-engineering

Herbert Simon 1962 Architecture of Complexity

概念

隐式循环架构

Implicit Loop Architecture

gather context · take action · verify work · repeat · 行为由约束塑造

  • concepts/agentic-systems
  • concepts/harness-engineering
  • concepts/fork-explore-commit

Anthropic Building Agents · Karpathy LLM-OS

概念

因陀罗网

Indra's Net

全息互映 · 每个节点包含全局信息 · 自相似性的哲学发现

  • concepts/fractal-architecture
  • concepts/fractal-dimension

Avatamsaka Sutra · Cook 1973

概念

归纳问题

Problem of Induction

从"过去如此"到"未来如此"没有理性基础 · 休谟之叉

  • concepts/causation-hume
  • concepts/custom-and-habit
  • concepts/falsificationism

Hume Enquiry 1748 · SEP Problem of Induction

概念

推断性因果理论

Inferential Theories of Causation

从规律到推断路径 · 因果关系的推理结构刻画

  • concepts/inus-conditions
  • concepts/causal-models
  • concepts/difference-making

SEP Regularity and Inferential Theories · Hempel & Oppenheim 1948

概念

干预主义理论

Interventionist Theory of Causation

因果关系 = 主动操控的系统性回答 · Woodward 2003

  • concepts/causal-models
  • concepts/causal-intervention
  • concepts/counterfactual-theory

Woodward Making Things Happen 2003 · Pearl Causality 2000

概念

INUS 条件

INUS Conditions

不充分但非冗余的不必要但充分条件的部分 · Mackie 1965

  • concepts/difference-making
  • concepts/causal-models
  • concepts/constant-conjunction

Mackie Causes and Conditions 1965

实体

James Woodward

James Woodward

Pittsburgh大学 · Making Things Happen 2003 · 干预主义因果理论 · 模块性 · 类型影响焦点

  • concepts/interventionist-theory
  • concepts/causal-models
  • entities/judea-pearl

Woodward 2003 Making Things Happen · SEP Metaphysics of Causation

实体

Jeremy Howard

Jeremy Howard

fast.ai联合创始人 · Answer.AI创始人 · llms.txt标准 · AI教育民主化 · top-down实践法

  • concepts/llms-txt
  • entities/answer-ai

fast.ai · Answer.AI · llmstxt.org

实体

Jerry A. Fodor

Jerry A. Fodor

1935-2017 · Rutgers/MIT · 思维语言 · 模块性 · Fodor&Pylyshyn 1988 联结主义挑战

  • concepts/language-of-thought
  • concepts/systematicity
  • concepts/compositionality

Fodor & Pylyshyn 1988 Cognition · Fodor 1975 The Language of Thought

实体

J. L. Mackie

J. L. Mackie

1917-1981 · INUS条件 · 规则性理论最重要技术贡献 · 因果场 · 道德错误论

  • concepts/inus-conditions
  • concepts/regularity-theory
  • entities/david-hume

SEP Regularity and Inferential Theories · Mackie 1965 · Mackie 1974

实体

Judea Pearl

因果革命的奠基者 · 图灵奖 2011 · UCLA CS 教授

  • concepts/causal-dag
  • concepts/do-calculus

Causality (2009); The Book of Why (2018)

实体

Karl Popper

Karl Popper

1902-1994 · 证伪主义 · 接受归纳问题不可解 · 佐证vs确认 · 科学发现的逻辑

  • concepts/falsificationism
  • concepts/induction-problem
  • entities/david-hume

Popper 1935 Logic of Scientific Discovery · SEP Problem of Induction

概念

知识提取与忠实性

Knowledge Extraction and Fidelity

忠实性 ≠ 准确率 · 描述的是网络实际做什么

  • concepts/mechanistic-interpretability
  • concepts/activation-intervention

Garcez & Lamb 2020 Neurosymbolic AI 3rd Wave

概念

因果之梯

Ladder of Causation

关联 · 干预 · 反事实 · 三级不可还原

  • concepts/causal-intervention
  • concepts/do-calculus
  • concepts/causal-dag

Pearl & Mackenzie The Book of Why 2018

实体

LangGraph

LangGraph

LangChain · StateGraph显式编排 · 隐式循环的对比范式 · Durable Execution · 企业采用

  • concepts/implicit-loop-architecture
  • entities/claude-agent-sdk
  • entities/codex

LangGraph Documentation

概念

思维语言

Language of Thought

Mentalese · 句法结构化的心理表征 · 系统性的结构保证

  • concepts/compositionality
  • concepts/physical-symbol-system
  • concepts/linear-representation-hypothesis

Fodor 1975 Language of Thought · Fodor & Pylyshyn 1988

实体

Lewis Fry Richardson

Lewis Fry Richardson

英国数学家气象学家 · 海岸线悖论经验发现者 · Richardson效应命名原型 · 数值天气预报先驱

  • concepts/richardson-effect
  • concepts/coastline-paradox
  • entities/benoit-mandelbrot

Wikipedia Coastline Paradox · Richardson 1961

概念

线性表征假说

Linear Representation Hypothesis

特征以线性方式编码在激活空间中 · 机制可解释性的方法论基础

  • concepts/mechanistic-interpretability
  • concepts/causal-inner-product
  • concepts/activation-intervention

Park Choe Veitch 2023 ICML · Gurnee Tegmark 2023

实体

Linux Foundation

Linux Foundation

开源中立治理机构 · A2A协议托管方 · Kubernetes模式 · PyTorch Foundation

  • concepts/a2a-protocol
  • entities/google

Google A2A Protocol 2024

概念

LLM 操作系统

LLM OS

LLM 不是公用事业 · 是软件生态系统 · 我们在 1960 年代

  • concepts/llm-os-analogy
  • concepts/context-management
  • concepts/implicit-loop-architecture

Karpathy 2025 YC Talk Software Is Changing Again

概念

LLM-OS 类比

LLM-OS Analogy

LLM = CPU · Agent = OS Kernel · Context = RAM · 结构性同构

  • concepts/implicit-loop-architecture
  • concepts/context-management
  • concepts/harness-engineering

Karpathy 2023 Intro to LLMs · 2026 Tweet

概念

LLM 安全

LLM Security

越狱 · 提示注入 · 数据投毒 · 猫鼠游戏永不终止

  • concepts/guardrails
  • concepts/llm-os-analogy
  • concepts/harness-engineering

Karpathy 2023 Intro to Large Language Models

概念

LLM 训练流水线

LLM Training Pipeline

预训练 → 微调 → RLHF · 三阶段标准心智模型

  • concepts/bitter-lesson
  • concepts/augmented-llm
  • concepts/llm-os-analogy

Karpathy 2023 Intro to Large Language Models

概念

llms.txt

llms.txt

网站根路径的 Markdown 文件 · 告诉 LLM 这个网站是什么

  • concepts/aci
  • concepts/context-engineering
  • concepts/agent-skills

Jeremy Howard Answer.AI 2024

概念

长时运行 Agent

Long-Running Agents

跨 session 持续工作 · 每个新 session 天然失忆 · 轮班工程师比喻

  • concepts/harness-engineering
  • concepts/feature-tracking
  • concepts/context-management

Anthropic Effective Harnesses Long-Running Agents

实体

Manus

Manus

通用自主agent · 5维Context Engineering框架 · prefill/decode 100:1 · logit masking

  • concepts/context-engineering
  • concepts/prefix-caching
  • concepts/context-management

Manus Context Engineering Blog 2025

实体

Max Tegmark

Max Tegmark

MIT物理学家 · 数学宇宙假说 · Future of Life Institute · 时空世界模型共同作者

  • concepts/spatiotemporal-world-model
  • entities/wes-gurnee

Gurnee & Tegmark ICLR 2024 arXiv:2310.02207

实体

Maximilian Schall

Maximilian Schall

NLP研究者 · 轨迹偏差命名者 · Draft-Conditioned解法 · RANLP 2025

  • concepts/trajectory-bias
  • concepts/constrained-decoding
  • concepts/structured-outputs

RANLP 2025 arXiv:2603.03305

实体

MCP

Model Context Protocol

Anthropic开放协议 · Agent↔工具标准接口 · A2A互补 · Mobile-MCP扩展

  • concepts/augmented-llm
  • concepts/a2a-protocol
  • entities/anthropic

Anthropic Building Effective Agents 2024

概念

手段-目的分析

Means-Ends Analysis

GPS 差异驱动搜索 · 子目标递归 · CoT/ReAct 的历史前身

  • concepts/heuristic-search
  • concepts/agentic-systems
  • concepts/implicit-loop-architecture

Newell & Simon 1972 Human Problem Solving

概念

机制可解释性

Mechanistic Interpretability

特征 · 归因图 · CLT · 幻觉机制 · CoT 不忠实

  • concepts/linear-representation-hypothesis
  • concepts/othello-world-model-hypothesis
  • concepts/neurosymbolic-ai

Anthropic 2024 Mapping the Mind of an LLM

概念

中介分析

Mediation Analysis

TE · CDE · NDE · NIE · 直接效应与间接效应的分解

  • concepts/ladder-of-causation
  • concepts/frontdoor-criterion
  • concepts/interventionist-theory

Pearl 2001 Direct and Indirect Effects

实体

MemGPT

Memory-GPT

UC Berkeley · 虚拟上下文管理 · LLM as OS · 层次化存储 · arXiv:2310.08560

  • concepts/virtual-context-management
  • concepts/context-management
  • entities/aios

MemGPT arXiv:2310.08560 (2023)

概念

AI 元认知

Meta-Cognition in AI

监控 · 评估 · 调整 · NSAI 5% 盲区 · AlphaGeometry 唯一覆盖

  • concepts/neurosymbolic-ai
  • concepts/mechanistic-interpretability
  • concepts/meta-harness

arXiv 2501.05435 Neurosymbolic AI 2024 Systematic Review

概念

元 Harness

Meta-Harness

Session · Harness · Sandbox 三层接口 · 脑-手解耦 · OS 虚拟化类比

  • concepts/harness-engineering
  • concepts/permission-modes
  • concepts/long-running-agents

Anthropic Claude Code Architecture

概念

元归纳

Meta-Induction

wMI 策略 · 对象-元归纳层 · 遗憾学习 · Bandit 类比

  • concepts/induction-problem
  • concepts/no-free-lunch
  • concepts/bayesian-induction

Schurz 2008 The Meta-Inductivist

概念

准可分解性

Near-Decomposability

短期独立 · 长期聚合耦合 · 频率分离 · Simon 1962 钟表匠

  • concepts/hierarchical-systems
  • concepts/fractal-architecture
  • concepts/context-management

Simon 1962 The Architecture of Complexity

概念

必然联系

Necessary Connection

休谟批判 · 三来源逐一排除 · 心灵的决定感 · 康德回应

  • concepts/causation-hume
  • concepts/constant-conjunction
  • concepts/regularity-theory

Hume 1748 Enquiry Concerning Human Understanding

实体

Neel Nanda

Neel Nanda

Google DeepMind · TransformerLens开发者 · MATS导师 · OthelloGPT线性化 · 机制可解释性

  • concepts/mechanistic-interpretability
  • concepts/othello-world-model-hypothesis
  • concepts/linear-representation-hypothesis

Nanda et al. BlackboxNLP @ EMNLP 2023

实体

Nelson Goodman

Nelson Goodman

美国哲学家 · Grue问题 · 新归纳之谜 · 构造主义世界观 · 符号理论

  • concepts/grue-problem
  • concepts/induction-problem
  • entities/david-hume

Goodman 1955 Fact Fiction and Forecast

概念

神经符号 AI

Neurosymbolic AI

System 1+2 集成 · 存在/全称量化互补 · Fodor-Pylyshyn 挑战

  • concepts/neurosymbolic-ai-taxonomy
  • concepts/system-1-vs-system-2
  • concepts/knowledge-extraction-fidelity

Garcez & Lamb 2023 Neurosymbolic AI The 3rd Wave

概念

神经符号 AI 分类法

Neurosymbolic AI Taxonomy

Kautz 六类 · 耦合谱 · 从 Symbolic_Neuro 到 Neuro[Symbolic]

  • concepts/neurosymbolic-ai
  • concepts/system-1-vs-system-2
  • concepts/constrained-decoding

Kautz 2022 The Third AI Summer AI Magazine

概念

No Free Lunch 定理

No Free Lunch Theorem

均匀分布下泛化误差 1/2 · 休谟归纳问题的数学版 · 归纳偏置不可回避

  • concepts/induction-problem
  • concepts/bayesian-induction
  • concepts/scaling-laws

Wolpert 1996 Neural Computation

实体

OpenAI

OpenAI

GPT系列 · Codex agent平台 · Harness Engineering方法论 · Agent Legibility · Structured Outputs

  • concepts/harness-engineering
  • concepts/implicit-loop-architecture
  • entities/codex

OpenAI Harness Engineering / Structured Outputs 2024

实体

OpenHands

OpenHands

开源多agent编码平台 · CodeActAgent · SWE-EVO基准框架 · 模型×框架性能函数

  • concepts/implicit-loop-architecture
  • entities/swe-bench
  • entities/codex

SWE-EVO arXiv:2512.18470

概念

编排器-工作者

Orchestrator-Workers

动态任务分解 · 区别于预定义并行 · A2A 跨系统扩展

  • concepts/parallelization
  • concepts/agentic-systems
  • concepts/harness-engineering

Anthropic Building Effective Agents 2024

概念

Othello 世界模型假说

Othello World Model Hypothesis

序列训练 → 棋盘表征涌现 · 99% 精度 · 表征 ≠ 规划

  • concepts/mechanistic-interpretability
  • concepts/linear-representation-hypothesis
  • concepts/world-models

Li et al. 2022 OthelloGPT · Nanda et al. 2023 · Yuan & Søgaard 2025

概念

并行化

Parallelization

Sectioning 分段 · Voting 投票 · 预定义子任务 · 置信度提升

  • concepts/orchestrator-workers
  • concepts/agentic-systems
  • concepts/guardrails

Anthropic Building Effective Agents 2024

实体

Paul Smolensky

Paul Smolensky

约翰斯·霍普金斯 · 调和理论 · 亚概念层理论 · 张量积表征 · 优化理论音系学

  • concepts/subconceptual-level
  • concepts/compositionality
  • entities/jerry-fodor

Smolensky 1988 Behavioral and Brain Sciences

概念

权限模式

Permission Modes

六模式自主度谱 · auto 分类器 · bypassPermissions 核选项

  • concepts/harness-engineering
  • concepts/guardrails
  • concepts/agent-sandboxing

Anthropic Claude Code Permissions 2024

概念

Allow/Ask/Deny 规则层次

Permission Rules Hierarchy

Deny-First 语义 · 评估顺序固定 · 跨层级不可覆盖

  • concepts/permission-modes
  • concepts/claude-code-permission-system
  • concepts/guardrails

Anthropic Claude Code Permissions 2024

概念

宠物与牲畜

Pets vs Cattle

可处置性 · 组件独立故障恢复 · 脑-手解耦使 Cattle 化成为可能

  • concepts/meta-harness
  • concepts/agent-sandboxing
  • concepts/reliability-decay

Randy Bias 2011 Cloud Computing

概念

物理符号系统假说

Physical Symbol System Hypothesis

Newell & Simon 1976 · 通用智能充要条件 · GOFAI 理论基础

  • concepts/heuristic-search
  • concepts/problem-space
  • concepts/neurosymbolic-ai

Newell & Simon 1976 Computer Science as Empirical Inquiry

概念

可能世界语义学

Possible World Semantics

最接近世界 · 比较相似性 · 非回溯解释 · Lewis 1973

  • concepts/counterfactual-theory
  • concepts/interventionist-theory
  • concepts/ladder-of-causation

Lewis 1973 Counterfactuals

概念

先占

Preemption

早期 · 晚期 · 凌驾 · 反事实理论的核心测试场

  • concepts/counterfactual-theory
  • concepts/process-theories
  • concepts/causal-models

Hall 2004 Two Concepts of Causation · Schaffer 2000

概念

前缀缓存

Prefix Caching

KV Cache 复用 · 10× 成本差距 · 静态前缀稳定优先 · Compaction 对立

  • concepts/context-management
  • concepts/harness-engineering
  • concepts/context-compression

Anthropic / OpenAI / Kimi Prefix Caching Docs 2024

概念

概率因果

Probabilistic Causation

概率提升 · 净效应与路径效应 · 先占挑战

  • concepts/counterfactual-theory
  • concepts/preemption
  • concepts/regularity-theory

Suppes 1970 · Eells 1991 · Hitchcock 2001

概念

因果概率

Probability of Causation

PN · PS · PNS · but-for 法律标准 · 单调性识别

  • concepts/ladder-of-causation
  • concepts/counterfactual-theory
  • concepts/structural-causal-model

Pearl 2009 Causality Chapter 9

概念

探针分类器

Probing Classifiers

线性/非线性探针 · 高性能 ≠ 模型使用 · 因果验证必要

  • concepts/linear-representation-hypothesis
  • concepts/mechanistic-interpretability
  • concepts/othello-world-model-hypothesis

Belinkov 2022 Probing Classifiers Review

概念

问题空间

Problem Space

状态-算子-目标 · 表示决定难度 · Newell & Simon 1976

  • concepts/heuristic-search
  • concepts/means-ends-analysis
  • concepts/physical-symbol-system

Newell & Simon 1976 Computer Science as Empirical Inquiry

概念

过程理论

Process Theories of Causation

守恒量传递 · 因果过程 vs 伪过程 · 先占强项 · 预防盲区

  • concepts/preemption
  • concepts/counterfactual-theory
  • concepts/regularity-theory

Salmon 1984 · Dowe 2000 Physical Causation

概念

提示链

Prompt Chaining

顺序步骤 · Gate 检查 · 用延迟换准确率 · 最简单 Workflow 模式

  • concepts/agentic-systems
  • concepts/routing
  • concepts/orchestrator-workers

Anthropic Building Effective Agents 2024

概念

命题固着

Propositional Fixation

McCarthy 神经网络局限 · 有限一阶逻辑 · 全称量化天花板

  • concepts/neurosymbolic-ai
  • concepts/neurosymbolic-ai-taxonomy
  • concepts/constrained-decoding

McCarthy 引用于 Garcez & Lamb 2023 Neurosymbolic AI

概念

规则性理论

Regularity Theory of Causation

休谟传统 · 恒常连结 · INUS 条件 · 当代演进

  • concepts/constant-conjunction
  • concepts/inus-conditions
  • concepts/necessary-connection

Hume 1739 · Mackie 1965 · Baumgartner 2013

概念

可靠性衰减

Reliability Decay

超线性衰减 · 错误正相关 · 领域 > 模型大小 · RDC 斜率定量任务分解收益

  • concepts/long-running-agents
  • concepts/error-cascade
  • concepts/harness-engineering

arXiv 2603.29231 Beyond Pass@1 Reliability Framework

概念

可靠性曲面

Reliability Surface

三维评估 · k 一致性 · ε 鲁棒性 · λ 容错性 · pass@1 不够用

  • concepts/reliability-decay
  • concepts/error-cascade
  • concepts/harness-engineering

arXiv 2601.06112 ReliabilityBench 2026

概念

Agent 的 REPL

REPL for Agents

驻留进程状态 · Undo 一等原语 · 双模式 · vs Subcommand vs MCP

  • concepts/implicit-loop-architecture
  • concepts/tool-design
  • concepts/context-management

CLI-Anything HARNESS.md

实体

Rich Sutton

Rich Sutton

强化学习奠基人 · The Bitter Lesson 2019 · 通用计算方法总是胜出 · 持续剥离假设

  • concepts/harness-engineering
  • concepts/meta-harness
  • entities/anthropic

Sutton The Bitter Lesson 2019

概念

Richardson 效应

Richardson Effect

尺子越短 · 边界越长 · 趋零发散 · 分形维数的经验来源

  • concepts/coastline-paradox
  • concepts/fractal-dimension
  • concepts/statistical-self-similarity

Richardson 1951 · Mandelbrot 1967

概念

路由

Routing

分类后导向专用流程 · 关注点分离 · 模型选择 · 客服分流

  • concepts/agentic-systems
  • concepts/prompt-chaining
  • concepts/orchestrator-workers

Anthropic Building Effective Agents 2024

概念

规则循环

Rule Circularity

用推理规则 R 辩护 R 自身 · 前提循环 vs 规则循环 · 反归纳的自我支持难题

  • concepts/induction-problem
  • concepts/uniformity-principle
  • concepts/meta-induction

Carroll 1895 · SEP Induction Problem

概念

缩放定律

Scaling Laws

N×D 可预测 · 无饱和迹象 · Gold Rush 理论基础 · 多步误差级联抵消

  • concepts/llm-training-pipeline
  • concepts/error-cascade
  • concepts/bitter-lesson

Karpathy 2023 Intro to Large Language Models

概念

设置作用域层次

Settings Scope Hierarchy

五级优先级 · 托管设置不可覆盖 · Deny 跨层封锁 · auto 模式特殊规则

  • concepts/permission-modes
  • concepts/permission-rules-hierarchy
  • concepts/claude-code-permission-system

Anthropic Claude Code Permissions 2024

概念

Software 3.0

Software 3.0

Prompt 即程序 · 自然语言编程 · 三代范式 · Karpathy 2025

  • concepts/llm-os
  • concepts/context-engineering
  • concepts/agentic-systems

Karpathy 2025 YC Talk Software Is Changing Again

概念

软件演进基准

Software Evolution Benchmark

SWE-EVO · 单步 72.8% → 多步 ~20% · Release Notes 输入 · 误差级联暴露

  • concepts/error-cascade
  • concepts/long-running-agents
  • concepts/harness-engineering

arXiv 2512.18470 SWE-EVO

概念

时空世界模型

Spatiotemporal World Model

Gurnee & Tegmark 2024 · R²=0.91/0.84 · 线性探针 · 涌现空间+时间表征

  • concepts/othello-world-model-hypothesis
  • concepts/probing-classifiers
  • concepts/mechanistic-interpretability

Gurnee & Tegmark ICLR 2024 arXiv:2310.02207

概念

SSM 混合架构

SSM Hybrid Architecture

Mamba-3 · Jamba 1/8注意力 · 线性复杂度 · 长上下文代理推理优化

  • concepts/long-running-agents
  • concepts/harness-engineering
  • concepts/prefix-caching

Mamba-3 2025 · Jamba AI21 Labs 2024

概念

统计自相似性

Statistical Self-Similarity

自然界分形 · 统计属性跨尺度恒定 · Richardson 效应基础 · 严格 vs 统计

  • concepts/richardson-effect
  • concepts/fractal-dimension
  • concepts/scale-free-networks

Mandelbrot 1967 · Richardson 1961

概念

结构因果模型

Structural Causal Model

SCM ⟨U,V,E⟩ · 三阶梯统一框架 · 结构方程 ≠ 回归 · Pearl 2010 元框架

  • concepts/causal-dag
  • concepts/do-calculus
  • concepts/probability-of-causation

Pearl 2010 Causality 2nd Ed · arXiv:1304.1108

概念

结构化输出

Structured Outputs

Schema→CFG→Token 掩码 · JSON mode vs Structured 11.97%→0.1% · 轨迹偏差隐性代价

  • concepts/harness-engineering
  • concepts/context-engineering
  • concepts/prompt-chaining

Anthropic Claude Structured Outputs · OpenAI Structured Outputs 2024

概念

亚概念层

Subconceptual Level

Smolensky 1988 · 微特征 · 三层框架 · 神经-符号中间层 · 分布式表征

  • concepts/neurosymbolic-ai
  • concepts/physical-symbol-system
  • concepts/mechanistic-interpretability

Smolensky 1988 On the Proper Treatment of Connectionism

概念

须弥芥子

Sumeru Mustard Seed

大小无碍 · 维摩诘经 · 华严宗 · 自相似性哲学先驱 · 布莱克/莱布尼茨共鸣

  • concepts/indra-net
  • concepts/statistical-self-similarity
  • concepts/compositionality

维摩诘所说经·不思议品 · 华严宗

实体

SWE-Bench

SWE-Bench

Princeton NLP · 编码Agent评估事实标准 · GitHub Issue修复 · SWE-Bench Verified · 72.8%饱和

  • concepts/software-evolution-benchmark
  • concepts/error-cascade
  • entities/openhands

SWE-EVO arXiv:2512.18470

概念

符号接地问题

Symbol Grounding Problem

Harnad 1990 · 字典循环 · 中文房间 · 三种解决路径 · LLM 部分接地争议

  • concepts/physical-symbol-system
  • concepts/neurosymbolic-ai
  • concepts/spatiotemporal-world-model

Harnad 1990 Physica D · Searle 1980

概念

系统一与系统二

System 1 vs System 2

Kahneman · Karpathy 2023 LLM 只有 System 1 · CoT/o1 是工程化 System 2 · 非二分告诫

  • concepts/meta-cognition-ai
  • concepts/evaluator-optimizer
  • concepts/neurosymbolic-ai-taxonomy

Kahneman 2011 · Karpathy 2023 Intro to LLMs

概念

认知系统性

Systematicity

Fodor & Pylyshyn 1988 · 非偶然的能力关联 · 联结主义的核心挑战 · LLM 实证验证

  • concepts/compositionality
  • concepts/physical-symbol-system
  • concepts/neurosymbolic-ai

Fodor & Pylyshyn 1988 Cognition · Dhar & Søgaard 2024

概念

任务生命周期

Task Lifecycle

A2A 状态机 · submitted→working→4终态 · input-required 人类介入 · 三种追踪模式

  • concepts/a2a-protocol
  • concepts/long-running-agents
  • concepts/context-management

Google A2A Protocol 2024

概念

目的论解释

Teleological Explanation

亚里士多德 · 规律性需要解释 · 非心理学目的论 · 三因合一 · 目的因优先性

  • concepts/four-causes
  • concepts/generation-verification-loop
  • concepts/aristotle

Stanford Encyclopedia of Philosophy — Aristotle on Causality

实体

Thomas Bayes

Thomas Bayes

英国牧师数学家 · 贝叶斯定理 · 逆概率方法 · 1764遗著 · 归纳问题的概率回应

  • concepts/bayesian-induction
  • concepts/induction-problem
  • entities/david-hume

Bayes 1764 Philosophical Transactions

概念

个例因果与类型因果

Token vs Type Causation

2×2矩阵 · 三种关系立场 · 非对称存在案例 · 净效应vs分量效应

  • concepts/probabilistic-causation
  • concepts/regularity-theory
  • concepts/structural-causal-model

SEP Metaphysics of Causation · Lewis 1973 · Eells 1991

概念

工具设计

Tool Design

格式选择 · 最小可行工具集 · Poka-yoke · 双模式双输出 · token 效率

  • concepts/aci
  • concepts/context-engineering
  • concepts/repl-for-agents

Anthropic Building Effective Agents · CLI-Anything

概念

轨迹偏差

Trajectory Bias

Schall & de Melo RANLP 2025 · 约束解码累积扰动 · 格式正确≠答案正确 · CRANE 解法

  • concepts/structured-outputs
  • concepts/system-1-vs-system-2
  • concepts/neurosymbolic-ai

Schall & de Melo RANLP 2025 · CRANE arXiv:2502.09061

概念

齐一性原则

Uniformity Principle

未来将与过去相似 · 休谟二难 · 训练分布=测试分布 · 分布漂移的哲学基础

  • concepts/induction-problem
  • concepts/causation-hume
  • concepts/grue-problem

Stanford Encyclopedia of Philosophy — Problem of Induction

概念

虚拟上下文管理

Virtual Context Management

MemGPT · OS 虚拟内存类比 · RAM/磁盘层次 · Page fault → 记忆检索

  • concepts/context-management
  • concepts/context-rot
  • concepts/long-running-agents

Packer et al. 2023 MemGPT arXiv:2310.08560

实体

Wes Gurnee

Wes Gurnee

MIT · LLM可解释性 · 时空表征R²=0.911/0.835 · 稀疏探针 · ICLR 2024共同作者

  • concepts/spatiotemporal-world-model
  • concepts/linear-representation-hypothesis
  • entities/max-tegmark

Gurnee & Tegmark ICLR 2024 arXiv:2310.02207

实体

Wesley Salmon

Wesley Salmon

科学哲学家 · 过程因果理论 · 标记传递 · 因果-机制解释模型 · 1925-2001

  • concepts/process-theories
  • concepts/preemption
  • entities/james-woodward

Salmon 1984 Scientific Explanation and the Causal Structure of the World

概念

世界模型

World Models

LeCun AMI · 预测下一状态而非下一token · 物理推理 · LLM涌现证据链

  • concepts/othello-world-model-hypothesis
  • concepts/spatiotemporal-world-model
  • concepts/agentic-systems

LeCun AMI Labs · World Models Race 2026

02 源头 — 82

每个源头都是一篇论文、博客或设计文档的结构化摘要。