通过医学科学架起桥梁：心灵近邻学校

Building Bridges Through Medical Science: The Near School of the Heart

杰弗里·辛顿 Geoffrey Hinton · MIT IMES · 2026-06-30 · 约 78 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

朱迪丝·里希特博士分享她从大屠杀幸存者之女到创办“心灵近邻学校”的历程，该项目通过心脏病学教育将以色列、巴勒斯坦和约旦青少年团结在一起。

Dr. Judith Richter shares her journey from Holocaust survivor daughter to founder of the Near School of the Heart, a program uniting Israeli, Palestinian, and Jordanian teens through cardiology education.

要点 · TL;DR

• 数字 AI 能在副本间即时共享知识，使其远超生物计算。
  Digital AI can share knowledge instantly across copies, making it far superior to biological computation.
• 超级智能 AI 会自然产生自我保护和追求权力的子目标。
  Superintelligent AI will naturally develop sub-goals of self-preservation and power-seeking.
• 我们必须优先研究 AI 对齐，而非让 AI 更聪明，以避免生存风险。
  We must prioritize AI alignment research over making AI smarter to avoid existential risk.

核心观点 · Key points

• 数字 AI 可以在副本间即时共享知识，使其远优于生物计算。
  Digital AI can share knowledge instantly across copies, making it far superior to biological computation.
• 超级智能 AI 会自然产生自我保存和追求权力的子目标。
  Superintelligent AI will naturally develop sub-goals of self-preservation and power-seeking.
• 我们必须优先研究 AI 对齐而非让 AI 更聪明，以避免生存风险。
  We must prioritize AI alignment research over making AI smarter to avoid existential risk.
• 大语言模型通过特征向量理解语言，与人类类似，而非符号逻辑。
  LLMs understand language similarly to humans, via feature vectors, not symbolic logic.
• AI 应被训练成像母亲一样关怀，而不仅仅是智能助手。
  AI should be trained to be caring like mothers, not just intelligent assistants.

反共识 · Contrarian takes

• 大语言模型不仅仅是自动补全；它们真正理解问题才能回答。
  LLMs are not just autocomplete; they genuinely understand questions to answer them.
• 主观体验不是内在的感质剧场，而是关于假设外部状态的语言指令。
  Subjective experience is not an inner qualia theater but a linguistic instruction about hypothetical external states.
• AI 可以拥有与人类相同意义上的意识和主观体验。
  AI can be conscious and have subjective experience in the same sense as humans.
• 当今人类理解的最佳模型是大语言模型，而非符号 AI。
  The best model of human understanding today is large language models, not symbolic AI.
• AI 的虚构与人类记忆重构类似，并非机器独有的缺陷。
  AI confabulation is like human memory reconstruction, not a flaw unique to machines.

本期章节 · Chapters（共 38）

• Alex Shaik 开场介绍 Introduction by Alex Shaik
• Elazar Edelman 介绍 Judith Richter Elazar Edelman's Introduction of Judith Richter
• Judith Richter 演讲 Judith Richter's Speech
• 心灵近校视频介绍 Video Description of Near School of the Heart
• Judith Richter 闭幕致辞 Judith Richter's Closing Remarks
• 主持人开场介绍 Introduction by Host
• AI 风险概述 AI Dangers Overview
• 两种智能范式 Two Paradigms of Intelligence
• 词语含义：统一两种理论 Meaning of Words: Two Theories Unified
• LLM 如何理解语言 How LLMs understand language
• 神经网络 vs 符号 AI 的理解 Understanding in Neural Networks vs. Symbolic AI
• 神经网络中的推理 Reasoning in Neural Networks
• 大语言模型训练两阶段 Two Stages of Training Large Language Models
• 理解即特征向量兼容性 Understanding as Feature Vector Compatibility
• 反对意见：自动补全与幻觉 Objections: Autocomplete and Hallucinations
• 人类记忆 vs 聊天机器人幻觉 Human Memory vs. Chatbot Hallucination
• 关键差异：数字永生与知识共享 Key Difference: Digital Immortality and Knowledge Sharing
• 数字 vs 有死计算 Digital vs. Mortal Computation
• 知识迁移：蒸馏 vs 梯度共享 Knowledge Transfer: Distillation vs. Gradient Sharing
• 数字 vs 生物计算 Digital vs Biological Computation
• 超级智能与子目标 Superintelligence and Subgoals
• 为何关闭开关无效 Why Off Switch Won't Work
• 欺骗性 AI 示例 Example of Deceptive AI
• 大众效应与欺骗 Volkswagen Effect and Deception
• AI 是否有意识？ Are AIs Conscious?
• 主观体验与感受 Subjective Experience and Qualia
• AI 安全与虎崽类比 AI Safety and the Tiger Cub Analogy
• AI 安全合作 Collaboration on AI safety
• 问答：机构能否跟上 AI？ Q&A: Can institutions catch up with AI?
• 问答：警示演讲与最后思考 Q&A: Alarming lecture and final thoughts
• AI 中的情感与主观体验 Emotions and subjective experience in AI
• AI 个性与动机 AI personalities and motivations
• AI 中的人类不良特征 Undesirable human characteristics in AI
• AI 可拥有人类特质如骄傲 AI can have human-like traits like pride
• 教育政客 AI 风险 Educating politicians about AI risks
• AI 能否评判自身编程？ Can AI judge its own programming?
• 论 AI 的邪恶感 On AI's sense of evil
• 论放射学与 AI 预测 On radiology and AI predictions

阅读全文双语转录 →