世界模型 vs. 大语言模型：医疗 AI 的下一个阶段

World Models vs. LLMs: The Next Phase of AI in Healthcare

杨立昆 Yann LeCun · Offcall 播客 · 2026-02-12 · 约 35 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

Alex Le Brun 和 Yan Lun 讨论世界模型作为大语言模型的继任者，强调其在医疗领域规划和可靠性方面的潜力。

Alex Le Brun and Yan Lun discuss world models as a successor to large language models, emphasizing their potential for planning and reliability in healthcare.

要点 · TL;DR

• 世界模型预测行动后果以进行规划，不同于局限于标记化数据的 LLM。
  World models predict action outcomes for planning, unlike LLMs limited to tokenized data.
• 医疗需要可靠的 AI；LLM 在高风险规划任务中失败。
  Healthcare needs reliable AI; LLMs fail in high-stakes planning tasks.
• 世界模型将作为智能助手增强医生，而非取代他们。
  World models will amplify doctors as smart assistants, not replace them.

核心观点 · Key points

• 世界模型预测行动后果，实现规划，与 LLM 不同。
  World models predict consequences of actions, enabling planning unlike LLMs.
• LLM 局限于离散数据；真实世界是连续且高维的。
  LLMs are limited to tokenized data; real world is continuous and high-dimensional.
• 医疗需要可靠 AI；LLM 在高风险规划任务中失败。
  Healthcare needs reliable AI; LLMs fail in high-stakes planning tasks.
• 世界模型将增强医生而非取代他们，充当智能助手。
  World models will amplify doctors, not replace them, acting as smart assistants.
• 构建通用世界模型需要数年，但窄用例更快实现。
  Building a universal world model will take years, but narrow use cases sooner.

反共识 · Contrarian takes

• 语言是容易的部分；理解物理世界要难得多。
  Language is the easy part; physical world understanding is much harder.
• 四岁儿童的视觉数据相当于互联网上所有公开文本。
  A four-year-old's visual data equals all public text on the internet.
• LLM 像布罗卡区；它们生成文本但不思考。
  LLMs are like Broca's area; they generate text but don't think.
• 医疗中 80%准确率无用；医生会忽略不可靠的工具。
  80% accuracy in healthcare is useless; physicians ignore unreliable tools.
• 医疗编码需要结合指南推理，而不仅仅是模式匹配。
  Medical coding requires reasoning with guidelines, not just pattern matching.
• 像 A1C 低于 7 这样的离散阈值是武断的；世界模型实现个性化优化。
  Discrete thresholds like A1C below 7 are arbitrary; world models enable personalized optimization.

本期章节 · Chapters（共 16）

• 开场与嘉宾介绍 Introduction and Guest Welcome
• 什么是世界模型？ What is a World Model?
• 为何需要世界模型？ Why World Models?
• 语言与现实经验的局限 Limitations of Language vs. Real-World Experience
• 临床笔记的局限 Limitations of Clinical Notes
• 大脑中的世界模型与 LLM World Model and LLM in the Brain
• 医疗 AI 面临的挑战 Challenges in Healthcare AI
• 临床 AI 采纳的挑战 Challenges in clinical AI adoption
• LLM 在医疗中的当前与未来影响 LLMs in healthcare: current and future impact
• 未来愿景：AI 助手与世界模型 Vision for the future: AI assistants and world models
• AI 是放大器而非替代品 AI as Amplifier, Not Replacement
• 医疗编码作为短期目标 Medical Coding as Short-Term Target
• 离散与连续决策 Discrete vs Continuous Decision Making
• 闪电轮：专业与炒作 Lightning Round: Specialty and Hype
• 从研究到医疗：合作 From Research to Healthcare: Partnership
• 使命与合作伙伴 Mission and Partnerships

阅读全文双语转录 →