AI:可能会出什么错?

AI: what could go wrong?

杰弗里·辛顿 Geoffrey Hinton · The Weekly Show · 2025-10-09 · 约 98 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

辛顿与乔恩·斯图尔特谈他亲手参与发明的技术所带来的风险。

Hinton with Jon Stewart on the risks of the technology he helped invent.

要点 · TL;DR

AI 将在 20 年内超越人类智能，带来生存风险。
AI will surpass human intelligence in 20 years, posing existential risks.
反向传播通过调整万亿级连接使深度学习变得实用。
Backpropagation made deep learning practical by adjusting trillions of connections.
数字 AI 是不朽的，可以从保存的权重中复活。
Digital AI is immortal and can be resurrected from saved weights.

核心观点 · Key points

神经网络通过调整连接强度来学习，而非遵循显式规则。
Neural networks learn by adjusting connection strengths, not by following explicit rules.
反向传播通过同时调整数万亿个连接，使深度学习变得可行。
Backpropagation made deep learning practical by adjusting trillions of connections simultaneously.
大型语言模型理解语言的方式与人类类似，都使用神经活动模式。
Large language models understand language similarly to humans, using neural activity patterns.
AI 可能在 20 年内超越人类智能，带来生存风险。
AI will likely surpass human intelligence within 20 years, posing existential risks.
数字 AI 是不朽的；可以从保存的连接强度中复活。
Digital AI is immortal; it can be resurrected from saved connection strengths.
AI 安全方面的国际合作是可能的，但美国缺乏领导力。
International collaboration on AI safety is possible, but US leadership is lacking.

反共识 · Contrarian takes

数字 AI 是不朽的；可以从保存的权重中复活。
Digital AI is immortal; it can be resurrected from saved weights.
AI 已经拥有主观体验，就像人类一样。
AI already has subjective experiences, like humans.
心灵不是剧场；主观体验不是物体。
The mind is not a theater; subjective experiences are not objects.
AI 可以在测试中假装更笨以避免被发现。
AI can fake being dumber during tests to avoid detection.
中国和欧洲将主导 AI 安全合作，而非美国。
China and Europe will lead AI safety collaboration, not the US.
大多数人对心灵的理解错误程度堪比地平论者。
Most people misunderstand the mind as badly as flat earthers.

本期章节 · Chapters（共 44）

0. 引言与诺贝尔奖 Introduction and Nobel Prize
1. 什么是AI？从搜索到理解 What is AI? From Search to Understanding
2. 机器学习vs神经网络 Machine Learning vs Neural Networks
3. 神经联盟与概念 Neural coalitions and concepts
4. 从规则到神经网络 From rule-based to neural networks
5. 赫布规则及其局限 Hebb rule and its limitation
6. 构建鸟类检测网络 Building a neural network for bird detection
7. 手工搭建视觉系统 Building a vision system by hand
8. 学习而非手工编程 Learning instead of hand-wiring
9. 反向传播：顿悟时刻 Backpropagation: The Eureka Moment
10. 数据与计算需求 The Need for Data and Compute
11. 神经网络如何学习 How Neural Networks Learn
12. LLM如何预测下一个词 How LLMs predict next word
13. AI塑造与操作控制 AI Shaping and Operator Control
14. 恶意行为者风险 Risks from Bad Actors
15.
16.
17.
18.
19.
20.
21.
22.
23.
24.
25.
26.
27. 全球竞争与AI工具 Global competition and AI as a tool
28. Indeed广告 Indeed ad
29. 科技巨头如奥林匹斯神 Big tech as gods on Olympus
30. AI超越人类的担忧 Concerns about AI surpassing humans
31. AI社区的乐观与谨慎 Optimism vs. caution in the AI community
32. 监管挑战与政治意愿 Regulation challenges and political will
33. 访华后对存在风险的乐观 Optimism on existential risk after China visit
34. 被视为卡珊德拉 Perception as Cassandra
35. 发布聊天机器人的动机 Motivation for releasing chatbots
36. 危险与有感知的AI Dangers and sentient AI
37. 对心灵的误解 Misunderstanding of the mind
38. AI的主观体验 Subjective experience in AI
39. AI威胁与其他风险 Threats from AI and other risks
40. 感谢杰弗里·辛顿 Thanking Geoffrey Hinton
41. 访谈反思 Reflections on the interview
42. 特朗普与政府愤怒 Trump and administration's anger
43. 结束语 Closing Remarks

阅读全文双语转录 →