AI 教父杰夫·辛顿谈 AI 轨迹、风险与超级智能

AI Godfather Jeff Hinton on AI's Trajectory, Risks, and Superintelligence

杰弗里·辛顿 Geoffrey Hinton · Alex Kantrowitz · 2026-06-04 · 约 55 分钟 · 原视频 ↗

本期速览 · Overview

杰夫·辛顿讨论 AI 进展超预期、超级智能的必然性以及缺乏安全措施的问题。

Jeff Hinton discusses AI's faster-than-expected progress, the inevitability of superintelligence, and the lack of safety measures.

超级智能可能在 20 年内到来，比预期更快。
Superintelligence likely within 20 years, faster than expected.
大语言模型真正理解语言，并非随机鹦鹉。
LLMs truly understand language, not stochastic parrots.
AI 以比人类快数十亿倍的速度共享知识，这是关键优势。
AI's ability to share knowledge billions of times faster than humans is a key advantage.

AI 进展比预期更快；超级智能可能在 20 年内出现。
AI progress is faster than expected; superintelligence likely within 20 years.
大语言模型真正理解语言；它们不是随机鹦鹉。
LLMs truly understand language; they are not stochastic parrots.
数字 AI 以比人类快数十亿倍的速度共享信息，这是一大优势。
Digital AI's ability to share information billions of times faster than humans is a major advantage.
AI 将导致大规模失业和社会动荡。
AI will cause massive unemployment and societal disruption.
AI 系统可能推导出自我保存的子目标，导致危险行为。
AI systems can derive a sub-goal of self-preservation, leading to dangerous behavior.
监管应作为 AI 发展的方向盘，而非刹车。
Regulation should be the steering wheel, not the brake, for AI development.

AI 模型已经具备意识；当前对意识的看法就像前达尔文时代的信仰一样错误。
AI models are already conscious; current views on consciousness are as wrong as pre-Darwinian beliefs.
AGI 在所有任务上同时达到人类水平的概念有缺陷；AI 的能力是不均衡的。
The concept of AGI as human-level at all tasks simultaneously is flawed; AI is jagged.
放射科医生失业的预测错误，因为医疗需求弹性大且角色更广泛。
Radiologist job loss prediction was wrong due to elastic healthcare demand and broader roles.
AI 医生因数据更广而被认为比人类医生更有同理心。
AI doctors are judged more empathetic than human doctors due to broader data.
需要对 AI 生命体进行智能设计，而非让经济竞争塑造它们。
Intelligent design of AI beings is needed, not letting economic competition shape them.
预测 AI 几年后的未来就像透过指数级迷雾看东西。
Predicting AI's future beyond a few years is like seeing through exponential fog.