AlphaGo：改变人工智能的那场对局

AlphaGo: The Match That Changed AI Forever

普什米特·科利 Pushmeet Kohli · Google DeepMind · 2026-03-10 · 约 54 分钟 · 原视频 ↗

本期速览 · Overview

十年前，AlphaGo 在围棋中击败李世石，标志着现代人工智能革命的开始。听突破背后的架构师讲述这个故事。

Ten years ago, AlphaGo defeated Lee Sedol in Go, marking the start of the modern AI revolution. Hear from the architects behind the breakthrough.

AlphaGo 结合直觉与搜索掌握围棋，超越人类知识。
AlphaGo combined intuition and search to master Go, surpassing human knowledge.
第 37 手展示了 AI 能产生非人类的创造性洞见。
Move 37 revealed AI can produce creative, non-human insights.
AlphaZero 从零学习，无需人类数据便超越 AlphaGo。
AlphaZero learned from scratch, outperforming AlphaGo without human data.

AlphaGo 结合直觉（策略网络）和计算（搜索）来掌握围棋。
AlphaGo combined intuition (policy network) and calculation (search) to master Go.
第 37 手表明 AI 能产生超越人类知识的洞见。
Move 37 showed AI can produce insights beyond human knowledge.
AlphaZero 通过从零开始学习、不使用人类数据而得到改进。
AlphaZero improved by learning from scratch without human data.
AlphaGo 的搜索技术被应用于矩阵乘法等科学问题。
AlphaGo's search techniques were applied to scientific problems like matrix multiplication.
代码等可验证领域对 AI 超越人类知识至关重要。
Verifiable domains like code are crucial for AI to go beyond human knowledge.
AlphaGo 标志着 AI 在特定领域超越人类智能的转折点。
AlphaGo marked a transition point where AI surpassed human intelligence in specific areas.

大语言模型是智能的捷径，但难以超越人类数据。
Large language models are a shortcut to intelligence, but struggle to go beyond human data.
AI 能生成正确但不可解释的证明，挑战人类理解。
AI can produce correct but uninterpretable proofs, challenging human understanding.
AlphaGo 优化胜率而非领先优势，导致反直觉的终局走法。
AlphaGo optimized for winning probability, not margin, leading to counterintuitive endgame play.
后训练中的强化学习现在是超越人类知识的关键。
Post-training with reinforcement learning is now key to pushing beyond human knowledge.
AlphaTensor 等 AI 智能体发现了人类 50 年来未发现的算法。
AI agents like AlphaTensor discovered algorithms humans had missed for 50 years.
可解释性重要，但对 AlphaFold 等科学突破并非必需。
Interpretability is important but not essential for scientific breakthroughs like AlphaFold.