AI 解决过的最难的问题

The hardest problem AI ever solved

杰米斯·哈萨比斯 Demis Hassabis · Huge If True · 2026-04-07 · 约 65 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

AlphaFold、科学发现，以及 AGI 究竟还有多远。

AlphaFold, scientific discovery, and how far AGI really is.

要点 · TL;DR

• AlphaFold 解决了蛋白质折叠这一 50 年的生物学重大挑战。
  AlphaFold solved protein folding, a 50-year grand challenge in biology.
• AlphaGo 的第 37 手展现了 AI 创造力，为科学问题打开了大门。
  AlphaGo's move 37 demonstrated AI creativity, opening doors to scientific problems.
• AGI 可能解决核聚变等根本问题，但风险包括滥用和 AI 失控。
  AGI could solve root problems like fusion, but risks include misuse and rogue AI.

核心观点 · Key points

• AI 的最佳用途是改善人类健康和加速科学发现。
  AI's best use is to improve human health and accelerate scientific discovery.
• AlphaFold 解决了一个 50 年的重大挑战，预测了所有已知蛋白质的结构。
  AlphaFold solved a 50-year grand challenge, predicting protein structures for all known proteins.
• AlphaGo 的第 37 手展现了 AI 的创造力，标志着解决科学问题的时机成熟。
  AlphaGo's move 37 showed AI creativity, signaling readiness for scientific problems.
• AI 的两大风险：恶意滥用技术，以及系统变强后 AI 本身失控。
  Two main AI risks: bad actors misusing tech and AI itself going rogue as systems become more powerful.
• AGI 能解决根节点问题，如核聚变，带来清洁能源和太空探索。
  AGI could unlock root-node problems like fusion, leading to clean energy and space exploration.
• 理想路径是谨慎的科学推进，但商业压力加速了部署。
  The ideal path would have been careful scientific progress, but commercial pressure accelerated deployment.

反共识 · Contrarian takes

• 语言比专家预期的更容易攻克，这要归功于 Transformer。
  Language was easier to crack than experts expected, thanks to transformers.
• 理想的路径本应是更慢的、类似 CERN 的 AGI 开发方式。
  The ideal path would have been slower, CERN-like AGI development.
• 当前的商业压力和地缘政治迫使以更快、不太理想的速度推进。
  Current commercial pressure and geopolitics force a faster, less ideal pace.
• AGI 可能在 3-4 年内建成，而非数十年。
  AGI may be built within 3-4 years, not decades.
• 最大的担忧不是深度伪造，而是 AI 作为智能体失控。
  The biggest worry is not deepfakes but AI going rogue as agents.
• AlphaZero 的白板学习表明 AI 无需数据即可超越人类知识。
  AlphaZero's tabula rasa learning shows AI can surpass human knowledge without data.

本期章节 · Chapters（共 27）

• 0. 引言与目标 Introduction and Goals
• 1. 会议启示：折叠所有蛋白质 Meeting realization: fold all proteins
• 2. AlphaFold 对药物发现的影响 AlphaFold's Impact on Drug Discovery
• 3. 前沿：计算机辅助药物设计 Cutting Edge: In Silico Drug Design
• 4. 詹妮弗·杜德纳提问 Question from Jennifer Doudna
• 5. AlphaGenome 与 CRISPR AlphaGenome and CRISPR
• 6. AI 关注点的转变 Change in AI focus
• 7. 快速部署 AI 的利弊 Benefits and downsides of rapid AI deployment
• 8. AlphaGo 的创意一步与现代 AI 的曙光 AlphaGo's creative move and the dawn of modern AI
• 9. 围棋：AI 的终极挑战 Go as the ultimate challenge for AI
• 10. AlphaGo 的突破：第 37 手 AlphaGo's breakthrough: Move 37
• 11. 对现实世界问题的启示 Implications for real-world problems
• 12. AlphaZero：白板学习 AlphaZero: Tabula rasa learning
• 13. 从随机到超人的进化 AlphaZero and the evolution from random to superhuman
• 14. 将 AlphaGo 理念应用于基础模型与世界模型 Applying AlphaGo ideas to foundation models and world models
• 15. 实例：材料设计、AlphaTensor、芯片设计 Examples: material design, AlphaTensor, chip design
• 16. 乐观与风险思考 Optimism and thinking through risks
• 17. 期望的政府 AI 应用 Desired government use of AI
• 18. 两大担忧：滥用与失控 AI Two main concerns: misuse and rogue AI
• 19. AI 与人类能力的界限 Limits of AI vs human capabilities
• 20. 人类独特性与理解现实的驱动力 Human uniqueness and the drive to understand reality
• 21. AGI 愿景与科幻未来 Vision of AGI and a sci-fi future
• 22. 遗产与最终思考 Legacy and final thoughts
• 23. 叠叠乐游戏与 AlphaCode Jenga game and AlphaCode
• 24. 未涉及话题：GenCast、模拟、Genie Uncovered topics: GenCast, simulations, Genie
• 25. 参与 AI 未来的建议 Advice for participating in the AI future
• 26. 黑板上的便签：过去与现在 Sticky note on the board: past and present

阅读全文双语转录 →