从扑克到外交：Noam Brown 谈 AI 与博弈论

From Poker to Diplomacy: Noam Brown on AI and Game Theory

诺姆·布朗 Noam Brown · No Priors 播客 · 2023-04-25 · 约 61 分钟 · 原视频 ↗

本期速览 · Overview

Noam Brown 讲述他从金融到 AI 研究的历程，专注于博弈论智能体以及塑造他工作的关键时刻。

Noam Brown discusses his journey from finance to AI research, focusing on game-theoretic agents and the pivotal moments that shaped his work.

要点 · TL;DR

仅靠扩展规模不够，推理和推理时计算对 AI 进步至关重要。
Scaling alone is insufficient; reasoning and inference-time compute are critical for AI progress.
像外交这样的游戏要求 AI 用自然语言与人类合作和谈判。
Games like Diplomacy require AI to cooperate and negotiate with humans in natural language.
自我对弈结合人类数据使 AI 在复杂游戏中超越人类表现。
Self-play combined with human data enables AI to surpass human performance in complex games.

核心观点 · Key points

仅靠 Scaling（规模扩张）是不够的；推理和推理时算力对 AI 进步至关重要。
Scaling alone is insufficient; reasoning and inference-time compute are critical for AI progress.
像 Diplomacy 这样的游戏要求 AI 用自然语言与人类合作和谈判。
Games like Diplomacy require AI to cooperate and negotiate with humans in natural language.
自我对弈结合人类数据使 AI 在复杂游戏中超越人类表现。
Self-play combined with human data enables AI to surpass human performance in complex games.
由于语言模型的进步，图灵测试不再是衡量 AI 智能的有用指标。
The Turing test is no longer a useful measure of AI intelligence due to progress in language models.
通用推理和样本效率仍然是实现 AGI（通用人工智能）的关键挑战。
General reasoning and sample efficiency remain key challenges for achieving AGI.

反共识 · Contrarian takes

数据不是瓶颈；推理时扩展算力比预训练更有前景。
Data is not the bottleneck; scaling compute during inference is more promising than pre-training.
没有规划的原始神经网络无法在围棋等游戏中达到顶级人类水平。
Raw neural nets without planning cannot match top human performance in games like Go.
在价格或薪资谈判等受限领域，AI 可以比人类谈判得更好。
AI can negotiate better than humans in constrained domains like price or salary negotiations.
扑克中的超额下注曾被视为错误，现在却是 AI 发现的有效策略。
Overbets in poker, once considered mistakes, are now effective strategies discovered by AI.
扑克中的纳什均衡保证期望上不输，这一概念曾遭许多人质疑。
Nash equilibrium in poker guarantees not losing in expectation, a concept many doubted.

本期章节 · Chapters（共 28）

0. 引言与诺姆背景 Introduction and Noam's background
1. AlphaGo 与模式匹配 AlphaGo and pattern matching
2. 有趣的机器人互动 Interesting Bot Interaction
3. 图灵测试相关性 Turing Test Relevance
4. 通用智能的衡量标准 Measures for General Intelligence
5. 推理的有前景方向 Promising Directions for Reasoning
6. 数据与规模化 Data and Scaling
7. 规模化极限与推理时计算 Scaling Limits and Inference-Time Compute
8. 外交数据与自我对弈 Diplomacy Data and Self-Play
9. AI 中的非语言交流与人类规范 Non-verbal communication and human norms in AI
10. 游戏作为 AI 进展的基准 Games as benchmarks for AI progress
11. 人类仍占主导的领域 Domains where humans still dominate
12. 通用性与样本效率 Generality and Sample Efficiency
13. 金融系统中的 AI AI in Financial Systems
14. AI 与人类谈判 AI Negotiation with Humans
15. 下一研究重点：推理 Next Research Focus: Reasoning
16. 领域特定规划算法的局限 Limitations of domain-specific planning algorithms
17. 研究中的局部最优风险 Risk of local minima in research
18. 给研究者的建议：敢于冒险 Advice for researchers: take risks
19. 外交游戏概述 Overview of Diplomacy game
20. 被人类误认为人类的机器人 Bots that humans thought were human
21. 外交与合作 Diplomacy and Cooperation
22. 扑克 AI 突破 Poker AI Breakthroughs
23. 扑克 AI 中的搜索发现 Discovery of Search in Poker AI
24. 多人扑克与可扩展搜索 Multiplayer Poker and Scalable Search
25. 对人类扑克策略的影响 Impact on Human Poker Strategy
26. 最优玩法与纳什均衡 Optimal Play and Nash Equilibrium
27. 扑克中的纳什均衡 Nash Equilibrium in Poker

阅读全文双语转录 →