Noam Brown 分享他从开发西塞罗到赢得世界外交锦标赛的历程，探讨 AI 在游戏中的演进及其对 AI 安全的影响。
Noam Brown discusses his journey from building Cicero to winning the World Diplomacy Championship, the evolution of AI in games, and the implications for AI safety.
要点 · TL;DR
推理模型需要足够的基础能力才能从思考中受益。 Reasoning models require sufficient base capability to benefit from thinking.
自我对弈仅在零和博弈中收敛到极小化极大，不适用于一般领域。 Self-play converges to minimax only in zero-sum games, not general domains.
深度研究表明推理模型在数学之外的非可验证领域也能成功。 Deep research shows reasoning models succeed in non-verifiable domains beyond math.
核心观点 · Key points
推理模型需要基础模型具备一定能力才能从思考中受益。 Reasoning models need a certain level of base model capability to benefit from thinking.
在两人零和博弈中，自我对弈收敛到极小化极大均衡，但在其他领域失效。 Self-play in two-player zero-sum games converges to a minimax equilibrium, but fails in other domains.
深度研究表明推理模型在非可验证领域也能成功，不仅限于数学和编程。 Deep research shows reasoning models succeed in non-verifiable domains, not just math and coding.
扩展推理算力面临收益递减和实际时间瓶颈。 Scaling test-time compute faces diminishing returns and wall-clock time bottlenecks.
数据效率是关键未解问题；模型样本效率低于人类。 Data efficiency is a key unsolved problem; models are less sample-efficient than humans.
反共识 · Contrarian takes
博弈论最优策略并非总是理想；剥削性玩法可能更有利可图。 Game theory optimal strategies are not always desirable; exploitative play can be more profitable.
AI 模型的辅助框架和路由器很可能被规模淘汰，而非永久存在。 Harnesses and routers for AI models will likely be washed away by scale, not permanent.
多智能体系统不应仅使用自我对弈；在零和博弈之外这是误导。 Multi-agent systems should not just use self-play; it's misguided outside zero-sum games.
显式世界模型可能非必需；隐式世界模型随规模涌现。 Explicit world models may not be needed; implicit world models emerge with scale.
人类智能并非窄带；文明进步源于长期合作与竞争。 Human intelligence is not a narrow band; civilization's progress comes from cooperation and competition over time.
本期章节 · Chapters（共 36）
引言与外交背景Introduction and Diplomacy Background
安全感知与外交基准Safety Perception and Diplomacy as Benchmark
氛围变化与轨迹Vibe changes and trajectory
深度研究在非可验证领域的证明Deep research as proof in non-verifiable domains
快慢思考类比局限Thinking fast and slow analogy limitations
AI 中的系统 1 与系统 2 思维System 1 vs System 2 Thinking in AI
模型路由器与扩展Model Routers and Scaling
给开发者的快速演进建议Advice for Developers on Rapid Evolution
伊利亚的尝试与推理时机Ilya's Attempt and Timing of Reasoning
预训练扩展的极限Limits of Pre-training Scaling
推理范式的兴起The Rise of Reasoning Paradigm
推理的内部辩论Internal Debate on Reasoning
扩展范式与数据效率Scaling paradigm and data efficiency
伊利亚·苏茨克维的洞见Insights from Ilya Sutskever
AI 编程：Codex 与 WinSurfCoding with AI: Codex and WinSurf
AGI 时刻与推理模型AGI moments and reasoning models
开发周期的断裂环节Broken parts of development cycle
软件工程之外的 AI 未来Future of AI beyond software engineering
对齐：安全与指令遵循Alignment: Safety vs Instruction Following
OpenAI 的多智能体团队Multi-Agent Team at OpenAI
扑克、GTO 与剥削Poker, GTO, and Exploitation
扑克中的剥削与博弈论最优Exploitative vs. Game Theory Optimal in Poker
世界模型与多智能体系统World Models and Multi-Agent Systems
超越 AlphaGo 的自对弈目标函数Objective Function for Self-Play Beyond AlphaGo
对 Sora 与自回归图像生成的印象Impression of Sora and Autoregressive Image Generation
图像生成的扩散与自回归对比Diffusion vs Autoregressive for Image Generation
对机器人技术与类人机器人的思考Thoughts on Robotics and Humanoids
紧跟研究前沿Keeping up with research
研究的环境限制因素Environmental limiters on research
测试时计算扩展瓶颈Test-time compute scaling wall
完美的人类生物学模型Perfect Models of Human Biology
中期训练与后期训练Mid-Training vs Post-Training
采访格雷格·布罗克曼Interviewing Greg Brockman
推荐游戏Recommended Games
AI 在不完美信息游戏中的应用AI for Imperfect Information Games
不完美信息游戏中的未解问题Unanswered questions in imperfect-information games