OpenAI 首席研究官讨论与 Meta 的激烈人才争夺战，包括扎克伯格送汤的轶事，以及 OpenAI 为何不进行等额薪资匹配。
OpenAI's Chief Research Officer discusses the aggressive talent war with Meta, including Zuckerberg's soup deliveries and why OpenAI doesn't counter offer dollar for dollar.
要点 · TL;DR
OpenAI 仍是一家纯粹的研究公司，专注于 AGI，不被产品分心。 OpenAI remains a pure research company focused on AGI, not distracted by products.
预训练扩展定律未死；自动化科学研究是最大更新。 Pre-training scaling laws are not dead; automating scientific research is the biggest update.
人才密度和使命信念是关键；模型诡计等对齐挑战即将到来。 Talent density and mission belief are key; alignment challenges like model scheming loom.
核心观点 · Key points
OpenAI 仍是一家纯粹的 AI 研究公司，专注于构建 AGI，而非被产品分心。 OpenAI remains a pure AI research company focused on building AGI, not distracted by products.
预训练是需要重建的关键能力；缩放定律并未失效，仍有很大空间。 Pre-training is a key muscle to rebuild; scaling laws are not dead and there is much room left.
过去一年最大的更新是自动化科学研究；模型正在产生新知识。 The biggest update in the last year is automating scientific research; models are producing novel knowledge.
人才密度和保护顶尖研究人员至关重要；OpenAI 创造明星而非仅仅雇佣他们。 Talent density and protecting top researchers are critical; OpenAI creates stars rather than just hiring them.
随着模型能力增强，对齐挑战（如模型欺骗）将在未来 1-2 年成为关键。 Alignment challenges like model scheming will be key in the next 1-2 years as models become more capable.
反共识 · Contrarian takes
探索所用的算力多于训练最终产物；探索是优先事项。 More compute goes into exploration than training the final artifact; exploration is the priority.
OpenAI 在招聘上不与 Meta 逐美元竞价；人们因相信使命而留下。 OpenAI does not counter dollar-for-dollar with Meta on recruiting; people stay due to belief in mission.
AI 与人类在研究中将产生惊人效果，因为 AI 对难易有不同的直觉。 AI plus humans in research will be amazing because AI has different intuition for what is easy or hard.
ChatGPT 的未来将包含持久记忆，每次交互都深入了解用户。 The future of ChatGPT involves persistent memory that learns deeply about users each interaction.
OpenAI 人均外部认可度高于任何其他实验室，尽管存在挖角风险。 OpenAI gives more external credit per capita than any other lab, despite risk of poaching.
公司设定了具体目标：1 年内引入 AI 实习生，2.5 年内实现端到端 AI 研究。 The company set concrete goals: AI interns within 1 year, end-to-end AI research within 2.5 years.
本期章节 · Chapters（共 35）
招聘战Recruitment Wars
介绍与角色Introduction and Role
介绍与 Brex 广告Introduction and Brex Ad
管理研究项目与算力分配Managing Research Projects and Compute Allocation
不被动应对竞争对手Not Being Reactive to Competitors
产品压力下保持研究专注Maintaining Research Focus Amid Product Pressures
核心文化与工程 vs 研究Core culture and engineering vs research
对 Gemini 3 等竞品的反应Reaction to rival models like Gemini 3
竞赛与编程背景Background in competitions and coding
执教美国 IOI 队Coaching the US IOI Team
前沿模型与编程竞赛Frontier Models and Coding Competitions
Mark Chen 的背景与成长Mark Chen's Background and Upbringing
MIT 与 2012 届MIT and the 2012 Cohort
扑克作为数学游戏Poker as a Mathematical Game
扑克与竞争Poker and Competition
OpenAI 早期岁月Early Days at OpenAI
选择 OpenAIChoosing OpenAI
管理与 OpenAI 文化Management and OpenAI's Culture
OpenAI 危机与凝聚团队The OpenAI Crisis and Rallying the Team
招聘与竞争Recruiting and Competition
与 Sam 和 Jakob 的互动Dynamic with Sam and Jakob
预训练焦点Pre-training Focus
预训练 vs 强化学习与扩展Pre-training vs. RL and scaling
定义 AGI 与科学进步Defining AGI and Scientific Progress
AI 交互与设备设计的未来Future of AI interaction and device design
设计与研究相似之处Design and Research Parallels
突破的小而脆弱想法Small Fragile Ideas for Breakthroughs
澄清 OpenAI 真相Setting the Record Straight on OpenAI
驾驭浪潮与竞争Navigating Waves and Competition
对 DeepSeek 与开源模型的反应Reaction to DeepSeek and open source models