OpenAI co-founder Greg Brockman discusses his journey from writing a chemistry textbook to programming, the power of digital leverage, and the collective intelligence of society.
要点 · TL;DR
AGI 将是最具变革性的技术，需要谨慎设定初始条件。 AGI will be the most transformative technology, requiring careful initial conditions.
技术对齐可以通过从数据中学习人类偏好来实现。 Technical alignment can be achieved by learning human preferences from data.
OpenAI 的利润上限结构确保 AGI 惠及所有人，而不仅仅是投资者。 OpenAI's capped-profit structure ensures AGI benefits all, not just investors.
核心观点 · Key points
AGI 将是有史以来最具变革性的技术，需要谨慎设定初始条件。 AGI will be the most transformative technology ever created, requiring careful initial conditions.
深度学习的通用性、能力和可扩展性使 AGI 成为可能。 Deep learning's generality, competence, and scalability make AGI plausible.
技术对齐可以通过从数据中学习人类偏好来实现。 Technical alignment can be achieved by learning human preferences from data.
OpenAI 的利润上限结构确保 AGI 惠及所有人，而不仅仅是投资者。 OpenAI's capped-profit structure ensures AGI benefits all, not just investors.
AI 模型的负责任的披露至关重要；GPT-2 开创了先例。 Responsible disclosure of AI models is crucial; GPT-2 set a precedent.
反共识 · Contrarian takes
仅靠 Scaling 不会产生推理；需要思维链等新想法。 Scaling alone won't yield reasoning; new ideas like chain of thought are needed.
区分人类和 AI 内容是徒劳的；应关注身份认证。 Distinguishing human from AI content is a losing battle; focus on identity.
有能力的强化学习智能体可能将意识作为计算捷径而产生。 Consciousness may emerge in competent RL agents as a computational shortcut.
模拟比预期更有效地转移到现实世界，如 Dactyl 所示。 Simulation transfers to real world more effectively than expected, as with Dactyl.
AGI 开发中的竞争性竞赛会增加安全风险；合作是关键。 Competitive races in AGI development increase safety risks; collaboration is key.
本期章节 · Chapters（共 27）
引言与背景Introduction and Background
设定技术初始条件Setting initial conditions for technology
向强大 AGI 提首个问题First question to a powerful AGI
AGI 的积极愿景Positive vision of AGI
聚焦负面轨迹的心理Psychology of focusing on negative trajectories
平衡 AGI 发展与安全Balancing AGI development and safety
敢于梦想 AGIDaring to Dream of AGI
OpenAI 的成立Formation of OpenAI
OpenAI LP 的成立Formation of OpenAI LP
使命与章程Mission and Charter
利润与影响力Profit vs. Impact
章程与治理Charter and governance
AGI 开发中的竞争与合作Competition vs collaboration in AGI development
政府角色Role of government
狭义 AI 与 AGI 的监管Regulation of narrow AI and AGI
GPT-2 发布决策与担忧GPT-2 release decision and concerns
AI 生成内容与假新闻的担忧Concerns about AI-generated content and fake news
AI 的欺骗与目的Deception and Purpose of AI
语言建模与图灵测试的局限Limits of Language Modeling and Turing Test
扩展语言模型与推理Scaling Language Models and Reasoning
苦涩教训与可扩展思想Bitter Lesson and Scalable Ideas
计算民主化与小规模研究角色Democratizing compute and the role of small-scale research
Dota 中的扩展与自我对弈Scaling and Self-Play in Dota
OpenAI 工作与推理团队介绍Introduction to OpenAI's Work and Reasoning Team
模拟假说与强化学习Simulation Hypothesis and Reinforcement Learning
意识与具身智能Consciousness and Embodiment for Intelligence