OpenAI 首席科学家 Jakub Pachocki 和技术研究员 Shimon Cedor 分享了他们从波兰高中到塑造 AI 未来的旅程，探讨了前沿研究的挑战以及强大 AI 系统的深远影响。
Jakub Pachocki and Shimon Cedor, OpenAI's chief scientist and technical fellow, discuss their journey from high school in Poland to shaping AI's future, the challenges of pioneering research, and the profound implications of powerful AI systems.
要点 · TL;DR
AGI 是一系列里程碑，而非单一事件。 AGI is a sequence of milestones, not a single event.
扩展推理模型将在未来几年加速 AI 进步。 Scaling reasoning models will accelerate AI progress in coming years.
迭代部署有助于在实际使用中学习和降低风险。 Iterative deployment helps learn and mitigate risks in real-world use.
核心观点 · Key points
AGI 是一系列里程碑，而非单一事件。 AGI is a sequence of milestones, not a single event.
扩展推理模型将在未来几年加速 AI 进步。 Scaling reasoning models will accelerate AI progress in coming years.
随着系统变得更强大，AI 对齐是一个紧迫的挑战。 AI alignment is a pressing challenge as systems become more powerful.
迭代部署有助于在实际使用中学习和降低风险。 Iterative deployment helps learn and mitigate risks in real-world use.
自动化 AI 研究是 OpenAI 正在努力实现的主要里程碑。 Automated AI research is a major milestone OpenAI is working toward.
AI 实验室的治理结构必须在早期就精心设计。 Governance structures for AI labs must be carefully designed early on.
反共识 · Contrarian takes
AI 系统可以纯粹从数据中学习语义，而非语法规则。 AI systems can learn semantics purely from data, not grammar rules.
神经网络即使有 bug 也常能学习，导致静默错误难以发现。 Neural networks often learn despite bugs, making silent errors hard to detect.
衡量 AI 能力正成为瓶颈；基准测试落后于实际能力。 Measuring AI capabilities is becoming a bottleneck; benchmarks lag behind.
AI 安全进步现在由能力需求而非仅伦理所保证。 AI safety progress is now guaranteed by capability needs, not just ethics.
AGI 一词正变得无意义，这可能意味着我们已接近。 The term AGI is becoming meaningless, which may mean we are close.
早期 OpenAI 充满冒名顶替综合征和尴尬氛围，与今天不同。 Early OpenAI had impostor syndrome and awkward vibes, unlike today.
本期章节 · Chapters（共 17）
开场与 OpenAI 早期Introduction and Early Days at OpenAI
AI 觉醒与 AlphaGo 影响AI awakening and AlphaGo impact
加入 OpenAIJoining OpenAI
早期 OpenAI 文化与技术不确定性Early OpenAI culture and technical uncertainty
GPT 发现背后的真实故事The real story behind GPT discovery
AI 日常：调试与研究自然现象Day-to-day work on AI: debugging and studying a natural phenomenon
最喜欢的协作与调试Favorite Collaboration and Debugging
教模型自己的思维方式Teaching models their own way of thinking
定义 AGI 及其临近性Defining AGI and its proximity
AI 实验室的责任Responsibility of AI labs
迈向 AGI 与迭代部署Transition to AGI and Iterative Deployment
控制强大 AI 的挑战The challenge of controlling powerful AI