ChatGPT 联合创始人、前 OpenAI 副总裁 Liam Fetus 分享他从物理学到 AI 的历程，以及他正在构建的面向原子的 AI 基础实验室。
Liam Fetus, co-creator of ChatGPT and former VP at OpenAI, discusses his journey from physics to AI and his new venture building an AI foundation lab for atoms.
要点 · TL;DR
AI 应用于物理世界需要闭环实验和数据。 AI for physical world requires closed-loop experiments and data.
语言模型在材料科学中协调专用模型。 Language models orchestrate specialized models in materials science.
AI 在软件领域的自我改进无法泛化到其他领域。 AI self-improvement in software does not generalize to other fields.
核心观点 · Key points
将 AI 与物理世界连接对于加速科学和技术至关重要。 Connecting AI to the physical world is essential for accelerating science and technology.
实验和数据的闭环系统是 AI 驱动材料发现的关键。 A closed-loop system of experiments and data is key for AI-driven materials discovery.
语言模型作为材料科学中专用模型的编排层。 Language models serve as an orchestration layer for specialized models in materials science.
AI 在软件工程中的自我改进正在发生，但在其他领域尚未实现。 AI self-improvement in software engineering is happening now, but not yet in other domains.
缩放定律和大规模实验将推动物理科学的进步，类似于 AI 领域。 Scaling laws and large-scale experiments will drive progress in physical sciences, similar to AI.
反共识 · Contrarian takes
智能不是标量；AI 系统具有尖峰能力，在狭窄领域表现出色。 Intelligence is not a scalar; AI systems have spiky capabilities, excelling in narrow domains.
AI 在软件工程中的自我改进不会自动泛化到生物学或其他领域。 AI self-improvement in software engineering does not automatically generalize to biology or other fields.
文献中的实验数据可能不可靠，同一属性的数值跨越多个数量级。 Experimental data from literature can be unreliable, spanning orders of magnitude for the same property.
从量子力学到流体动力学的泛化有限；模型需要特定领域的数据。 Generalization from quantum mechanics to fluid dynamics is limited; models need domain-specific data.
在 AI 驱动的材料科学中，算力成本通常超过物理基础设施成本。 Compute cost often exceeds physical infrastructure cost in AI-driven materials science.
本期章节 · Chapters（共 10）
引言与背景Introduction and Background
测试时推理与物理世界连接Test-time inference and physical world connection
系统架构与延迟System Architecture and Latency
商业化策略Commercialization Strategy
愿景与未来影响Vision and Future Impact
跨学科合作与物理科学规模化Interdisciplinary Collaboration and Scaling in Physical Sciences
AGI 与智能的尖峰性AGI and the Spikiness of Intelligence
软件工程与 AI 研究中的闭环系统Closed-loop systems in software engineering and AI research
机器人技术在闭环系统中的作用Role of robotics in closed-loop systems
对 AI 与机器人超越周期的期待Excitement about AI and robotics beyond Periodic