AI 连接物理世界：Periodic Labs 创始人 Liam Fedus 谈材料科学革命AI Meets the Physical World: Periodic Labs Co-Founder Liam Fedus on Revolutionizing Materials Science
The Takeaway: 要真正加速科学和技术进步，就必须将 AI 系统与物理世界相连，通过实验驱动的闭环学习来推进材料工程。
Periodic Labs 联合创始人 Liam Fedus（前 OpenAI 后训练 VP，曾在 Google Brain 工作）拥有物理学背景，曾参与暗物质研究和早期 transformer 等创新。他解释说，物理学家大量进入 AI 领域，是因为该领域提供高杠杆且严谨的思考方式，尤其在高能物理瓶颈后。
Fedus 强调，语言模型虽强大，但 2022 年的 ChatGPT 技术还不足以实现物理世界的连接；如今的推理和代理能力才使之可能。数据方面，他们利用现有模型的数万亿 token 先验知识，但在特定化学空间中需要实验数据来提供真实 grounding，并形成交互闭环：分析异常、与模拟和文献一致性，然后指导下一次实验。
架构上，语言模型作为 orchestration layer，指导专门针对原子系统的对称感知神经网络，作为工具和奖励函数。Fedus 指出，物理科学将类似于 LLM 时代，建立 scaling properties，通过自动化和智能系统大规模实验。
他最兴奋的是多学科团队的迭代：物理学家、化学家与顶尖 AI 研究者紧密合作，改变传统研究方式。
“Science ultimately isn't sitting in a room thinking really hard. You have to conduct experiments. You have to learn from them. You have to interface with reality.”
在十年愿景中，成功将带来原子级合成代理，加速半导体、航空航天和能源等领域，让物理世界的发展跟上数字世界的步伐。
Periodic Labs 联合创始人 Liam Fedus（前 OpenAI 后训练 VP，曾在 Google Brain 工作）拥有物理学背景，曾参与暗物质研究和早期 transformer 等创新。他解释说，物理学家大量进入 AI 领域，是因为该领域提供高杠杆且严谨的思考方式，尤其在高能物理瓶颈后。
Fedus 强调，语言模型虽强大，但 2022 年的 ChatGPT 技术还不足以实现物理世界的连接；如今的推理和代理能力才使之可能。数据方面，他们利用现有模型的数万亿 token 先验知识，但在特定化学空间中需要实验数据来提供真实 grounding，并形成交互闭环：分析异常、与模拟和文献一致性，然后指导下一次实验。
架构上，语言模型作为 orchestration layer，指导专门针对原子系统的对称感知神经网络，作为工具和奖励函数。Fedus 指出，物理科学将类似于 LLM 时代，建立 scaling properties，通过自动化和智能系统大规模实验。
他最兴奋的是多学科团队的迭代：物理学家、化学家与顶尖 AI 研究者紧密合作，改变传统研究方式。
“Science ultimately isn't sitting in a room thinking really hard. You have to conduct experiments. You have to learn from them. You have to interface with reality.”
在十年愿景中，成功将带来原子级合成代理，加速半导体、航空航天和能源等领域，让物理世界的发展跟上数字世界的步伐。
The Takeaway: True acceleration in science and technology requires connecting AI systems to the physical world through experiment-driven closed-loop learning to advance materials engineering.
Periodic Labs co-founder Liam Fedus (former VP of Post-Training at OpenAI, previously at Google Brain) has a physics background, including dark matter research and early innovations like transformers. He explains why so many physicists enter AI: it offers high-leverage, principled thinking, especially after bottlenecks in high-energy physics post-Higgs.
Fedus stresses that while language models are powerful, 2022-era ChatGPT tech wasn't sufficient for physical-world connection; today's reasoning and agent capabilities make it possible. On data, they leverage trillions of tokens from existing models for foundational priors, but need experimental data for grounding in specific chemical spaces, forming an interactive closed loop: spotting aberrations, checking consistency with simulations and literature, then driving the next experiments.
Architecturally, language models act as an orchestration layer directing specialized symmetry-aware neural nets for atomic systems as tools and reward functions. Fedus notes physical sciences will mirror the LLM era by establishing scaling properties and running massive experiments via automation and intelligence.
He's most excited about multidisciplinary iteration: physicists and chemists working closely with top AI researchers, fundamentally changing decades-old research practices.
"Science ultimately isn't sitting in a room thinking really hard. You have to conduct experiments. You have to learn from them. You have to interface with reality."
In a ten-year vision, success means agency for atomic rearrangement and synthesis, accelerating semiconductors, aerospace, and energy so the physical world keeps pace with digital change.
查看原文 →
Periodic Labs co-founder Liam Fedus (former VP of Post-Training at OpenAI, previously at Google Brain) has a physics background, including dark matter research and early innovations like transformers. He explains why so many physicists enter AI: it offers high-leverage, principled thinking, especially after bottlenecks in high-energy physics post-Higgs.
Fedus stresses that while language models are powerful, 2022-era ChatGPT tech wasn't sufficient for physical-world connection; today's reasoning and agent capabilities make it possible. On data, they leverage trillions of tokens from existing models for foundational priors, but need experimental data for grounding in specific chemical spaces, forming an interactive closed loop: spotting aberrations, checking consistency with simulations and literature, then driving the next experiments.
Architecturally, language models act as an orchestration layer directing specialized symmetry-aware neural nets for atomic systems as tools and reward functions. Fedus notes physical sciences will mirror the LLM era by establishing scaling properties and running massive experiments via automation and intelligence.
He's most excited about multidisciplinary iteration: physicists and chemists working closely with top AI researchers, fundamentally changing decades-old research practices.
"Science ultimately isn't sitting in a room thinking really hard. You have to conduct experiments. You have to learn from them. You have to interface with reality."
In a ten-year vision, success means agency for atomic rearrangement and synthesis, accelerating semiconductors, aerospace, and energy so the physical world keeps pace with digital change.