AI for Atoms：Liam Fedus 讲述 Periodic Labs 如何用 AI 革新材料工程AI for Atoms: Liam Fedus on Revolutionizing Materials Engineering with Periodic Labs
核心要点：AI 必须通过闭环实验与物理世界连接，才能真正加速材料科学等领域的科技进步。
Periodic Labs 联合创始人 Liam Fedus 曾任 OpenAI 后训练副总裁，帮助将预训练的 GPT-4 转化为 ChatGPT，他凭借物理本科背景和 Google Brain 经验（参与了 transformer、Mixture of Experts 以及稀疏性技术）来构建针对原子的 AI 系统。他指出，许多物理学家进入 AI 领域是因为其高杠杆效应和严谨的思维方式，尤其是在 Higgs 发现后高能物理面临瓶颈的情况下。
科学需要通过实验与现实交互，而非仅靠思考。早期 ChatGPT 时代的模型还不足以胜任，但推理、测试时推理、工具使用和 agent 的进步使其成为可能。Periodic Labs 将 LLM 作为编排层，指导针对原子系统的专用对称性感知模型，所有操作都处于文献、模拟和真实实验数据的交互闭环中，从而 grounding 一切并驱动迭代发现。
物理学家、化学家和顶级 AI 工程师之间的这种多学科协作正在改变研究方式，像 ML 用 GPU 扩展一样扩展实验规模。Fedus 设想 AI 将赋予“原子重排合成的能力”，让半导体、航空航天和能源领域的物理创新速度赶上软件的步伐。他说：“我们正在努力赋予人类原子重排合成的能力”，加上机器人作为加速器，十年后生活将大不相同。
Periodic Labs 联合创始人 Liam Fedus 曾任 OpenAI 后训练副总裁，帮助将预训练的 GPT-4 转化为 ChatGPT，他凭借物理本科背景和 Google Brain 经验（参与了 transformer、Mixture of Experts 以及稀疏性技术）来构建针对原子的 AI 系统。他指出，许多物理学家进入 AI 领域是因为其高杠杆效应和严谨的思维方式，尤其是在 Higgs 发现后高能物理面临瓶颈的情况下。
科学需要通过实验与现实交互，而非仅靠思考。早期 ChatGPT 时代的模型还不足以胜任，但推理、测试时推理、工具使用和 agent 的进步使其成为可能。Periodic Labs 将 LLM 作为编排层，指导针对原子系统的专用对称性感知模型，所有操作都处于文献、模拟和真实实验数据的交互闭环中，从而 grounding 一切并驱动迭代发现。
物理学家、化学家和顶级 AI 工程师之间的这种多学科协作正在改变研究方式，像 ML 用 GPU 扩展一样扩展实验规模。Fedus 设想 AI 将赋予“原子重排合成的能力”，让半导体、航空航天和能源领域的物理创新速度赶上软件的步伐。他说：“我们正在努力赋予人类原子重排合成的能力”，加上机器人作为加速器，十年后生活将大不相同。
The Takeaway: AI must connect to the physical world via closed-loop experiments to truly accelerate science and technology in materials and beyond.
Liam Fedus, co-founder of Periodic Labs and former VP of post-training at OpenAI—who helped turn the pretrained GPT-4 into ChatGPT—brings his physics major background and Google Brain experience (where he contributed to transformers, mixture of experts, and sparsity techniques) to build AI systems for atoms. He notes that many physicists entered AI for its high leverage and principled thinking, especially after bottlenecks in high-energy physics post-Higgs discovery.
Science requires interfacing with reality through experiments, not just thinking. Early ChatGPT-era models weren't capable, but advances in reasoning, test-time inference, tool use, and agents now make it possible. Periodic Labs uses LLMs as an orchestration layer directing specialized symmetry-aware models for atomic systems, all within an interactive loop of literature, simulations, and real experimental data that grounds everything and drives iterative discovery.
This multidisciplinary collaboration between physicists, chemists, and top AI engineers is transforming research, scaling experiments like ML scaled with GPUs. Fedus envisions AI granting 'agency for atomic rearrangement synthesis,' speeding physical innovation in semiconductors, aerospace, and energy to match software's pace. "We're trying to give humanity this agency for atomic rearrangement synthesis," he explains, and with robotics as an accelerator, life will feel dramatically different in ten years.
查看原文 →
Liam Fedus, co-founder of Periodic Labs and former VP of post-training at OpenAI—who helped turn the pretrained GPT-4 into ChatGPT—brings his physics major background and Google Brain experience (where he contributed to transformers, mixture of experts, and sparsity techniques) to build AI systems for atoms. He notes that many physicists entered AI for its high leverage and principled thinking, especially after bottlenecks in high-energy physics post-Higgs discovery.
Science requires interfacing with reality through experiments, not just thinking. Early ChatGPT-era models weren't capable, but advances in reasoning, test-time inference, tool use, and agents now make it possible. Periodic Labs uses LLMs as an orchestration layer directing specialized symmetry-aware models for atomic systems, all within an interactive loop of literature, simulations, and real experimental data that grounds everything and drives iterative discovery.
This multidisciplinary collaboration between physicists, chemists, and top AI engineers is transforming research, scaling experiments like ML scaled with GPUs. Fedus envisions AI granting 'agency for atomic rearrangement synthesis,' speeding physical innovation in semiconductors, aerospace, and energy to match software's pace. "We're trying to give humanity this agency for atomic rearrangement synthesis," he explains, and with robotics as an accelerator, life will feel dramatically different in ten years.