# The First Open-Source Agentic AI Physicist
**作者**: Dr. Alex Wissner-Gross
**日期**: 2026-03-15T17:03:16.000Z
**来源**: [https://x.com/alexwg/status/2033227542392418458](https://x.com/alexwg/status/2033227542392418458)
---
The Singularity already writes code, trades stocks, and diagnoses tumors, but it has never done physics research end-to-end, until today. Physical Superintelligence PBC (PSI), a company I co-founded, is releasing Get Physics Done (GPD): the first open-source agentic AI physicist that can scope a physics problem, plan the research, carry out derivations and numerical checks, and verify its own results against the constraints that nature actually imposes. Built at PSI by physicists who needed it for their own work, GPD is now available to the global research community.
The field that gave us transistors (1947), nuclear energy (1951), and lasers (1960) still runs on the same artisanal production method it used during its last golden age: one theorist, one whiteboard, one career. Meanwhile, existing AI agents ship software, manage calendars, and close tickets, but none of them can tell you whether your Lagrangian is missing a boundary term. GPD can.
Watch it work:
GPD does three things that no tool has done before.
First, it is a copilot designed specifically for practicing physicists. You give it a research question. It asks clarifying questions to pin down scope, assumptions, notation, and verification targets. It builds a phased roadmap. Then it executes: derivations, numerical checks, literature work, and writing, producing LaTeX files, Python verification scripts, figures, and structured documentation that a working physicist can steer. GPD locks notation and sign conventions so that consistency holds as a project grows. No more discovering on page forty that your collaborator has been using the opposite metric signature since page three. The unit of work is a physics project, not a chat session.
Second, it is an AI peer reviewer for physics manuscripts. Before you submit, GPD can run a standalone review that checks dimensional consistency, limiting cases, symmetry constraints, conservation laws, and numerical stability. Physics has a built-in error-correction code written into the structure of reality itself; GPD actually uses it. It does not replace human referees. But it catches the classes of errors that consume referee time and delay publication: the errors you would rather find before submission than after. Every physicist who has waited months for a rejection over a sign error, or every experimentalist who built a six-month measurement program on a flawed prediction, knows why this matters.
Third, it introduces an autopilot mode for directed, autonomous physics research. Point GPD at a well-scoped problem and it will formulate the project, plan the phases, execute the derivations and numerical verification, and package the output with minimal human intervention. It compresses the time between asking a good question and getting a verified answer from weeks to hours.
Named as a nod to GSD (Get S*** Done), whose adoption proved that AI-native command workflows can work in practice, GPD is free, Apache 2.0 licensed, and runs inside Claude Code, Gemini CLI, Codex, and OpenCode.
Get started at github.com/psi-oss/get-physics-done.
Currently supported physics subfields and topics already include quantum field theory, quantum gravity, string theory, condensed matter, GR and cosmology, statistical mechanics, AMO, nuclear and particle, quantum information, fluid and plasma, mathematical physics, algebraic QFT, string field theory, classical mechanics, soft matter and biophysics, and astrophysics.
As Peter Diamandis and I argued in Solve Everything, physics is the key domino in the AI solution wavefront for every other field. Numerous breakthroughs downstream in chemistry, materials science, biology, and energy are bottlenecked on physics breakthroughs that no single human mind can reach alone. An open-source AI physicist that makes every working researcher more productive pushes that domino faster than any proprietary system could. If the community builds on GPD, extends it, and pressure-tests it against real problems, the compound effect accelerates the entire field.
For a century, the bottleneck on the next golden age of physics has been the scarcity of physicist-hours. That scarcity is not just a problem for physics departments; it is the reason the physical world has not kept pace with the digital one. And it ends today.
Star the repo, file issues, open PRs. The Singularity now has its first physicist on staff.
(Disclosure: I sit on the PSI board and have a financial interest in the company.)
## 相关链接
- [Dr. Alex Wissner-Gross](https://x.com/alexwg)
- [@alexwg](https://x.com/alexwg)
- [352K](https://x.com/alexwg/status/2033227542392418458/analytics)
- [Physical Superintelligence PBC](https://www.psi.inc/)
- [Get Physics Done](https://github.com/psi-oss/get-physics-done)
- [GSD (Get S*** Done)](https://github.com/gsd-build/get-shit-done)
- [github.com/psi-oss/get-physics-done](https://github.com/psi-oss/get-physics-done)
- [Solve Everything](https://solveeverything.org/)
- [Star the repo, file issues, open PRs.](https://github.com/psi-oss/get-physics-done)
- [PSI](https://www.psi.inc/)
- [Upgrade to Premium](https://x.com/i/premium_sign_up)
- [1:03 AM · Mar 16, 2026](https://x.com/alexwg/status/2033227542392418458)
- [352.9K Views](https://x.com/alexwg/status/2033227542392418458/analytics)
- [View quotes](https://x.com/alexwg/status/2033227542392418458/quotes)
---
*导出时间: 2026/3/16 23:27:46*
---
## 中文翻译
# 首个开源智能体 AI 物理学家
**作者**: Dr. Alex Wissner-Gross
**日期**: 2026-03-15T17:03:16.000Z
**来源**: [https://x.com/alexwg/status/2033227542392418458](https://x.com/alexwg/status/2033227542392418458)
---
“奇点”已经能写代码、炒股和诊断肿瘤，但它从未端到端地完成过物理研究——直到今天。Physical Superintelligence PBC（PSI），一家由我共同创立的公司，正式发布 Get Physics Done（GPD）：首个开源智能体 AI 物理学家，它能够界定物理问题、规划研究、执行推导和数值验证，并根据自然界实际施加的约束来核验自身结果。GPD 由 PSI 中需要它来完成自身工作的物理学家打造，现已面向全球研究界开放。
这个孕育了晶体管（1947）、核能（1951）和激光（1960）的领域，仍然沿用它上一个黄金时代所使用的同一种手工作坊式生产方式：一位理论家、一块白板、一整段职业生涯。与此同时，现有的 AI 智能体可以交付软件、管理日程、关闭工单，但没有一个能告诉你你的拉格朗日量是否漏掉了一个边界项。GPD 可以。
观看它的工作过程：
GPD 做到了三件前所未有的工作。
第一，它是专为执业物理学家设计的副驾驶。你给它一个研究问题，它会提出澄清性问题以确定范围、假设、符号约定和验证目标。它会制定分阶段的路线图，然后执行：推导、数值检验、文献工作和写作，产出 LaTeX 文件、Python 验证脚本、图表以及结构化文档，供一线物理学家掌控和引导。GPD 会锁定符号和符号约定，确保项目在扩展过程中保持一致性。再也不用在第四十页才发现合作者从第三页起就一直在用相反的度规符号。工作的单位是物理项目，而不是一次聊天会话。
第二，它是针对物理学稿件的 AI 同行评审员。在投稿之前，GPD 可以运行一次独立评审，检查量纲一致性、极限情形、对称性约束、守恒定律和数值稳定性。物理学内置了一套写进现实结构本身的纠错码；GPD 真正使用它。它不会取代人类审稿人，但它能抓住那些消耗审稿人时间、延误发表的类别错误：那些你宁愿在投稿前而不是投稿后发现的问题。每一位曾因一个符号错误等待数月才收到拒稿信的理论学家，或每一位曾基于一个有缺陷的预测启动六个月测量计划的实验学家，都明白这为什么重要。
第三，它引入了一种面向定向、自主物理研究的自动驾驶模式。将一个界定清晰的问题交给 GPD，它会制定项目方案、规划阶段、执行推导和数值验证，并在最少人工干预的情况下打包输出。它把从提出一个好问题到得到一个经过验证的答案之间的时间从数周压缩到数小时。
它的命名是对 GSD（Get S*** Done）的致敬——后者的普及证明了 AI 原生命令行工作流在实践中是可行的。GPD 免费、采用 Apache 2.0 许可证，可在 Claude Code、Gemini CLI、Codex 和 OpenCode 中运行。
请访问 github.com/psi-oss/get-physics-done 开始使用。
目前支持的物理学子领域和主题已包括：量子场论、量子引力、弦论、凝聚态、广义相对论与宇宙学、统计力学、原子分子光物理（AMO）、核物理与粒子物理、量子信息、流体与等离子体、数学物理、代数量子场论、弦场论、经典力学、软物质与生物物理，以及天体物理学。
正如 Peter Diamandis 和我在《Solve Everything》中所论证的，物理学是 AI 解决方案波前上撬动所有其他领域的关键多米诺骨牌。化学、材料科学、生物学和能源领域的众多下游突破，都受制于任何单个人类头脑无法企及的物理学突破。一个让每一位在岗研究人员都更具生产力的开源 AI 物理学家，比任何专有系统都能更快地推倒这块多米诺骨牌。如果社区基于 GPD 进行构建、扩展并在真实问题上对其施加压力测试，其复合效应将加速整个领域的进程。
一个世纪以来，下一个物理学黄金时代的瓶颈一直是物理学家工时的稀缺。这种稀缺不仅是物理系的问题，也是物理世界未能跟上数字世界步伐的原因。而这种稀缺，从今天起终结了。
为仓库点星、提交 issue、发起 PR。奇点现在有了它的第一位在编物理学家。
（披露：我是 PSI 董事会成员，并持有该公司的经济利益。）
## 相关链接
- [Dr. Alex Wissner-Gross](https://x.com/alexwg)
- [@alexwg](https://x.com/alexwg)
- [352K](https://x.com/alexwg/status/2033227542392418458/analytics)
- [Physical Superintelligence PBC](https://www.psi.inc/)
- [Get Physics Done](https://github.com/psi-oss/get-physics-done)
- [GSD (Get S*** Done)](https://github.com/gsd-build/get-shit-done)
- [github.com/psi-oss/get-physics-done](https://github.com/psi-oss/get-physics-done)
- [Solve Everything](https://solveeverything.org/)
- [为仓库点星、提交 issue、发起 PR。](https://github.com/psi-oss/get-physics-done)
- [PSI](https://www.psi.inc/)
- [Upgrade to Premium](https://x.com/i/premium_sign_up)
- [2026年3月16日 凌晨1:03](https://x.com/alexwg/status/2033227542392418458)
- [352.9K 次浏览](https://x.com/alexwg/status/2033227542392418458/analytics)
- [查看引用](https://x.com/alexwg/status/2033227542392418458/quotes)
---
*导出时间: 2026/3/16 23:27:46*
