AI编程工具的战争，大多讨论程序员写CRUD、网页、App的场景。但OpenAI最新实地报告把AI coding agents拉到一个被忽视的领域：科学计算。
7月28日OpenAI发布《Scientific computing in the age of agentic AI》，展示基因组学、气候建模、天文、粒子物理等领域研究者如何用AI coding agents加速科学软件开发。
被遗忘的"遗留代码"困境
科学计算有个公开秘密：大量关键科学软件是几十年前的老代码，没人敢动但所有人都在用。基因组学核心工具是2000年代Perl脚本跑在老服务器上，气候建模有大量Fortran 77代码，天文用IDL脚本——原作者早已离开，文档缺失，依赖过时，但没人敢重构怕结果不对。
因为科学软件正确性比一切都重要。商业软件出bug可以热修复；科学软件出bug可能导致Nature论文结论错误、药物研发方向被误导。所以科研人员对重构有极强心理障碍。
AI coding agents正在打破僵局。报告中基因组学研究者用Claude Code和GPT通读几万行无文档老Perl代码、解释功能、识别bug、自动翻译成Python——同时生成测试验证重构后输出与原代码一致。以前博士后几个月的重构，Agent辅助下几天做完。
"科学计算可能是AI coding agents被低估最深的场景。不是替代科学家，是把科学家从维护Fortran 77的噩梦里解放出来。"—— Dawn Vision编辑部
Agent重构科研工作方式
科学场景下Agent用法和商业开发很不同：商业开发让Agent从零写功能，科研里更多是"理解、翻译、验证、加速"——理解老代码、翻译成现代语言、写测试验证数值一致性、单线程改并行。Agent帮科学家从代码维护脱身，把时间花在需要科学直觉和创造力的部分。
这也解释了为什么AI编程工具在科研圈普及比预期快。科研人员不是程序员，学编程是为了解决科学问题。Agent充当"跨时代翻译器"——懂老语言也懂新语言，帮科研人员无痛迁移到现代工具链。跨学科研究也变容易：生物学家不用花一学期学R，气候科学家不用从头学Python遥感库，自然语言描述需求Agent写代码。
当然风险存在：Agent生成代码有微妙数值错误可能导致完全错误的科学结论。有经验的科研用户强调：Agent代码必须经严格单元测试和交叉验证，永远不能盲目信任。AI编程工具的影响力正从互联网公司代码库蔓延到大学实验室，那些跑了二十年的Fortran程序可能在一两年内被年轻科学家在Agent辅助下现代化。
科学家最宝贵的资源不是算力，是时间。AI正在把时间还给他们。
明天见。
Sources · 参考来源
声明：本文为 Dawn Vision 基于公开信息的二次创作与独立分析，标题、观点、行文均为原创，仅供参考，不构成任何投资建议或决策依据。如有侵权请联系删除。
本文基于 Dawn Vision 认知引擎处理的 6 个源信号生成，经编辑部人工审核。素材来源：OpenAI Blog。
相关入库笔记：OpenAI · 科学计算 · AI coding agent · 基因组学 · 科研软件 · 遗留代码重构 · Fortran · agentic科研
Most AI coding tool debates center on developers writing CRUD apps, web pages, and apps. But a new OpenAI field report pulls AI coding agents into an overlooked domain: scientific computing.
On July 28, OpenAI published "Scientific computing in the age of agentic AI," showcasing how researchers in genomics, climate modeling, astronomy, and particle physics use AI coding agents to accelerate scientific software development.
The Forgotten 'Legacy Code' Dilemma
Here's an open secret in scientific computing: vast amounts of critical scientific software is decades-old legacy code nobody dares touch but everyone relies on. Core genomics tools are 2000s Perl scripts on ancient servers; climate modeling has tons of Fortran 77; astronomy uses IDL scripts — original authors long gone, missing docs, outdated dependencies, yet nobody refactors for fear of producing different results.
Because in scientific software, correctness matters above all. A commercial bug gets hot-fixed; a scientific software bug could invalidate a Nature paper or mislead drug discovery. So researchers have extreme psychological barriers to refactoring.
AI coding agents are breaking this stalemate. In the report, genomics researchers use Claude Code and GPT to read tens of thousands of lines of undocumented Perl, explain functionality, identify bugs, auto-translate to Python — while generating tests to verify refactored outputs match originals. What took a postdoc months takes days with agent assistance.
"Scientific computing may be the most underappreciated scenario for AI coding agents. They don't replace scientists — they free them from maintaining Fortran 77."—— The Dawn Vision Editorial Desk
Agents Restructure How Research Works
Agents are used differently in science vs. commercial dev: instead of writing features from scratch, they're used for "understanding, translation, verification, acceleration" — understanding legacy code, translating to modern languages, writing tests for numerical consistency, parallelizing single-threaded code. Agents free scientists from code maintenance to focus on what requires scientific intuition and creativity.
This explains why AI coding tools are spreading faster in research than expected. Researchers aren't programmers; they code to solve scientific problems. Agents serve as "cross-era translators" — understanding both old and new languages, helping researchers painlessly migrate to modern toolchains. Interdisciplinary work gets easier: biologists don't need a semester of R; climate scientists don't need to learn Python remote sensing from scratch; describe needs in natural language and agents write code.
Risks exist: subtle numerical errors in agent-generated code could produce wrong scientific conclusions. Experienced researchers emphasize: agent code must go through rigorous unit testing and cross-validation — never blindly trust. AI coding tools' influence is spreading from tech company codebases to university labs. Those twenty-year-old Fortran programs may be modernized within a year or two by young scientists with agent assistance.
A scientist's most precious resource isn't compute — it's time. AI is giving them their time back.
See you tomorrow.
Sources · 参考来源
声明：本文为 Dawn Vision 基于公开信息的二次创作与独立分析，标题、观点、行文均为原创，仅供参考，不构成任何投资建议或决策依据。如有侵权请联系删除。
This article was generated by the Dawn Vision cognitive engine processing 6 source signals, with human editorial review. Sources: OpenAI Blog.
相关入库笔记：OpenAI · scientific computing · AI coding agents · genomics · research software · legacy code · Fortran · agentic research
科学计算可能是AI coding agents被低估最深的场景。不是替代科学家，是把科学家从维护Fortran 77的噩梦里解放出来。
—— Dawn Vision编辑部
Scientific computing may be the most underappreciated scenario for AI coding agents. They don't replace scientists — they free them from maintaining Fortran 77.
—— The Dawn Vision Editorial Desk
OpenAI · scientific computing · AI coding agents · genomics · research software · legacy code · Fortran · agentic research
Sources · 信源 Sources
本文基于 Dawn Vision 认知引擎处理的 6 个源信号生成，经编辑部人工审核。素材来源：OpenAI Blog。
This article was generated by the Dawn Vision cognitive engine processing 6 source signals, with human editorial review. Sources: OpenAI Blog.
