Can Vibe Coding Work for Science? Argonne Puts the Idea to the Test
Nearly a year after OpenAI cofounder Andrej Karpathy coined the phrase in a February 2025 social media post, “vibe coding” has become shorthand for using LLMs to turn ideas into code with minimal manual programming. Karpathy described the approach as one where developers “fully give into the vibes, embrace exponentials, and forget that the code even exists.”
Vibe coding has since gained traction in software prototyping, but its relevance for rigorous scientific workloads remains an open question. Now, researchers at Argonne National Laboratory are putting vibe coding tools to the test to see whether they can move scientific ideas into working code more quickly.
At a recent Argonne hackathon, researchers from across the laboratory gathered to evaluate commercial AI-assisted coding tools such as Cursor and Warp across scientific problems ranging from nuclear energy optimization to particle physics, testing whether the tools could handle real research workloads. Argonne’s interest in the approach reflects its long-standing role in computational science. The laboratory is home to Aurora, one of the world’s most powerful supercomputers, and has played a central role in advancing AI-driven scientific computing across the DOE research ecosystem.
Rick Stevens
Rick Stevens, associate laboratory director for Computing, Environment and Life Sciences at Argonne, described vibe coding as a productivity hack that can help scientists “work as fast as they can think” in a recent article published by the laboratory. “You’re unhobbled from your coding speed,” he said.
Stevens said vibe coding allows researchers to interact with LLMs in real time, often through spoken or natural language prompts, and receive usable code or analysis in seconds or minutes rather than hours. He likens the experience to working alongside an AI co-scientist and encourages researchers to spend time with the tools, experiment, and learn what works: “Think, play, and have a blast!”
In his view, the value of these systems extends beyond speed, freeing scientists from routine coding tasks and allowing them to focus their effort on the more human aspects of science, like formulating ideas, asking better questions, and exercising scientific judgment. “With these tools, you’re not bottlenecked by writing code,” he said. “Now, you’re focused on ideas.”
In the laboratory’s telling, the hackathon felt less like a typical coding session and more like an open experiment: “At the hackathon, the vibe in the room was playful. The group was a mix of coders and non-coders from a variety of disciplines. Instead of quietly pecking away at their keyboards, researchers were laughing, bouncing ideas off each other, and confidently speaking commands to their laptops,” the lab wrote.
Several teams at the hackathon explored practical ways the tools could be put to work. Nuclear engineer Yeni Li said vibe coding is useful for building software to create AI models of nuclear power plants, which will help plant engineers and managers predict the best times for maintenance. “These tools will help us do a few days of work in a single afternoon,” said Li.
In bioscience, researchers examined how natural language prompts could help automate Python-based data processing pipelines, lowering the barrier for scientists with limited coding experience. Molecular biologist Rosemarie Wilton tested the approach by using AI-assisted tools to generate command codes for routine analysis workflows, noting that the approach can be a time saver for novice coders.
Rosemarie Wilton (right) and Sarah Owens tested AI workflows in biosciences at the 1000 Scientist Jam in 2025 (Credit: Argonne National Laboratory)
In scientific infrastructure, researchers explored whether vibe coding could help modernize legacy software more quickly. Zachary Sherman, a software developer who manages open-source Python tools for the Atmospheric Radiation Measurement program, said he is looking to use vibe coding tools to translate older codebases written in multiple languages into Python. Sherman noted that many atmospheric science tools and data interfaces are outdated and difficult to maintain. He said vibe coding could accelerate the creation of modern Python tools for interacting with atmospheric datasets, while also making it easier to troubleshoot problems and reduce costs.
In particle physics, researchers examined whether vibe coding could help scale the merging of massive datasets generated by simulations and experiments. Nuclear physicist Chiara Bissolotti and theoretical particle physicist Tim Hobbs are working with data from sources including simulations and experiments conducted at facilities such as CERN and the planned Electron-Ion Collider at Brookhaven National Laboratory. Hobbs said the challenge lies in whether datasets from different sources can be made to “talk to each other,” uncovering shared patterns or pointing toward new theoretical predictions. Bissolotti said vibe coding could help by generating the scaffolding needed to make those comparisons more scalable and efficient.
For large, interdisciplinary modeling efforts tied to national security and public health, computational scientist Jonathan Ozik said vibe coding could improve collaboration across teams with different technical backgrounds. Ozik uses supercomputers and simulations to study complex systems such as biological processes, health care interventions, and the spread of infectious diseases in urban environments. He said AI-assisted tools could help summarize code and context when switching between projects.
By lowering the effort required to engage with complex codebases, Ozik said vibe coding could expand the range of ideas researchers are willing to pursue: “If you have fewer perceived barriers, you create new possibilities. Things that were previously infeasible in science will become common.”
For now, vibe coding remains an exploratory approach for scientific workflows. Its role is still being defined, but these early experiments suggest it could become a useful complement to existing methods, particularly in lowering barriers to prototyping, collaboration, and iteration across disciplines. As AI capabilities advance, it will be fascinating to see how researchers will adapt these tools to work faster, think more creatively, and tackle problems that once seemed out of reach.
Note: This article is based on reporting and interviews published by Argonne. AIwire thanks the laboratory for sharing its work.
Related
Can Vibe Coding Work for Science? Argonne Puts the Idea to the Test
Nearly a year after OpenAI cofounder Andrej Karpathy coined the phrase in a February 2025…
AI to Augment More Jobs Than It Replaces Through 2030, Forrester Says
A recent jobs report from Forrester suggests that while artificial intelligence will reshape parts of…
Berkeley Lab: EcoFABs Could Help Fuel AI in Agriculture
Jan. 23, 2026 — Harnessing the power of artificial intelligence to study plant microbiomes —…
Nokia and Hypertec Deploy Nibi Supercomputer at University of Waterloo
Jan. 23, 2026 — Nokia, a world-class leader in IP networking and AI-HPC data center fabric,…
TACC: Scaling Up Computing Education in a Time of AI
Jan. 23, 2026 — Artificial intelligence is not only a leading driver of the U.S….
Sandia: Physicists Employ AI Labmates to Supercharge LED Light Control
Fourfold improvement bridges basic science with next-generation technology ALBUQUERQUE, N.M., Jan. 22, 2026 — In 2023,…
