Tips from Early Adopters

In recent advances in climate research, Zeke Hausfather has found innovative ways to visualize the urgency of climate change. Last year, he tapped into AI technology to create eye-catching climate-data visualizations, seeking fresh and impactful methods to convey the rapid warming of Earth. Collaborating with an AI tool, Hausfather generated striking tree-ring-style graphics that encapsulated temperature changes over time, complete with vibrant color coding. The creative journey took a fascinating turn when Hausfather inquired how these visualizations could be transformed into a three-dimensional format.

The result was a captivating thermal helix animation, vividly illustrating temperature trends spiraling upward through time, resembling a tornado (see ‘A new view’). This novel representation is refreshing in a landscape saturated with the classic ‘hockey-stick’ graph of rising global temperatures, combining aesthetics with urgency. Despite his capabilities as a coder, Hausfather found himself unable to create this visualization without the help of AI.

As a researcher at Berkeley Earth, a climate data non-profit in California, Hausfather is part of a growing community using AI tools for innovative solutions. With the advent of large language models (LLMs), individuals can now directly query their computers to develop and implement code that caters to their varied needs, from graphics to data analysis.

This relaxed approach to coding has been coined “vibe coding,” a term introduced by Andrej Karpathy, co-founder of OpenAI, last year. This method emphasizes interaction with AI-driven tools to build projects collaboratively, allowing users to refine their requests until the outcomes align with their vision. In its most simplified form, vibe coding does not necessitate looking at the underlying code—focus remains on the end product. The boundaries of what constitutes vibe coding can be quite fluid, as experienced coders often oscillate between initiating a project independently and enlisting AI for assistance.

The publication Nature consulted a diverse group of scientists—ranging from seasoned coders to novices—who are exploring how AI can enhance their research capabilities. Many these researchers integrate AI into their workflows, with some taking the initiative to push the boundaries of its applications. While they agree on the impressive capabilities of existing AI coding tools, they also caution about their limitations and share cautionary tales about potential pitfalls.

All aboard

Vibe coding represents a significant milestone in the evolution of computer interaction. Going back to the 1960s, the use of punch cards for machine communication laid the groundwork for the development of coding languages like BASIC and Python, which made computer instructions more intuitive. Subsequent innovations in software tools have empowered non-coders to create with ease, exemplified by Microsoft’s Word allowing formatting changes without coding knowledge. The latest leap comes from LLMs, which offer unmatched speed and versatility in generating code, albeit with the occasional tendency to produce inaccuracies or mistakes.

While any LLM can be utilized for coding tasks, several systems specifically designed for coding have emerged recently, including GitHub Copilot, Anysphere’s Cursor, Anthropic’s Claude Code, Google’s Gemini Code Assist, and OpenAI’s Codex. These platforms can create functional applications from a single-sentence prompt, though they may still produce errors. For instance, Anthropic’s Claude Opus 4.7 is currently excelling in the Vibe Code Bench—a benchmark measuring AI-generated web applications—achieving an accuracy of just 71%.

Like many AI products, AI coding tools continue to evolve. The editions released in the past year often resemble helpful project managers, according to Hausfather. Users can provide extensive descriptions of their goals, prompting the AI to respond with coding plans, suggested verification tests, interface design options, and a comprehensive codebase complete with documentation. Hausfather notes that today’s AI-generated code often rivals the quality of his own: “as bug-free as my code.”

In the software development community, reliance on AI tools has reached significant levels. A recent survey conducted by DX, a Utah-based company focusing on developer productivity, indicates that over 90% of software developers engage with AI coding assistants at least once a month, with AI-generated code now making up more than a quarter of customer-facing code.

While the exact number of researchers adopting vibe coding remains unclear, interest is clearly on the rise. Argonne National Laboratory in Lemont, Illinois, recently hosted a one-day vibe-coding hackathon, attracting the maximum capacity of 200 participants.

Manuel Corpas, a genomicist and health-data scientist at the University of Westminster in London, describes a genuine enthusiasm for vibe coding among his peers. He managed to vibe-code a project named ClawBio in just two days, showcased during a hackathon at Imperial College London in early March. This project serves as a library of valuable bioinformatics code snippets, offering functionalities such as data extraction from scientific figures and personalized medication suggestions based on genomic sequences stored locally. AI agents can leverage this library to incorporate these functionalities into their tasks.

Following its launch, ClawBio saw an impressive 5,000 downloads within two weeks, and the community promptly contributed dozens of new skills, all of which were developed through vibe coding.

Good news first

Rosemarie Wilton, a molecular biologist at Argonne National Laboratory, entered the vibe coding realm without prior coding experience. Familiar with established software for analyzing data sets related to viruses discovered in wastewater, she attended the lab’s hackathon to explore how AI coding tools might assist her work.

AI ‘scientists’ joined these research teams: here’s what happened

Wilton found the experience enlightening. Without a graduate student to assist her, she was pleased to find that AI tools could independently perform tasks such as running data through multiple software programs, cross-checking results, or generating output graphs in specified formats. These AI systems were capable of operating autonomously for extended durations.

Previously, to set up new data-processing pipelines, she would initiate each step manually or consult the Data Science and Learning division at Argonne for coding help. Now, AI accelerates this exploratory phase, allowing her to refine methods more efficiently. If she identifies a successful approach, she plans to collaborate with the division for proper coding before submitting processed data to the health department.

Additionally, Wilton notes that the accessibility of vibe coding has made learning some programming less daunting. “I can learn a lot from it, even without much experience in Python,” she reflects. “It has opened up my world.”

Speed and agility stand out as key benefits of AI coding for everyone interviewed by Nature. Hausfather points out that even seemingly simple tasks like adjusting the log scale on a graph or layering additional information into a chart can be counterintuitive. The magic of being able to command his computer using plain language for such tasks is invaluable, he states. Vibe coding has also allowed him to design and launch websites within a day, a feat he had not previously achieved.

Tim Hobbs, a theoretical physicist at Argonne who participated in the hackathon, utilizes AI regularly, given the coding demands of his role. He investigates physics that exceeds the standard model of particle physics by analyzing large data sets from particle accelerators for theory testing or discovery. Vibe coding has enabled him to experiment with various mathematical approaches, identifying the most effective without significant effort.

“It’s like passing a problem to a highly skilled graduate student,” Hobbs remarks. “I can quickly explore different ideas and eliminate less promising ones.” He does underline, however, that he verifies any critical code, such as that for research publication.

Hobbs has been impressed with the clarity and quality of AI-generated code, noting its well-written annotations, often surpassing the standard of code produced by human authors in published papers. “Human-generated code tends to be messy since it reflects our imperfections,” he states.

In a paper¹ released earlier this year, Jesse Meyer, an analytical chemist at Cedars Sinai Medical Center in Los Angeles, sought to highlight the potential of AI. His team had developed software for biological data processing previously, but this time, he employed an LLM-powered app builder to vibe-code a pipeline for analyzing proteomics and similar ‘omics’ data.

Meet the academics refusing to use generative AI

Remarkably, it took him less than ten minutes, with just four well-structured prompts and a cost of under $2, to create a tool that might take traditional coders months, if not years, to develop. “The threshold for exploring new ideas has been significantly lowered,” Meyer observes. He envisions a future where researchers publish not only code but also their prompts, or ‘vibe blueprints’, to aid others.

However, he stresses that while this showcases AI’s capabilities, vibe coding for critical tasks requires extensive verification. In the introduction of his publication, he included a disclaimer stating: “Vibe coding does not replace a solid understanding of statistical analysis or computational logic.”