AI Tools Help Researchers Explore “Vibrating Code” in the Quantum Realm

“Vibecoding” is expanding into the field of quantum computing, a discipline traditionally known for its demanding programming requirements. To make quantum programming more accessible, researchers at Pasqal, a quantum-computing start-up based in Paris, have developed an artificial intelligence agent that converts English-language prompts into quantum code and can run that code on a quantum computer.

The agent is described in a preprint posted last month on the arXiv server.¹ Although AI-generated quantum programs often require feedback from experts to work correctly, the developers say progress is accelerating. Their work could eventually make quantum computers — machines that use quantum phenomena to speed up certain calculations — accessible to a much broader research community.

Christophe Jurczak, a co-author of the study and a co-founder of Pasqal, said the AI agent enabled him to conduct experiments that would normally require a team of physicists with specialist knowledge of quantum computing. “And you can do it yourself from your couch in Dallas, Texas.”

Bringing AI vibecoding to quantum computing

In conventional software development, “vibecoding” refers to an extreme form of AI-assisted programming. A user describes what they want software to do, and an AI tool generates and runs a functional program, often refining it in response to the user’s feedback.

Advanced large language models (LLMs), including Anthropic’s Claude, have demonstrated an understanding of quantum-computing concepts. Many researchers now use these systems to generate both conventional software and quantum code. To determine whether an LLM could support quantum vibecoding, Jurczak and his colleagues trained a frontier LLM to understand the technical specifications of Pasqal’s quantum hardware.

The researchers focused on quantum simulations, one of the most promising applications of quantum computers. These simulations can model the behaviour of physical systems, including chemical catalysts and materials with unusual magnetic properties. In the future, quantum computers could use such simulations to predict material properties that are too complex for classical computers to calculate efficiently.

For their study, the Pasqal team evaluated the AI agent using three tests. In each case, the researchers selected a physics paper describing a phenomenon that could, in principle, be simulated on a quantum computer. They then asked the agent to generate and run quantum code that could reproduce the reported results.

Two tests involved materials whose atoms can point up or down, behaving like tiny bar magnets according to the orientation of neighbouring atoms. To simulate these systems, the AI agent first had to translate the physical model into operations compatible with a Pasqal quantum computer, which stores information in arrays of atoms trapped by laser light.

This crucial “translation” step usually requires researchers with expertise in both the physics of the material being modelled and quantum programming, said physicist Loïc Henriette, Pasqal’s chief technology officer and a co-author of the study.

Before sending the quantum code to real hardware, the agent tested it on a “virtual” quantum computer running on a classical computer. When the code passed the simulation, it was automatically submitted to one of Pasqal’s quantum computers in Dhahran, Saudi Arabia, or to another system in Sherbrooke, Canada.

Across all three tests, the authors reported that the agent showed “a strong grasp of hardware constraints”. In one test, the researchers deliberately asked it to run a simulation that they knew was too complex for the available Pasqal machine in the cloud. The agent correctly explained why the task could not be completed. In other tests, however, the researchers said the AI system needed substantial human input to produce a “physically accurate implementation”.