Can AI Write Quantum Programs, and How Well Does it Perform?

Research demonstrates that OpenAI’s ChatGPT (GPT-4) can generate quantum circuits, although performance varies depending on the hardware. The model exhibits greater proficiency in composing circuits for IBM’s quantum machines compared to those utilising Xanadu’s photonic devices, suggesting differing levels of accessibility for quantum programming via large language models.

The increasing availability of quantum computing resources via cloud platforms presents a challenge: how to bridge the gap between the complex skill set required to program these machines and the potential user base. Researchers are now investigating whether large language models, which have already demonstrated utility in simplifying complex tasks in fields such as software development, can offer a solution for quantum programming. A team from the Darwin Deason Institute for Cyber Security at Southern Methodist University, comprising Elena R. Henderson, Jessie M. Henderson, Joshua Ange, and Mitchell A. Thornton, addresses this question in their work, “Programming Quantum Computers with Large Language Models”. They examine the capacity of OpenAI’s ChatGPT (GPT-4) to generate quantum circuits suitable for execution on both IBM’s superconducting transmon qubits and Xanadu’s photonic quantum computers, finding a notable performance difference favouring the former platform.

Large language models (LLMs) are increasingly investigated for their potential as assistive tools within quantum computing, and recent research assesses the capacity of OpenAI’s ChatGPT (GPT-4) to generate functional quantum circuits. This study addresses a gap in current investigation, moving beyond explorations of LLMs in classical software development to evaluate their utility within the quantum realm. The research focuses on practical application, evaluating the LLM’s ability to produce functional quantum programs for two distinct hardware platforms: IBM’s superconducting systems and Xanadu’s photonic devices.

The results reveal a significant performance disparity between the platforms, with ChatGPT demonstrating substantially greater proficiency in generating circuits compatible with IBM’s hardware. This suggests that the current architecture and established programming paradigms of IBM’s systems, coupled with the prevalence of the Qiskit software development kit, provide a more readily accessible framework for the LLM. Qiskit is an open-source software development kit for working with quantum computers, providing tools for creating, compiling, and running quantum circuits.

Conversely, ChatGPT exhibits limitations when applied to Xanadu’s photonic quantum computing platform, Strawberry Fields. This stems from the fundamentally different approach to quantum computation employed. Unlike qubit-based systems, which utilise discrete units of quantum information, Xanadu utilises continuous variable quantum computing (CVQC). CVQC encodes information in continuous degrees of freedom, such as the amplitude and phase of light, requiring a different set of algorithms and programming techniques, presenting a greater challenge for the LLM.

The research highlights the importance of squeezed states of light, a quantum state essential for reducing noise in CVQC systems. It demonstrates how effectively ChatGPT can incorporate these concepts into its generated code. Squeezed states reduce uncertainty in one variable at the expense of increased uncertainty in another, improving the signal-to-noise ratio in quantum computations. This highlights the need for LLMs to possess a deep understanding of the underlying physics and engineering principles of different quantum computing platforms. The observed performance gap underscores the importance of tailoring LLM training and prompting strategies to the specific characteristics of each quantum computing architecture.

Future research should focus on developing LLM training datasets that specifically address the unique challenges of CVQC, potentially involving the creation of synthetic datasets of quantum circuits designed for CVQC architectures. Additionally, researchers should explore techniques for incorporating domain-specific knowledge about CVQC into the LLM’s training process, such as using knowledge graphs or other structured representations of quantum information. Developing more sophisticated prompting strategies could also improve the LLM’s performance on CVQC tasks.

The findings indicate that while LLMs hold promise as educational tools and potential aids in quantum algorithm development, their effectiveness is heavily influenced by the underlying quantum computing platform and the availability of well-structured software resources, suggesting that a one-size-fits-all approach to LLM integration in quantum computing is unlikely to succeed. The study demonstrates that LLMs can successfully generate basic quantum circuits, but their capacity to handle complex algorithms or platform-specific nuances remains limited, highlighting the need for further research to enhance the capabilities of LLMs in quantum programming. This work provides a valuable baseline for future research exploring the integration of LLMs into the quantum computing ecosystem and emphasizes the importance of developing accessible and well-documented software tools to maximise their potential.

Researchers should also investigate the potential of using LLMs to automate the process of quantum circuit optimisation, potentially identifying and eliminating redundant quantum gates or simplifying complex circuits. LLMs could also generate test cases for quantum circuits, helping to ensure they are functioning correctly, or assist in the development of new quantum algorithms by generating candidate algorithms based on a given set of specifications.