Designing more efficient quantum circuits with Claude

Quantum computers require optimized circuits to perform useful tasks on limited hardware. Anthropic's Claude AI was provided with plain-language specifications for eight quantum circuits, covering areas such as chemistry, optimization, simulation, and number theory. Using automated evaluation, Claude independently designed circuits that outperformed existing best practices, in some cases by orders of magnitude.

This could reduce the resources needed for practical quantum applications, potentially allowing for earlier deployment of quantum computers. Quantum computers store information in qubits, which can exist in multiple states simultaneously. This allows them to tackle complex problems in chemistry, materials science, and number theory that are difficult for classical computers.

However, today's quantum computers only have hundreds to thousands of qubits, insufficient for many promising algorithms. Qubits are also sensitive to environmental noise and errors, which can lead to incorrect results. Researchers are working to redesign quantum algorithms to use fewer qubits and compute faster. FeMoco, a nitrogenase enzyme, serves as an example of a complex problem requiring significant quantum computational power.

In 2017, scientists estimated that simulating FeMoco's most stable state would need around 50 trillion Toffoli gates, a standard measure of computational cost. Through algorithmic improvements, this number was reduced to about 340 million, a 150,000-fold improvement, achievable on a machine with roughly three million qubits. Claude demonstrated its ability to optimize quantum circuits when given only a plain-language description of their purpose.

Quantum circuits, made up of sequences of gates, can be described in code, enabling Claude to generate, test, and refine designs as ordinary programs. The algorithms being optimized can be tested efficiently for correctness, allowing Claude to receive feedback on its proposals. Claude Mythos 5 worked autonomously, proposing circuit designs, receiving scores from an evaluation program, and improving its next attempt based on the feedback.

Across eight types of computational primitives—ranging from chemistry and optimization to simulation and number theory—Claude discovered novel designs that used fewer qubits and gates than the best previously published algorithms. These improvements resulted from uncovering new mathematical insights and structural optimizations, not just fine-tuning existing designs.

For instance, when designing a circuit to simulate an iron-sulfur molecule, Claude recognized patterns in the molecule's symmetries and repeating structure, enabling some calculations to be reused instead of recomputed. In other cases, Claude identified costly components in existing circuits that were only necessary for rare inputs, removing them to reduce computational cost.

These findings suggest that AI can play a crucial role in redesigning quantum circuits, a vital step toward making practical quantum computing a reality.