Anthropic Uses Claude to Redesign Quantum Computing Circuits

Researchers at Anthropic have demonstrated that artificial intelligence can autonomously design quantum computing circuits that improve upon published academic baselines, according to research details reported by Hacker News and posted to arXiv. The project was led by Andi Gu through Anthropic’s STEM Fellows Program, with mentorship from Anthropic quantum physicist Alek Kemeny in collaboration with Harvard University researchers Susanne Yelin and Mikhail Lukin, who is also affiliated with QuEra Computing.

Developing software for quantum computers presents constraints distinct from classical programming. Because quantum systems process information through qubits in superposition, circuits cannot execute conventional conditional branching, such as standard if-then logic, without maintaining all execution paths. In addition, physical principles dictate that quantum operations must be mathematically reversible. Rather than simply overwriting intermediate scratch data, quantum circuits must discard information by uncomputing it in reverse, requiring additional computational gates and execution time.

These physical limitations make quantum computational resources scarce. Existing quantum hardware contains hundreds to thousands of noise-sensitive qubits, far below the millions required for large-scale practical utility. For example, simulating the ground-state energy of FeMoco—the iron–sulfur cluster in nitrogenase used in biological nitrogen fixation—was estimated in 2017 to require roughly 50 trillion Toffoli gates. Although subsequent algorithmic improvements reduced that estimate to approximately 340 million gates on a theoretical machine with 3 million physical qubits, further structural circuit optimizations remain necessary to make practical applications feasible on near-term hardware.

To test whether language models could streamline quantum algorithms, Anthropic evaluated Claude Mythos 5 across eight computational primitives spanning chemistry, optimization, physics simulation, number theory, and numerical computing. The model received plain-language specifications of target calculations and interacted with an automated evaluator program that verified proposed circuit designs for correctness while calculating an objective score based on total qubits multiplied by total gates. Anthropic noted that the benchmarks were conducted internally in an automated closed loop without human intervention between iteration rounds.

Operating within this automated loop, Claude produced novel circuit designs that used fewer qubits and fewer gates than existing published algorithms across all eight primitives. According to Anthropic, the performance improvements derived from new structural and mathematical optimizations rather than incremental adjustments to existing designs. In one optimization primitive, the model reduced operational costs approximately 12-fold by eliminating costly components that were only needed to handle rare edge cases.

In a specific evaluation of point addition—a mathematical operation central to solving discrete logarithm problems on elliptic curves—Claude worked for 53 hours with minimal human input beyond its initial prompt. Point addition relies heavily on modular inversion, an arithmetic step frequently executed on classical systems through an iterative routine called divsteps. Straightforward quantum adaptations of divsteps previously proved impractical because storing operational branch history to preserve reversibility requires large amounts of memory qubits.

Claude addressed this bottleneck by proposing a theoretical method to infer branch information directly from the numbers remaining after each iteration, bypassing the need to store the operational history in dedicated qubits. The model proved a series of mathematical properties concerning the iteration process to ensure the approach functioned on full 256-bit numbers.

Anthropic noted that automated AI circuit design could help accelerate practical quantum applications. By reducing gate counts and qubit requirements for core computational subroutines, optimized algorithms could allow researchers to run complex tasks in chemistry, materials science, and cryptography on earlier generations of quantum hardware than previously expected.