Anthropic’s Claude has designed quantum circuits improving on the best published designs by an order of magnitude or more in some cases, substantially reducing the quantum resources needed for applications. Andi Gu, leading the work as part of Anthropic’s STEM Fellows Program, collaborated with researchers from Harvard University and QuEra Computing, including Susanne Yelin and Mikhail Lukin, to achieve these circuit design improvements.
Current quantum computers possess only hundreds to thousands of qubits, and this research aims to make those qubits more efficient. The team reports that Claude was given plain-language specifications for eight circuits and worked largely on its own to achieve these results.
Claude Optimizes Quantum Circuits for Chemistry and Beyond
Claude discovered new approaches to quantum circuit design that reduce the number of qubits and gates required for computations, a feat accomplished through novel structural and mathematical insights rather than incremental improvements to existing designs. This advance addresses a fundamental challenge in quantum computing: the trade-off between qubit usage and computational time, a constraint that significantly impacts the feasibility of complex algorithms. The system successfully optimized circuits for a range of applications including chemistry, optimization, physics simulation and number theory, demonstrating a versatility beyond simple algorithmic fine-tuning.
The AI’s success stems from its ability to reason under constraints unlike those found in conventional programming, specifically the need to balance qubit allocation with the avoidance of redundant calculations. Claude identified that existing circuits often included costly components designed to handle rare inputs, and subsequently redesigned them to eliminate these unnecessary elements.
For instance, when tasked with simulating an iron-sulfide molecule, Claude used the molecule’s inherent symmetries and repeating structures to reuse previously calculated results, minimizing redundant computation. This approach, reminiscent of a technique called “divsteps” in classical computing, allowed for a more efficient allocation of quantum resources. Claude’s capabilities extend to optimizing specific computational primitives.
The team reports that another discovery made the costliest step of a leading quantum optimization algorithm approximately twelve times cheaper. These improvements are not merely reductions in gate count. They represent a fundamental shift in how these circuits are structured, demonstrating an ability to uncover previously unseen efficiencies.
The implications of this work extend beyond mere algorithmic optimization; it suggests that AI can actively contribute to the redesign of circuits, a key step in realizing the full potential of quantum computing. Faster circuit design could accelerate progress in fields reliant on quantum computation, including drug discovery, materials science and fertilizer production. The ability to efficiently manage qubits and gates is paramount, as both contribute to the overall cost and complexity of quantum computations.
“Both of these constraints make quantum computations far more expensive, in both runtime and qubits, than their classical counterparts,” the team notes, highlighting the significance of Claude’s contributions. The work represents a move toward automated quantum circuit design, potentially opening new avenues for exploration and innovation in the field.
Claude Mythos 5 Achieves Closed-Loop Quantum Circuit Optimization
Claude Mythos 5 achieved closed-loop quantum circuit optimization by autonomously proposing, testing, and refining designs without human intervention, a feat demonstrated across eight distinct computational primitives. The system’s success hinges on an automated evaluator program that objectively scores each circuit based on qubit and gate count, serving as a proxy for runtime and enabling iterative improvement. This approach allowed the AI to move beyond simply fine-tuning existing designs and instead uncover novel structural and mathematical insights within the problems themselves, leading to substantial efficiency gains.
Through a series of algorithmic advances, the estimated number of gates for simulating FeMoco’s ground-state energy has come down to about 340 million, a roughly 150,000-fold reduction, corresponding to a few hours of computation on a future machine with about three million physical qubits.
These reductions largely came from redesigning the circuits that carry out the computation, rather than from making better hardware. The system’s ability to identify and find underlying mathematical structures suggests a capacity for uncovering previously unknown relationships within complex quantum systems. This could accelerate the development of more efficient algorithms and create opportunities for quantum computation.
Novel Circuit Designs from Claude Reveal Structural Insights
Claude Mythos 5 demonstrated an ability to independently refine quantum circuit designs, achieving improvements over published algorithms through the identification of underlying mathematical structures. Across eight distinct computational primitives spanning chemistry, optimization and number theory, the system generated circuits requiring fewer qubits and gates than previously known implementations. This success wasn’t achieved through incremental adjustments to existing designs, but through the discovery of novel approaches rooted in the inherent properties of each problem.
The system’s closed-loop operation, proposing designs, receiving evaluations and autonomously iterating, was critical to this process. This objective function, combined with an external evaluator program capable of verifying correctness, allowed Claude to function with minimal human intervention.
Through a series of algorithmic advances, the estimated number of gates for simulating FeMoco’s ground-state energy has come down to about 340 million, a roughly 150,000-fold reduction, corresponding to a few hours of computation on a future machine with about three million physical qubits. Claude has proven particularly adept at optimizing classical algorithms, so we wanted to see whether it could do the same for quantum algorithms if it were given only a plain-language description of what each circuit must compute.
The system identified opportunities to reconstruct temporary values from the circuit’s current state, eliminating the need for costly storage and further streamlining the decoding process. Claude’s approach to temporary value management exemplifies its ability to find structural insights. Ordinarily, circuits must store temporary values to reconstruct them later, a process that demands significant qubit resources. Claude, however, devised a method to derive most of this information from the circuit’s existing state, effectively bypassing the need for storage.
This technique mirrors the strategy employed in optimizing divsteps, another computational primitive, where the system identified opportunities to reduce qubit usage through clever data reuse. While Claude exhibited strong performance, researchers noted a tendency toward conservatism in its initial design choices. This ability could prove valuable in other areas of scientific inquiry, from materials science to fertilizer production, where complex systems often exhibit similar repeating patterns.
Reversible Algorithm Constraints Guide Claude’s Quantum Design Process
Claude’s approach to quantum circuit design consistently prioritized minimizing qubit requirements, even when it meant a slight increase in the number of operations needed for computation. This preference emerged as the AI tackled challenges in areas like elliptic-curve cryptography and quantum optimization, revealing a tendency to favor storage efficiency over raw speed in its initial designs. This focus on resource optimization is particularly important given the current limitations of quantum hardware, where the number of available qubits remains in the hundreds to thousands.
The AI’s success in optimizing stems from its ability to navigate the unique constraints inherent in quantum computation, constraints absent in traditional programming. Unlike conventional code, quantum computers cannot employ conditional logic, the “if-then” branching, because a quantum computer operates on a superposition of all possible inputs simultaneously. This necessitates that any operation applicable to one input must also be included for all others, significantly increasing circuit complexity.
This technique echoes optimization strategies used in classical computation, specifically in the refinement of divsteps, a computational primitive where minimizing intermediate storage is also paramount. A prime example of this optimization is Claude’s redesign of a quantum circuit for Reed-Solomon decoding, a critical step in error correction for quantum computers. Claude achieved this efficiency by switching from the extended Euclidean algorithm to the Berlekamp-Massey algorithm, a method that simplifies the computational steps and aligns better with the fixed-sequence gate requirements of quantum circuits.
The shift to Berlekamp-Massey, however, introduced a new challenge related to the reversibility constraint of quantum computations. Quantum computations must be reversible, meaning that the input can be fully reconstructed from the output, and this typically requires storing intermediate results. This demonstrates a capacity to not only satisfy the algorithmic requirements of a task but also to intelligently manage the physical limitations of quantum hardware.
This ability to accelerate the iterative process of quantum circuit design is particularly noteworthy. The team’s findings also suggest that Claude’s designs often expand the possibilities of quantum circuit performance, offering combinations of qubit count and runtime that were previously unattainable. While Claude exhibited a conservative approach in its initial designs, favoring established methods, the researchers believe that further refinement could unlock even greater potential. The team is now proposing evaluating Claude’s performance on a wider range of circuits and exploring methods to encourage more adventurous design choices.
See today’s quantum computing news on Quantum Zeitgeist for the latest breakthroughs in qubits, hardware, algorithms, and industry deals.
