Now loading...
Quantum computers require precisely engineered circuits to execute effective algorithms on constrained hardware. Researchers provided Claude with clear specifications for eight circuits, which pertain to fields like chemistry, optimization, physics simulation, and number theory. Alongside an automated evaluation tool that checks and scores each design, Claude operated independently and discovered designs that surpass the best currently available circuits in some cases by as much as tenfold. These advancements can significantly lower the quantum resources necessary for practical applications, suggesting that functional quantum computers may arrive sooner than anticipated.
This initiative was spearheaded by Andi Gu, part of Anthropic’s STEM Fellows Program, with guidance from Alek Kemeny, a quantum physicist at Anthropic, and collaboration with Susanne Yelin and Mikhail Lukin from Harvard University and QuEra Computing. The full research paper can be accessed here.
In contrast to classical computers, which manage data in bits with values of 0 or 1, quantum computers use qubits, which can be realized through single atoms, superconducting circuits, or light particles. Qubits leverage quantum phenomena like superposition, allowing quantum systems to exist in multiple states at once, facilitating the solving of complex issues in chemistry, materials science, and number theory that would be intractable for traditional computers.
As it stands, existing quantum computers possess a few hundred to several thousand qubits—insufficient for executing many algorithms that could yield practical benefits over classical systems. Moreover, these qubits are extremely sensitive to environmental noise such as temperature changes, vibrations, and cosmic rays. A slight disruption can alter a qubit’s state, leading to accumulated errors that result in incorrect computations.
Due to the limited availability of both qubits and computing time, researchers are focused on redesigning quantum algorithms to use fewer resources while expediting computations. One significant example is the simulation of FeMoco, an iron-sulfur cluster integral to the nitrogenase enzyme that bacteria use to convert atmospheric nitrogen into ammonia. Understanding this process could aid scientists in developing improved industrial catalysts.
In 2017, estimates indicated that about 50 trillion Toffoli gates would be needed to compute the energy of FeMoco’s most stable configuration. Meanwhile, recent advancements have reduced this requirement to approximately 340 million gates, a decrease more than 150,000 times, allowing for a computation period of just a few hours on a future machine equipped with around three million physical qubits.
Most of these reductions stemmed from reconfiguring the circuits responsible for the computations, rather than enhancements in hardware. Claude demonstrated a particular aptitude for optimizing classical algorithms, prompting an investigation into whether it could similarly refine quantum algorithms based solely on plain-language descriptions of circuit requirements.
Claude is uniquely well-equipped for this role for several reasons. Quantum algorithms are expressed through quantum circuits, which consist of sequences of gates, such as the previously mentioned Toffoli gates. These circuits can be coded, letting Claude use its experience with programming to write, execute, and modify circuit designs like standard programs. Additionally, the type of quantum algorithms targeted for optimization can be efficiently tested for accuracy, producing definitive outputs that can be verified by a classical computer, allowing an external evaluation program to validate Claude’s suggestions. Furthermore, Claude can receive a clear quantitative score to enhance: the product of the used qubits and the number of gates, which serves as a proxy for estimating how long a circuit will take to operate.
Results showed that Claude Mythos 5 could effectively operate within a closed-loop system, suggesting circuit designs, receiving feedback, scoring from the evaluation program, and using that feedback for improvements without any human intervention between iterations. Across eight computational primitives for chemistry, optimization, quantum simulation, number theory, and numerical computing, Claude identified innovative designs that employed fewer qubits and fewer gates compared to previously published algorithms.
Improvements stemmed from revealing new structural and mathematical insights into the problems rather than modifying existing designs. For instance, while creating a circuit to simulate an iron-sulfide molecule, Claude recognized that the molecule’s symmetries and repetitive structure allowed certain results to be calculated once and applied multiple times rather than recalculated for every section of the system. In some instances, it discovered that existing circuits contained costly components necessary only for extremely rare inputs, allowing Claude to redesign them.
Among Claude’s findings, one novel method enables modular inversion, which is the most resource-intensive step in elliptic-curve calculations. Another breakthrough made the most expensive step in a leading quantum optimization algorithm approximately twelve times cheaper.
These results imply that artificial intelligence can now assist in the redesign of quantum circuits, a critical step in accelerating the execution of quantum algorithms. Enhancing circuit design could bring quantum computing solutions, from new medicine developments to cleaner fertilizer production methods, within reach more rapidly.
