Anthropic’s Claude Designs More Efficient Quantum Circuits With Automated Feedback

Insider Brief

• Researchers from Anthropic and Harvard combined Claude with an automated testing system to design more efficient quantum circuits for eight computing tasks.
• The system reduced the number of quantum operations and qubits needed in seven tasks. For the eighth, it developed a new way to do arithmetic. The researchers checked the results using ordinary computers, not quantum hardware.
• The researchers encouraged Claude to explore more possibilities. However, the real-world benefits still depend on hardware needs, error correction and assumptions about how often a system can fail.

Researchers from Anthropic and Harvard used Claude to design more efficient quantum circuits. Their work shows how AI models can develop algorithms when they work with automated tests and receive clear feedback about how well their solutions perform.

Across eight tasks, the system created seven designs that needed fewer costly quantum operations and fewer qubits than the starting designs used in the study. The eighth task produced a new way to do arithmetic that offered a better balance between computing time and memory use.

The researchers tested the circuit designs using ordinary computers rather than running experiments on quantum hardware. The main finding is that researchers can use a language model to suggest solutions, measure how well they work and improve them based on clear test results.

The study involved researchers affiliated with Harvard University, Anthropic and QuEra Computing. In its accompanying blog post, Anthropic identifies the model as Claude Mythos 5.

A Search Guided by Tests

Each search started with written instructions describing the problem and an empty program. Claude suggested several different ways to solve it, worked on them at the same time and shared promising ideas between the different approaches.

A separate automated system checked the proposed circuits and gave them scores based on the number of Toffoli gates multiplied by the highest number of qubits used at once. Toffoli gates are computing operations that can be especially costly in many proposed quantum computers that use error correction.

The researchers chose circuits whose behavior could be tested efficiently on ordinary computers. This let the system receive feedback without needing a quantum processor.

The challenge was more complicated than simply making ordinary computer code run faster. Quantum routines must keep the information needed to undo their calculations. They cannot simply erase temporary results by writing new information over them. Removing those temporary results requires extra operations.

Claude repeatedly found ways to work out earlier information from the current state of a calculation, reducing the amount of storage needed. It also rearranged operations so that different parts of a calculation could share the same arithmetic steps.

According to the paper, one redesigned routine for decoding information used about 440,000 Toffoli gates, down from 5.1 million in the original design. It also reduced the highest number of qubits needed at once from 1,938 to 1,837.

Another result involved modular inversion, a type of arithmetic used in elliptic-curve calculations. Claude created a reversible version of an existing method designed for ordinary computers. It did this by finding ways to recover earlier decisions in a calculation without saving the full record of every step.

According to Anthropic's post, the work on elliptic curves took 53 hours. Human involvement was limited to giving the initial instructions and occasionally prompting Claude to explore more possibilities.

Human Guidance Still Mattered

Sometimes, Claude was too cautious, which limited how much it explored. According to Anthropic's post, it often preferred familiar methods and repeatedly checked improvements that already had strong support.

The researchers addressed this by requiring Claude to try several different approaches and sometimes encouraging it to search more widely. Anthropic said that after Claude found the new arithmetic method, it developed, implemented and checked the method on its own.

The results also have technical limits. Some of the savings depend on accepting a small chance of failure for certain groups of inputs. In one comparison, the researchers used a straightforward reversible design because earlier research had not described an existing reversible circuit for that task.

The numbers of gates and qubits are also estimates of practical costs. The actual costs depend on the hardware used and how error correction is carried out.

The study suggests that AI-guided discovery works especially well when researchers can clearly define a problem, test possible solutions cheaply and measure how much each solution improves. Applying the same approach to a wider range of scientific problems will depend on creating testing systems that are just as reliable.