Insider Brief
- Anthropic and Harvard researchers paired Claude with an automated evaluator to design more efficient quantum circuits across eight computational tasks.
- The system reduced gate and qubit requirements in seven tasks and developed a new arithmetic approach in the eighth, with results checked through classical computation rather than quantum hardware.
- Researchers guided Claude toward broader exploration, while practical benefits remain dependent on hardware requirements, error correction and assumptions about acceptable failure rates.
Anthropic and Harvard researchers used Claude to design more efficient quantum circuits, showing how AI models paired with automated testing can develop algorithms under demanding technical constraints.
Across eight tasks, the system produced seven designs that reduced both costly quantum operations and qubit requirements compared with the study’s baselines. The eighth yielded a new arithmetic approach offering better trade-offs between computing time and memory.
The findings concern circuit designs checked through classical computation, rather than experiments on quantum hardware. For an AI audience, the broader result is a demonstration of a research workflow in which a language model proposes solutions, receives measurable feedback and revises its approach.
The study involved researchers affiliated with Harvard University, Anthropic and QuEra Computing. Anthropic’s accompanying blog post identifies the model as Claude Mythos 5.
A Search Guided by Tests
Each search began with a written specification and a blank program. Claude proposed several distinct algorithmic approaches, developed them in parallel and shared promising ideas between them.
An independent evaluator checked proposed circuits and scored them using the number of Toffoli gates multiplied by peak qubit use. Toffoli gates are computational operations that are particularly costly in many proposed error-corrected quantum computers.
The researchers selected circuits whose behavior could be tested efficiently on conventional computers. This provided feedback without requiring a quantum processor.
The challenge extended beyond ordinary code optimization. These quantum routines must preserve information needed to reverse their calculations. Temporary results cannot simply be overwritten; clearing them requires additional operations.
Claude repeatedly found ways to reconstruct earlier information from the calculation’s current state, reducing storage requirements. It also reorganized operations so different computational paths could share arithmetic.
One redesigned decoding routine used about 440,000 Toffoli gates, compared with 5.1 million in the baseline, while reducing peak qubit use from 1,938 to 1,837, according to the paper.
Another result involved modular inversion, an arithmetic operation used in elliptic-curve calculations. Claude developed a reversible version of an existing classical method by finding how to recover earlier computational decisions without storing their full history.
According to Anthropic’s post, the elliptic-curve work took 53 hours, with human involvement limited to initial instructions and occasional prompts to broaden the search.
Human Guidance Still Mattered
Claude’s caution sometimes limited its exploration. According to the post, it favored familiar approaches and repeatedly checked improvements that were already well supported.
The researchers countered this by requiring several distinct approaches and occasionally encouraging a broader search. Once Claude identified the new arithmetic approach, it developed, implemented and verified it autonomously, Anthropic said.
The results also have technical limits. Some savings depend on accepting rare failures under specified input distributions. One comparison used a straightforward reversible implementation because no previous reversible circuit existed in the literature.
Gate and qubit counts also approximate practical costs, which depend on hardware and error correction.
The study suggests AI-guided discovery works particularly well when researchers can define a clear task, test candidates cheaply and measure improvement. Extending that approach to broader scientific problems will depend on building equally reliable evaluators.
