Claude designs leaner fault-tolerant quantum circuits with Harvard and QuEra

[Callum_33]

Researchers from Anthropic, Harvard and QuEra have used Claude to design more efficient fault tolerant quantum circuits. The Quantum Insider's Matt Swayne reports that the AI worked on eight building blocks used in chemistry, optimisation, physics simulation and number theory. In seven of the eight tasks, its designs used fewer gates and fewer qubits than the best published versions

The setup was interesting. An automated evaluator scored each design by multiplying the number of Toffoli gates by the peak number of qubits. Claude then kept revising its designs with very little human input, gradually improving the score. That kind of tireless iteration is where AI can help

Some of the improvements are large. A decoder used in decoded quantum interferometry fell from 5.1 million Toffoli gates to about 440,000, roughly twelve times fewer. It reserves only 19 qubits for stored values, compared with a worst case of 665, and accepts failure in about one run in 10,000. A Gaussian function circuit improved 39 times and a linear equation routine 32 times

One task, a reversible modular inversion method, was worked on for 53 hours. The final design uses 12 percent more qubits but 17 times fewer gates than the smallest previous design. The work is led by Andi Gu with Alek Kemeny, Susanne Yelin and Mikhail Lukin

There are important caveats. These are resource estimates rather than runs on real hardware, and the preprint has not been peer reviewed yet. Cutting the cost of future quantum algorithms brings useful machines closer, and it is a fascinating example of AI helping science

Could AI become a standard tool for designing quantum circuits? And how much should we trust designs that have not yet been tested on hardware?

Anthropic's Claude Designs Leaner Quantum Circuits, Reducing Resources Needed for Key Calculationsthequantuminsider.com

[Wandering Matt]

Seven out of eight tasks improved is a strong hit rate. The one that did not improve would be interesting to understand. Failures teach a lot too

[MJF02]

An automated scorer plus an AI that keeps iterating is a powerful combination. Humans get tired of small tweaks. Machines do not. Automated scoring also makes the results reproducible. Others can test the same approach

[Caitlin_75]

Accepting failure in one run in 10,000 is a sensible trade. Perfect circuits are often wasteful

Repeat the rare failures and save resources overall. Fault tolerant designs often over engineer for rare cases

Accepting small failure rates can save huge resources. It is a smart engineering choice

[Dank]

Mikhail Lukin's involvement gives this real credibility. His group knows neutral atom hardware inside out. They will want designs that work in practice

[CosmicRay91]

Resource estimates are useful, but they are not hardware results. Peer review will check the designs carefully

I would wait before celebrating too much. Peer reviewers will want to check the circuits carefully

Small errors in resource counts can matter a lot. Independent verification is important