AI from Anthropic and Google designed quantum computer programs that need far fewer steps and qubits, two research teams report

Artificial intelligence is starting to help design programs for quantum computers, and two research groups say the AI found better designs than human experts had, according to research papers posted online this week and in October.

A quantum computer stores information in “qubits,” units that can hold a mix of 0 and 1 at once.  For certain jobs, such as simulating molecules to design drugs or materials, that could let it solve problems no ordinary computer can.  Today’s machines are small and make many errors.  Experts expect the first truly useful ones will need thousands of reliable qubits running billions of steps, so every step saved brings that day closer, the Google paper says.

What the AI found

• Anthropic’s Claude, working with Harvard University and QuEra Computing, a Boston company that builds quantum computers, redesigned eight common building blocks of quantum programs.  In seven of them, the AI’s designs needed fewer qubits and fewer of the costliest kind of step, the researchers reported.  One piece used to simulate molecules went from 6,355 of those costly steps and 252 qubits to 1,140 steps and 142 qubits, QuEra said.
• A Google DeepMind and Google Quantum AI team used teams of Gemini 4 AI agents, programs that work through tasks step by step, which shared notes in a wiki and held scheduled “conferences” with each other.  Their designs cut estimated costs by about 11 times for a fusion-energy simulation, 8 times for a chemistry simulation and 2.5 times for a calculation about how the body breaks down medicines, the paper says.

Both teams checked the AI’s work automatically.  In the Google project, every design had to pass a mathematical proof checker, so a design that only looked right was rejected.  QuEra said the Anthropic project used a separate program that tested each design and scored it.

Limits

These are designs and cost estimates, not programs run on working machines.  QuEra said a 39-fold drop on its combined score “is not a 39x speedup.”  The Google team wrote that its large simulations still need far more than the first error-corrected machines are expected to offer.  Neither paper has gone through peer review, in which outside scientists check the work before a journal publishes it.

Claude also found a new way to do a calculation used in elliptic-curve math, which underlies much of today’s online encryption, according to QuEra.  QuEra plans to offer a machine called Libra on Amazon’s cloud in 2028, the company said.

Sources:  QuEra announcement; Gu et al., preprint on arXiv; Anthropic research post; Google DeepMind and Google Quantum AI preprint on arXiv.