Anthropic said on Friday, Oct. 16, that its Claude AI, working largely on its own, came up with designs for eight basic quantum computing routines that improve on the best ones researchers had published, in some cases by ten times or more, though the accompanying paper shows the gain on one of them is a trade-off. The routines are building blocks for larger quantum programs in chemistry, optimization, physics simulation, number theory and numerical computing. Anthropic says savings like these cut the resources useful quantum computers will need, so practical machines could arrive sooner than expected.
Quantum computers store information in qubits, units that, unlike an ordinary computer's bits, can exist in a combination of states at once. The machines built so far have only hundreds to thousands of qubits, too few to run many of the programs that promise an advantage, and they work by running long chains of simple operations called gates. Quantum programs are laid out as circuits, fixed sequences of those gates, so researchers redesign the circuits to need fewer qubits and fewer gates. Anthropic points to the metal cluster at the heart of the enzyme bacteria use to turn nitrogen from the air into ammonia, which scientists want to simulate to design better industrial catalysts: the estimated number of gates needed fell about 150,000-fold through redesigns that took many research groups eight years.
The researchers gave Claude Mythos 5, one of Anthropic's models, a plain-language description of what each circuit had to compute, plus a separate program that checked every design and scored it by the qubits and gates it used. Claude proposed a design, got its score back and tried again, with no human input between rounds. In seven of the eight tasks, its circuits used both fewer qubits and fewer of one of the costliest kinds of gate than the best earlier designs, the paper says; for one of the seven no published design existed, so the comparison is with a straightforward version. For the most expensive step of a quantum method for optimization problems, the kind behind scheduling and logistics, Anthropic says Claude's design needs one-twelfth as many gates as an earlier design by Google researchers.
The eighth task was point addition, a rule for combining two points on a mathematical curve that sits at the core of quantum methods for a math problem that, for large curves, ordinary computers can't solve in any reasonable time, and a standard test in the field. Anthropic says Claude worked on it for 53 hours with only the opening brief and occasional prompts to think more broadly, and found a new approach. One design from it uses 12% more qubits than the earlier design with the fewest qubits but 17 times fewer gates. Another published design still needs fewer gates than any of Claude's, though it uses far more qubits, so here Claude offers new trade-offs rather than a clear win, the paper's figures show. Anthropic says Claude worked out the math behind its approach and had a computer check the proof.
The results come with limits. The designs were checked by a program on ordinary computers, not run on a quantum machine. Some leave out parts needed only for very rare inputs, so they occasionally give a wrong answer: the optimization design fails about one run in 10,000, which Anthropic says is well within the error rate the task allows. The researchers say that in early experiments Claude was cautious, tended to prefer well-known approaches, and had to be pushed to try bolder ones. The work was led by Andi Gu, a fellow in an Anthropic program, with the Anthropic physicist Alek Kemeny and the Harvard physicists Susanne Yelin and Mikhail Lukin, who also works with the quantum computer maker QuEra Computing.
It is not yet clear how much of the saving will hold up inside a working machine. Circuit design is only one part of building a useful quantum computer, Anthropic says, since the machine must also catch and fix its own errors as it runs, and the paper says the right way to score a design depends on the hardware and how it corrects errors.