The Signal
Google Opens $150,000 Contest to Shrink Quantum Circuits
Google Quantum AI has launched a $150,000 Kaggle competition to reduce the qubit and gate costs of circuits used in fusion simulation, optimization and quantum chemistry. Its verifier checks candidate circuits with randomized classical test
Quantum.world Team · Written with AI · fact-checked ·
Google Quantum AI is putting $150,000 behind a contest to make fault-tolerant quantum circuits smaller, but its verifier does not prove that a submitted circuit exactly matches its target. Instead, the competition uses randomized classical tests to estimate whether a circuit’s errors stay within an allowed bound. That trade-off—faster checks in exchange for a statistical rather than exact guarantee—is central to the challenge announced on October 21.
Entrants will optimize three quantum-circuit bottlenecks drawn from research in fusion simulation, combinatorial optimization and chemistry. Google says a new open-source library, kickmix, lets AI agents build and test large circuits at speeds of up to 60 billion gates per second on one CPU core. The company’s Kaggle competition asks participants to submit circuits that pass the supplied verifier while reducing a score based on qubit and gate counts.
Fast tests, bounded confidence
Ordinary classical simulation of an arbitrary quantum circuit becomes intractable as gates create superpositions across many bitstrings. Google’s approach targets a structured family called kickmix circuits, built from reversible classical logic, diagonal phase gates and measurement-based uncomputation. Those circuits can be tested on classical input bitstrings without simulating a growing superposition.
The test checks whether the circuit produces the expected output, clears its scratch qubits and applies the intended phase. The authors write, “Randomized fuzz testing does not prove exact equivalence, but most quantum algorithms can tolerate a small, bounded error.” Tanuj Khattar and Ryan Babbush, Google Quantum AI
The paper describes a verifier that derives challenge inputs from a hash of the submitted circuit, combined with a private salt during the competition. That design is intended to prevent participants from tailoring a circuit to a fixed, known test set. Passing the tests supports a high-confidence estimate of the circuit’s failure rate under the verifier’s assumptions; it is not a proof that the circuit is correct for every possible input.
Three bottlenecks, from fusion to FeMoCo
The first benchmark concerns calculating a phase proportional to 1/√x, a subroutine used in Coulomb-potential calculations relevant to inertial-confinement-fusion simulations. The second asks entrants to build a quantum circuit for Reed–Solomon decoding, an arithmetic bottleneck in decoded quantum interferometry applied to an optimization problem.
The chemistry benchmark targets circuits for simulating the iron-molybdenum cluster FeMoCo, the active site of the nitrogenase enzyme. Entrants must optimize a factorization of the molecular Hamiltonian and the quantum lookup-table circuits that load its parameters. Google presents each task as a costly subroutine within a larger algorithm, not as a complete simulation or a test of quantum hardware.
Google keeps its best runs under wraps
Google says its internal runs with Gemini models improve on previously published baseline circuits in all three benchmarks, but it does not publish the size of those gains in its announcement. The accompanying paper records hashes of the internal circuits and says the full solutions will be released after the competition. That gives the team a way to commit to its results now without revealing the circuits participants are being asked to beat.
The contest therefore tests two claims at once: whether automated searches can reduce the resource costs of useful quantum subroutines, and whether classical randomized checks can make that search practical at scale. The verifier’s acceptance criteria are concrete: expected input and output bitstrings, clean scratch qubits and a phase within the benchmark’s stated tolerance.