A quantum computer is like an orchestra: each musician plays their part, but the slightest false note ruins the symphony. The physical qubits it is built from are noisy and error-prone. To protect computations, they are combined into logical qubits using quantum error correction. The traditional approach, the surface code, is reliable but wasteful: hundreds of physical qubits are needed for a single logical qubit. Quantum LDPC codes promise more, but they are hard to implement due to complex connectivity. New work offers a way out: the bicycle architecture, where modules are connected by long-range links, like conductor's batons reaching distant instruments.

At its core are two-dimensional bivariate bicycle codes: qubits and checks are placed on a torus, and each check is additionally linked to two distant qubits. For the gross code (144 physical qubits, 12 logical qubits, distance 12) and the two-gross code (288 physical qubits, 12 logical qubits, distance 18), explicit syndrome measurement circuits, logical operations via code deformation (quantum surgery), and injection of T-states from factories are constructed. Fault tolerance was verified by simulating circuit-level noise with error probability p. At p=10⁻³, the logical idle error for gross was 10⁻⁸·⁸, and for two-gross it was 10⁻²⁰·¹. These are many orders of magnitude lower than for a surface code of comparable distance.

End-to-end estimates for random circuits and simulations of the transverse-field Ising model showed that with 5 thousand physical qubits and p=10⁻³, the bicycle architecture performs 10 times more logical operations than the surface code. For the Ising model, 8138 qubits suffice at p≈7×10⁻⁴, compared to 58–73 thousand for the surface code. This is the first end-to-end demonstration that modularity and long-range links yield an order-of-magnitude gain in the number of logical qubits without compromising reliability.

Future prospects involve improving inter-module measurements, whose logical errors currently limit overall performance. New adapter designs, decoder optimization, and parallel measurements are possible. Another direction is increasing code distances: the family of bivariate bicycle codes allows scaling, as do other low-density parity-check codes, up to constant-rate codes. Finally, magic-state factories can be improved by using long-range links for cultivation or distillation. The bicycle architecture sets new standards for designing quantum processors, especially superconducting ones where long-range links are already being demonstrated, and for compilers.