Member-only story Quantum Sundays |76⟩ The 38% Qubit Cut — What AlphaEvolve Actually Found Inside a Quantum Code An AI agent deleted a ring of edges from a graph and saved 104 qubits. The interesting part is why that worked, and what it says about how machine-driven discovery is going to run. TL;DR. Google’s AlphaEvolve, an evolutionary coding agent, searched a design space of roughly 10⁵⁰ possible fermion-to-qubit encodings and found a way to simulate three small molecules on a fault-tolerant quantum computer using 168 data qubits instead of 272, a 38% cut, with error protection held at exact code distance 5. The 38% measures against qBraid’s own deliberately over-provisioned starting point. Against the conventional route, Jordan-Wigner mapping plus a surface code wrapped around every qubit, the same construction runs about 4.8 times leaner: 168 qubits against 800. The discovery itself fits in one sentence. Start from a nearly complete graph, then delete a ring of edges chosen from the pairs of modes that never physically interact. That drops every vertex degree by two, saves a qubit at each one, and leaves the molecular physics untouched. AlphaEvolve returned it as readable Python, scored by a verifier it was never allowed to modify, which is the part of this story most worth stealing.