Decoding error correction codes for cat qubits? We just took a first step with @nvidia NVQLink. 🔥 🔍 Quantum errors pile up fast, so they need to be fixed as they happen. That’s the job of decoding: a classical computer uses the syndrome measurements to infer the most likely errors, so their effect can be undone when interpreting the results. But decoding has to hit demanding performance targets in latency and throughput, or quantum computing won’t be feasible. That takes tight coupling between the quantum computer and the GPU. 🤝 That’s why NVQLink matters. It’s NVIDIA’s platform for tightly coupling QPUs and GPUs, and we’ve been testing it on the decoding of our LDPC-cat codes, Alice & Bob’s error-correcting codes for cat qubits. This first step is a proof of concept, run on synthetic error data: decoding our LDPC-cat codes on a GPU with NVQLink is feasible. 🙌🏻 We also found that batching pays off. Decoding many syndromes together makes latency grow sublinearly compared to decoding them one at a time. That’s a good sign for scaling, since bigger quantum computers will have far more errors to handle. 🛣️ Next: making it fast enough for real-time decoding, so error correction can keep pace with the quantum computer. ⚡ Stay tuned! The future of computing is GPUs and QPUs working side by side, and progress on decoding is another step toward the fast, scalable error correction that fault-tolerant quantum computing will need. ⚛️

Shared logical qubit definitions are how we stop comparing apples 🍎 to oranges 🍊 In June this year, we published our report the five criteria to define a logical qubit. Yesterday, @Microsoft Quantum and @QuantumOpticsLb published theirs, co-authored by John Martinis, 2025 Physics Nobel laureate and an advisor to Alice & Bob. We are happy to see the industry aligning! A logical qubit has to operate beyond break-even. It has to get more reliable as you add more qubits to it. It has to keep running QEC for the whole computation, not just a benchmark. And a single impressive number doesn't say much on its own. On all of this, we're aligned. 🤝 The paper puts a number on the finish line: logical error rates of 10⁻¹² to 10⁻¹⁵, depending on the application. We wish we'd thought of that. Having a number to reach for makes the results easier to place. Two notes in the margin 📝 🎯 Their paper allows post-selection, where you don't keep every run in the final result, and notes it can be useful as a near-term lever. We agree, and we agree it doesn't scale. Where we differ: for us, that's exactly why post-selected results shouldn't count as logical qubit claims. A logical error rate only means something if it holds across every run. Dropping the bad runs isn't fixing them. 🧮 Their definition requires a universal set of fault-tolerant gates with real-time decoding and feedback. Ours deliberately stopped at memory: a logical qubit’s ability to preserve information over time, and computing on it is the next question. We think both framings are needed, and we appreciate that they drew the line where they did. Haven't read our five criteria, or want a refresher? 😉👉 alice-bob.com/science/#five-… #QuantumComputing #LogicalQubit