A device can preserve a fragile state better than its individual components and still have a long way to go before it solves a useful problem. That gap is where engineering lives.
The distinction is easy to lose when quantum computing appears beside theories of emergent spacetime. Both involve quantum information. Their questions and evidence can nevertheless be quite different. A theoretical relation between entanglement and geometry does not certify a processor. A processor benchmark does not establish a theory of gravity.
To judge a technology claim, begin with the operation it actually performed. Then follow the chain from that operation to the result someone wants to buy.
A memory experiment has a defined job
Quantum error correction distributes a logical state across physical qubits. Repeated checks extract information about errors, and a classical decoding procedure interprets those checks. The aim is to protect the encoded information despite faults in the underlying hardware.
This is more demanding than buying additional components and assuming reliability will improve. The additional operations can introduce errors. The hardware, code and decoding procedure must work together well enough that increasing protection produces a net benefit.
Google Quantum AI and collaborators reported a concrete result in a paper published online in December 2024. Their Willow processors supported surface-code memories whose logical error rates decreased as code distance increased. A distance-seven memory used 101 qubits, including qubits for checks and leakage removal, and had a reported logical error of 0.143% per correction cycle. A separate distance-five memory incorporated real-time decoding. Research paper.
Code distance describes a property of the error-correcting code, not the physical distance between chips. The central result concerns the preservation of logical information under a specified protocol. It should not be retold as a measurement of business productivity.
The article’s April 2026 correction fixes labels in a repetition-code figure. Repetition-code results and surface-code results must remain distinct; the correction is part of the record. Author correction.
A memory experiment establishes a foundation. It does not, by itself, demonstrate a complete logical algorithm, an advantage on a customer’s workload or an affordable service. Those are further propositions with their own measurements.
AI can help interpret errors
There is a specific role for machine learning here. Given a history of noisy checks, a decoder must infer the correction relevant to the logical result. Learning patterns in those histories can improve that inference.
The AlphaQubit study, published in November 2024, used a recurrent transformer-based decoder. It reported stronger decoding accuracy than its compared alternatives on experimental Sycamore data for distance-three and distance-five surface codes. Tests extending to distance eleven used simulated data. The model was pretrained on synthetic samples and adapted using experimental samples; the study held back test data rather than grading only its training examples. Original study, experimental evaluation and simulation sections.
Its conclusion also identifies a practical constraint: the reported decoder’s throughput was slower than the target for superconducting hardware. Better accuracy and sufficiently fast operation are separate engineering requirements. The authors discuss routes toward improvement, rather than claiming that this constraint has already disappeared.
This is a useful example of AI contributing to scientific engineering. It learns a constrained decoding task with an observable error metric. It neither invents a new spacetime nor establishes that an unrestricted language model can reason reliably about every quantum experiment.
Simulation provides another important boundary. A simulated noise model can explore larger cases and isolate mechanisms. Real devices can contain errors that the model omits or misrepresents. A favorable simulated result deserves follow-up on hardware; it does not silently become that hardware result.
Follow the result through four questions
An engineering claim can be examined at four levels. These are questions to answer, rather than a universal development schedule.
What operation succeeded? Name the preparation, storage, gate or measurement task. Identify the metric and conditions. “More powerful” is too broad to tell a reader what happened.
How reliably can it be repeated? Include the faults, corrections and excluded runs. If a protocol discards difficult cases, account for those discarded cases when describing its useful output. Check stability across the durations and operating conditions the application would require.
Does the complete workload improve? A fast subroutine can sit inside an expensive workflow. Preparation, transfer, decoding and verification belong in the comparison. Use a relevant conventional method and require outputs of comparable quality.
Can the result be delivered at an acceptable cost? Include access, integration and maintenance. An impressive capability can be valuable for research before it makes economic sense as a routine service.
These questions explain why a narrowly described milestone can be both substantial and incomplete. The achievement does not become smaller when its scope is stated accurately. The missing steps become visible enough to work on.
Post-quantum cryptography is a different response
A company preparing for quantum computing need not start by purchasing a quantum processor. Post-quantum cryptography is designed to protect information against attacks by sufficiently capable quantum computers while running on conventional computers. The adjective describes the threat being addressed, not the machine required to run the protection. NIST’s explanation.
NIST’s current project page identifies the three principal standards released in 2024: ML-KEM for key establishment, and ML-DSA and SLH-DSA for digital signatures. It calls for migration work and distinguishes those published standards from additional algorithms still undergoing standardization. Current standards and migration overview.
For an organization, that creates concrete questions about where cryptography is used, what suppliers support and how changes can be tested. Those questions differ from forecasting the date of a cryptographically relevant quantum computer. An inventory and a migration plan can be assessed even while that date remains uncertain.
Implementation still needs the organization’s security expertise, applicable requirements and tested products. A name on an algorithm list does not validate a particular deployment. The useful claim must extend all the way to the system that actually protects the information.
The frontier needs an acceptance test
Imagine a vendor says its quantum-assisted system will improve a scheduling process. Before debating the label, identify the schedules it must produce, the constraints they must satisfy and the conventional approach it must beat. Measure the whole process, including the time required to verify a proposed answer.
This is a proposed way to evaluate a claim, not a scheduling experiment performed for this series. The same discipline would also apply to an ordinary software product. Quantum hardware adds important physical questions; it does not remove the need for a fair comparison.
The next frontier is more legible when each claim has a place. Theories can explain possible underlying structure. Experiments can establish physical performance. Engineering can turn performance into a dependable capability. Economic analysis can ask whether that capability improves a decision enough to justify its cost.
The last question requires attention to what the user gains. Counting qubits, model parameters or stored information cannot answer it on its own.
Loading comments…