A fault-tolerant quantum computer needs a decoder: something that reads the error syndrome and works out what went wrong fast enough to correct it before the errors compound. High-capacity neural decoders are the leading candidates because they stay accurate at large code distances.

The obstacle is how the cost scales. Larger code distances are exactly where you need these decoders, and they are also where generating syndrome training data and optimising the network become rapidly more expensive. The method works best in the regime that is hardest to build it for.

Neural transfer unification is the response, and the name states the strategy: rather than training a fresh decoder for each code distance, transfer what was learned at smaller distances into the larger ones. The barrier here is economic rather than theoretical, and reducing what each new distance costs is what turns a promising decoder into a usable one.