The usefulness-gap study challenges what Pearl mining proves in practice

Abhinaba Basu’s study reports that the observed Pearl network performed no useful AI computation during the measurement period. It argues that miners could submit random matrices, and that the dominant mining software appeared to contain no inference code.

The paper also estimates network composition, energy, GPU-rental and mining economics. Those figures belong to the author’s method and shouldn’t be detached from it.

Pearl can verify that qualifying matrix multiplication happened. That isn’t the same as verifying that a customer needed those matrices. Usefulness can emerge through integrated providers, but no valid proof guarantees it.

• Useful work as a share of all accepted work.
• A public definition of what counts as an outside workload.
• Independent replication of provider-side inference integration.
• Cost and latency before and after proof generation.
• Whether the useful share grows as commercial endpoints gain demand.

Dismissing the criticism because the protocol can support useful work would be a category error. Declaring the construction worthless because current mining isn’t useful would be another.