“John Hasse believes that a 10-20% performance advantage in large language model benchmarks is less important for military success than the quality and training of soldiers.”