The most efficient program for finding all the eigenvalues of a symmetric matrix is a combination of the Householder tridiagonalization and the QR algorithm. The latter, if carried out in a natural way, requires 4n additions, 10n multiplications, 2n divisions, and n square roots per iteration (n the order of the matrix). In 1963, Ortega and Kaiser showed that the process can be carried out using no square roots (and saving 7n multiplications). However, their algorithm is unstable and several modifications were suggested to increase its accuracy. We, too, want to give such a modification together with some examples demonstrating the achieved accuracy.