We compare linear zonal predictors of atmospheric turbulence for adaptive optics. Zonal prediction has the possible advantage of being able to interpret and utilize wind-velocity information from the wavefront sensor better than modal prediction. For simulated open-loop atmospheric data for a 2- meter 16-subaperture AO telescope with 5 millisecond prediction and a lookback of 4 slope-vectors, we find that Widrow-Hoff Delta-Rule training of linear nets and Back- Propagation training of non-linear multilayer neural networks is quite slow, getting stuck on plateaus or in local minima. Recursive Least Squares training of linear predictors is two orders of magnitude faster and it also converges to the solution with global minimum error. We have successfully implemented Amari's Adaptive Natural Gradient Learning (ANGL) technique for a linear zonal predictor, which premultiplies the Delta-Rule gradients with a matrix that orthogonalizes the parameter space and speeds up the training by two orders of magnitude, like the Recursive Least Squares predictor. This shows that the simple Widrow-Hoff Delta-Rule's slow convergence is not a fluke. In the case of bright guidestars, the ANGL, RLS, and standard matrix-inversion least-squares (MILS) algorithms all converge to the same global minimum linear total phase error (approximately 0.18 rad2), which is only approximately 5% higher than the spatial phase error (approximately 0.17 rad2), and is approximately 33% lower than the total 'naive' phase error without prediction (approximately 0.27 rad2). ANGL can, in principle, also be extended to make non-linear neural network training feasible for these large networks, with the potential to lower the predictor error below the linear predictor error. We will soon scale our linear work to the approximately 108-subaperture MMT AO system, both with simulations and real wavefront sensor data from prime focus.