Their goal isn't to make a powerful model. It's to show how well compute-optimal models do on test-loss as a function of increasing model size. This function ca

Their goal isn't to make a powerful model. It's to show how well compute-optimal models do on test-loss as a function of increasing model size. This function can be used with some caveats to forecast the test-loss of larger models for which compute-optimality becomes more important.