We deal with the practical construction of confidence intervals (CIs) for a real-valued, smooth parameter by targeted learning, when sample size is so large that the resulting computational problems cannot be skirted. We propose to carry out targeted learning on a sub-sample selected with unequal inclusion probabilities based on easy to observe summary measures of the data. As examples, we show how to use Sampford's and determinantal survey sampling designs. The inclusion probabilities can be optimized to the reduce the width of the CIs. A simulation study illustrates our results.