In this paper we demonstrate that a recurrent neural network meta-trained on an ensemble of arbitrary classification tasks can be used as an approximation of the Bayes optimal classifier. This result is obtained by relying on the framework of e-free approximate Bayesian inference, where the Bayesian posterior is approximated by training a neural network using synthetic samples. We denote the resulting model as neural ensembler. We show that a single neural ensembler trained on a large set of synthetic data achieves competitive classification performance on multiple real-world classification problems without additional training.