The podcast discusses two approaches to learning implicit representations in neural networks, including treating them as an oracle and using GAN type neural networks to have a Bayes net of semantic latent variables with dependencies between them.