The article "Unsupervised Cross-lingual Representation Learning at Scale" by Conneau et al. presents a method to teach computers to understand languages without direct translation. The researchers developed a large-scale model that can learn to represent words in different languages in a shared space. This allows the model to understand relationships between words in different languages, even if they don't have direct translations. The key finding is that this approach can help computers understand languages better and improve tasks like translation and language processing.