We are committed to building the theories and technologies that enable artificial intelligence, particularly foundation models, which lie at its core, to derive intelligence from data. With a special focus on language, one of the most challenging domains for AI, our ultimate goal is to elucidate the fundamental principles by which AI acquires and exercises linguistic competence.

AI-related technologies are now ubiquitous as practical tools throughout society, yet fundamental hurdles remain. For example, it is far from straightforward for humans to understand and explain how deep learning acquires and leverages useful knowledge and cues from data. At the same time, new research challenges are continuously coming into sharper focus, including issues of fairness arising from biased data, misinformation resulting from the misuse of AI technologies, and the enormous energy and financial costs involved in building foundation models. Moreover, for AI to be used in society with confidence, significant technical challenges remain, including alignment with human values, improvements in reliability, safety, and explainability, and the establishment of robust frameworks to evaluate these properties. By confronting this broad spectrum of both longstanding and emerging issues surrounding AI technologies, we strive to clarify their underlying principles and essential nature through rigorous theoretical and empirical investigation.