We are committed to addressing fundamental grand challenges in the fields of Science and AI. We do not restrict ourselves to any particular solution, but currently, we find the paradigm of AI + Science particularly promising—Science can help AI transition from empiricism to a more scientific approach, enabling researchers to understand and design AI more scientifically; in turn, AI can complement the shortcomings of traditional Science and bring new paradigms to it.

Grand Challenges in AI:

1. Science for AI: Endowing AI with human-like capabilities (AI does not necessarily have to be human-like, but it should at least possess human capabilities!). This includes visual reasoning, physical reasoning, the ability to learn from few samples, and the capacity for rapid adaptation. We aim to develop new models and algorithms from a more foundational perspective—representation learning, information theory (intelligence as compression), and inspiration from neuroscience—to blaze a trail different from the current scaling laws.

2. Science of AI: The scientific theory of AI. We hope to start from the methodology of scientific research, gain insights from highly controllable “toy” experiments, construct relevant AI theories (models), and use these models to inspire the design of better algorithms. Specifically, we aim to understand and control each element in the current pipeline more scientifically—models, objective functions, optimization, data, and various intriguing phenomena such as Grokking and Neural Scaling Laws.

Grand Challenges in Science:

1. AI for Fundamental Science: Emergence. Emergence is directly or indirectly related to many scientific puzzles—such as Hilbert's Sixth Problem, high-temperature superconductivity, and controlled nuclear fusion. These problems/systems are complex because the scientific tools for studying emergence are very limited, and AI now offers a new tool and brings new possibilities.

2. AI for Everyday Science: AI assistants. Assisting in daily tasks, such as intelligently performing data analysis and visualization, proposing hypotheses based on data, and designing validation schemes, among others.