Imagine an AI tackling the world's toughest math problems, the kind that stump even brilliant Olympiad contestants. That's the challenge researchers at Shanghai AI Lab and various universities took on with their creation of "Lean Workbook," a massive dataset designed to push the boundaries of what AI can achieve in formal mathematics. Large language models (LLMs) like those powering ChatGPT excel at many language tasks, even some math problems. But when it comes to formal theorem proving using languages like Lean, they struggle. A major roadblock is the lack of training data in these formal languages. Lean Workbook addresses this gap by translating thousands of natural language math problems, from middle school to Olympiad level, into Lean 4 code. This involved a clever system that generates and then filters potential translations using the Lean compiler itself and natural language processing techniques to ensure they mean the same thing. Human experts refined the most difficult translations, creating a reliable and challenging dataset. The result? A collection of nearly 57,000 formal-informal problem pairs, plus 21 new International Math Olympiad (IMO) problems now formalized in Lean. This represents a substantial increase in the type of complex math data LLMs can learn from. Initial tests show this training can significantly improve AI’s ability to translate and understand intricate math problems, and even find solutions. The Lean Workbook offers not just a wealth of problems but also the opportunity to elevate AI's mathematical reasoning to new heights, perhaps one day creating an AI mathematician capable of tackling the most challenging problems. While promising, this is just the start. Challenges remain, like handling problems with similar structures and expanding the dataset to encompass different levels of mathematical difficulty. Still, the Lean Workbook has laid a strong foundation for the next generation of AI theorem provers.