Imagine an AI not just crunching numbers, but actually coming up with new mathematical ideas. That's the tantalizing possibility explored in recent research examining whether Large Language Models (LLMs) can generate novel knowledge, particularly in spatial reasoning. Researchers put LLMs like Claude 3, ChatGPT, and Bing to the test with unusual spatial problems, including a little-known combinatorial game and a family of 24-sided polygons with unique properties. The goal was to see if these AI models could go beyond simply regurgitating information from their training data and actually create something new. The results were intriguing. Claude 3, in particular, showed flashes of genuine insight. In the game scenario, it devised a winning strategy that, to the researchers' knowledge, hadn't been documented before. With the polygons, Claude 3 correctly identified a non-trivial property: these shapes always have an even number of right angles. It even suggested that these polygons could tile a plane, a complex idea that turned out to be true. While other LLMs like Bing also offered some interesting ideas, Claude 3 consistently demonstrated a higher level of creative problem-solving. However, the research also highlighted the limitations of current LLMs. They still produce incorrect or irrelevant information, much like a math student still learning the ropes. The challenge for future LLM development lies in refining their ability to filter out these errors and focus on generating truly novel and useful insights. This research offers a glimpse into a future where AI could become a powerful partner in mathematical discovery, pushing the boundaries of human knowledge in ways we can only begin to imagine.