One of the biggest mysteries in AI right now is how large language models (LLMs) develop surprising new abilities as they grow. We can track their progress on specific tasks, but it's often impossible to tell when an LLM will suddenly become good at something it was previously terrible at—a phenomenon called "emergence." This unpredictability makes planning for the future of AI difficult, for both developers and policymakers. New research has found a clever way to peek into the future of LLM abilities. Imagine trying to predict if a child prodigy will become a great musician as an adult. You could give them some advanced training now and see how they respond. If they quickly master the complex material, they’re more likely to excel later on, even without further intensive training. This research applies the same principle to LLMs. By "fine-tuning" current models—giving them focused training on a specific task—researchers can shift the point of emergence to smaller, less powerful models. This creates a sort of "shortcut" to observing how future, larger models might behave. The research introduces the concept of "emergence laws." These laws, based on how the point of emergence shifts with different amounts of fine-tuning, allows for predicting when an LLM will spontaneously develop a new skill, even before it shows any signs of it. This is like having a formula to predict the child prodigy’s future musical success based on their response to early training. The results are promising. In tests on tasks like question answering and grammar checks, the method accurately predicted when larger LLMs would suddenly become proficient, sometimes using models with just a quarter of the computing power needed for the skill to emerge naturally. This is like predicting the child's success years before their talent blossoms. This new research has exciting implications. Beyond predicting which tasks future LLMs will excel at, it could also help evaluate the quality of training data more efficiently and anticipate potentially risky emergent capabilities, like the ability to create malicious software. It offers a way to make the development and deployment of AI safer and more predictable. While the current method predicts emergence much better than chance, there's still a long way to go before we can predict LLM capabilities with the same precision we predict the trajectory of a rocket. Improving data selection and understanding *why* fine-tuning shifts the point of emergence are key next steps in this exciting research area.