Imagine teaching a child to tie their shoes. Once they master it, they can use that skill in different contexts, like tying a ribbon or securing a knot. This ability to combine learned skills in novel ways is called compositional generalization, and it's something humans do effortlessly. But can AI models learn to do the same? A recent research paper explores this question, investigating whether smaller AI models can learn skill composition from examples. The research used a clever approach: generating text samples that combined various language skills – like rhetoric, reasoning, and even theory of mind – using GPT-4. These samples were then used to fine-tune smaller language models. The findings were surprising! Fine-tuning these smaller models on a limited dataset of combined skills actually improved their ability to compose *unseen* skills in new combinations. This suggests that the models weren't just memorizing specific combinations, but learning a higher-order meta-skill of composition. They were generalizing the concept of combining skills, much like a child learning to tie different kinds of knots. Even more intriguing, the research showed that training on combinations of just two or three skills led to improvements in composing four or five skills – combinations the models had never encountered during training. This opens up possibilities for more efficient training, focusing on smaller skill combinations to bootstrap the model's compositional abilities. The implications of this research extend beyond just language skills. If AI models can truly learn compositional generalization, it could pave the way for them to solve complex problems in various fields by combining their existing knowledge in creative ways. However, there are still limitations. The computational costs involved in training and evaluating these models are high, and more research is needed to explore compositional generalization in different contexts beyond language. Still, this research is a step forward in understanding how AI can evolve from “stochastic parrots,” simply mimicking training data, toward true compositional thinking.