Large Language Models (LLMs) are impressive, but their size and complexity make them expensive and raise privacy issues. This has sparked interest in smaller, more efficient models, but how can we make these smaller models reason as effectively as their larger counterparts? New research explores a clever technique called SKIntern, which focuses on teaching small language models (SLMs) to internalize knowledge, much like a human intern gradually absorbs information and becomes more independent. Initially, the SLM receives detailed guidance and symbolic knowledge from a larger LLM, similar to an intern receiving comprehensive training materials. This knowledge includes not just the "how" but also the "why" behind the answers. However, unlike existing methods that require constant access to this extra information, SKIntern focuses on internalizing this knowledge into the SLM’s parameters. Over time, the SLM learns to reason effectively without needing constant access to external resources. This process of gradual internalization significantly reduces computational overhead during inference—the stage where the model actually answers questions. The approach addresses the key challenge of making AI reasoning both effective and efficient, crucial for deployment in real-world scenarios with limited resources. Tests show that SKIntern not only boosts the reasoning performance of SLMs on a variety of tasks but also slashes the computational cost by up to 4 times compared to existing methods. This improvement was seen across different sizes of small models, even when training data was limited. Notably, SLMs enhanced with SKIntern even managed to outperform the larger "teacher" model on certain tasks! This innovative approach shows great promise for making AI reasoning more accessible and cost-effective, potentially unlocking numerous applications in areas where resources are constrained.