Training large language models (LLMs) like those powering ChatGPT is a computationally intensive process, demanding vast resources and energy. Researchers are constantly seeking ways to make this training more efficient, and a new paper explores a promising technique called "model growth." Imagine building with LEGOs: instead of starting from scratch each time, you could reuse smaller, pre-built sections to construct larger, more complex creations. This is the core idea behind model growth. The research focuses on how to effectively "stack" smaller, pre-trained transformer models (the building blocks of LLMs) to create larger ones. They systematically tested different stacking methods and found that a simple "depthwise stacking" approach, called Gstack, significantly accelerates training. Gstack essentially duplicates and stacks layers of a smaller model to create a deeper, larger model. This method not only speeds up training but also leads to improved performance on various language tasks. The researchers tested Gstack on models with up to 7 billion parameters and found that the benefits persisted even with massive amounts of training data. For example, they achieved a 54.6% speedup in training a 7-billion parameter model. They also developed guidelines for when and how to best apply Gstack, making it a practical tool for LLM training. This research opens up exciting possibilities for more efficient and sustainable AI development. By reusing existing knowledge, we can reduce the computational burden and environmental impact of training ever-larger language models, paving the way for more powerful and accessible AI in the future.