Large language models (LLMs) are revolutionizing how we write code, but training them to follow instructions usually requires tons of human-written examples or access to proprietary, expensive models like GPT-4. What if LLMs could learn to code by themselves? Researchers have introduced SelfCodeAlign, an innovative technique that lets LLMs generate their *own* training data. This self-alignment process begins by feeding the LLM high-quality code snippets, from which it extracts core coding concepts. Then, like a student brainstorming practice problems, the LLM uses these concepts to generate new coding tasks and even writes test cases to check its own answers! Successful solutions then become part of a growing dataset used to fine-tune the LLM's coding abilities. The results are impressive. A relatively small 7-billion parameter model trained with SelfCodeAlign outperformed the much larger CodeLlama-70B on the HumanEval+ benchmark, a standard test for code generation. This self-learning approach also proved effective across different sizes of LLMs, suggesting smaller, open-source models can achieve impressive coding skills without relying on massive datasets or closed-source technology. SelfCodeAlign is not without limitations. Currently, it favors medium-sized code samples and the self-generated test cases aren’t always perfect. However, this research opens exciting possibilities for the future of AI-assisted coding. Imagine LLMs continually refining their skills by learning from their own successes – a self-improving cycle that could lead to even more powerful and efficient code generation tools. The research has already led to the development of StarCoder2-Instruct, a fully transparent and open-source code LLM demonstrating the real-world potential of self-alignment. As this technology matures, we might see a shift towards more accessible and powerful AI coding assistants for everyone.