Microsoft's SkillOpt research has shown that AI skills can be transferred across various environments and model scales. The study found that a skill trained on one model can be successfully applied to another model, even if it was not trained on the same data. This ability to transfer skills was demonstrated in a test where a skill trained on a Codex model improved the performance of a Claude Code model. The success of skill transfer varies depending on the task type, with some tasks showing significant improvement and others showing little to no improvement. This development has implications for the development of more efficient and adaptable AI systems.