MoonBit, a new general-purpose programming language designed for cloud and edge computing, was initiated in late 2022, coinciding with the announcement of ChatGPT. Language models like GPT, capable of producing practical programs, are revolutionizing the way we write programs and interact with computers. However, significant challenges persist, such as the models’ inability to understand the global context of a whole project with its dependencies, the need for human verification and correction of generated code, and the lack of assurance in meeting basic requirements like syntactic correctness.

In this paper, we explore the design of the MoonBit language highlighting its AI integration, emphasizing the synergy between traditional code intelligence and large language model capabilities. We also introduce a real-time, semantics-based sampler to guide the inference process of language models. This approach ensures the generated programs are both syntactically correct and free from obvious semantic flaws, such as type errors. Crucially, this has been achieved with minimal impact on overall performance. Our evaluation demonstrates a notable improvement in code quality, achieved without sacrificing the models’ responsiveness.