LLM-assisted coding is not deterministic. Does it matter? (blog.vrypan.net)

A recent discussion highlights the distinction between determinism and predictability in AI-assisted coding, emphasizing that while both concepts are related, they are not synonymous. Determinism in systems means identical conditions yield the same results, whereas predictability refers to our ability to foresee those results based on current knowledge and tools. Traditional software development has never been strictly deterministic, as human developers often introduce unpredictability in the coding process. Similarly, AI models like large language models (LLMs) likewise generate code that can be unpredictable. This creates a complex landscape where the focus should shift from whether the coding process is deterministic to how predictable the outcomes can be in practice. This conversation holds significant implications for the AI/ML community, particularly in defining workflows that enhance predictability in coding. While LLMs may not be deterministic, their growing accuracy in generating code is of interest. By developing robust testing practices and verification processes akin to the DO-178C guidelines for safety-critical software, AI tools can be better integrated into software development. The ultimate goal is to determine which workflows yield more reliable results, whether those involve human developers, LLMs, or a hybrid approach, especially in environments characterized by inherent complexity and the potential for unforeseen issues.