LLM Programming and Theory Building (kevinlynagh.com)

In a recent exploration of Large Language Models (LLMs) for programming, the author introduced the concept of "Work" as a more flexible foundation for software architecture, moving away from the traditional task-subtask structure. This innovation stems from the author's struggle to streamline code generation and programming workflows using LLMs like Astra 6, Opus 5.5, and Sol 6.1. After numerous attempts to refine the code output from these models, the author realized that engaging more deeply with the structure and semantics of programming was essential for effective theory building and insight development. By conceptualizing "Work," the author unlocked a potentially more efficient and collaborative coding approach that allows for both structured tasks and ad-hoc executions in parallel. This exploration is significant for the AI/ML community as it touches on the broader implications of collaborative human-LLM interactions in programming. The struggle for clarity in LLM-generated code suggests a need for better engagement strategies when using these models. By proposing an iterative sketching method and questioning how LLMs can be used to fill gaps in structured programming, the conversation opens avenues for improving how AI can assist in programming without overwhelming users with excessive information. This insight echoes the long-standing belief that effective programming goes beyond mere execution; it is inherently a process of theory building and understanding.