Reducing Entropy in AI-Assisted Software Creation by Removing the Programming Language Layer: An Information-Theoretic Analysis of Token Utilization Efficiency

Programming languages have traditionally served asthe main symbolic bridge between human intent and executablemachine behavior. In AI-assisted software creation, however,large language models increasingly translate natural-languageintent into source code, tests, compiler artifacts, and deploymentassets. This paper examines whether the programming-languagelayer is always necessary once the primary generator of softwarebecomes an AI model rather than a human programmer. UsingShannon entropy, Kolmogorov complexity, algorithmic informationtheory, Solomonoff induction, and the Minimum DescriptionLength principle, we argue that source code can be both anentropy-reducing representation and an entropy-increasing intermediatelayer. It reduces ambiguity for humans and tools, butit also introduces syntax, boilerplate, framework conventions, andaccidental complexity. We propose Token Utilization Efficiencyas a lifecycle metric that includes input, output, repair, andverification tokens. A small reproducibility package comparesC source, compiler-derived LLVM IR, and executable binariesfor five benchmark tasks. The results show that direct lowerlevelartifacts are not automatically more compact or morereviewable; LLVM IR and binaries were larger than source inthis experiment. The paper therefore concludes that removing theprogramming-language layer is promising only when compactexecutable representations are paired with strong verificationartifacts and human-readable explanations.Index Terms—AI-assisted software engineering, entropy, tokenefficiency, Kolmogorov complexity, algorithmic information theory,code generation, LLVM IR, executable synthesis. Files