Combatting Software Sprawl with LLMs

Abstract: Especially in manufacturing, many development tasks involve translation—converting natural language (often regulations) into formal engineering models, or bridging gaps between formalisms used by engineers with different expertise. While these translations are conceptually simple, they have historically been too costly to automate, forcing humans to handle tedious, repetitive work.

A software process sprawls when it becomes clogged with such translation tasks, creating inefficiencies akin to urban sprawl’s fragmented, low-density development. Large Language Models (LLMs) now offer a solution: they excel at low-cost, largely accurate translation, effectively acting as skeleton keys against software sprawl. By reshaping the economics of translation, LLMs unlock new levels of automation, accelerating process velocity and freeing engineers to focus on creative, high-value work.                                                                      

In this talk, I define software sprawl and demonstrate how it stems from communication that blends formal and explanatory/regulatory channels. I introduce the Dual Channel Hypothesis to explain the interaction of these channels and show how LLMs can "polymerize” sprawl — automating its consolidation and streamlining development.

Speaker: Earl Barr is a professor of software engineering at the University College London.  He received his PhD at University California Davis.  Earl's research interests include artificial intelligence for software engineering (and vice versa), debugging, testing and analysis, game theory, and computer security. His recent work focuses on probabilistically quantifying program equivalence, probabilistic type inference, and dual channel constraints.  With the exception of a pandemic-imposed hiatus, Earl dodges vans and taxis on his bike commute in London.