Wysiwid: What you see is what it does (essenceofsoftware.com)

The evolving role of large language models (LLMs) in software development is prompting a fundamental rethink of how we write and structure code. While LLMs excel at generating code snippets and even entire modules, they struggle with the complexity, interdependencies, and subtle nuances of large codebases—often introducing errors and breaking existing functionality like novice programmers. This limitation highlights longstanding flaws in current software design practices and underscores the need for a return to modularity and clearer abstraction, not just as theoretical ideals but as practical necessities to enable effective human-AI collaboration. A promising new approach centers on “concept design,” where software is decomposed into fully independent modules called concepts that are defined by user-facing purposes rather than traditional object-oriented entity stacks. These concepts remain isolated, avoiding shared mutable state, and are coordinated via fine-grained synchronizations—mediators that manage interactions and enforce behavioral constraints across concepts. This design pattern, demonstrated through prototypes like Kodless and experimental implementations on benchmark apps, aligns natural modularity with LLM capabilities by constraining context size and improving code legibility. The “What You See Is What It Does” (WYSIWID) paradigm thus ensures that the code’s structure closely reflects user behaviors, enabling scalable, incremental AI code generation while preserving software integrity. This shift carries major implications for the AI/ML community: advancing software engineering fundamentals is not just a human goal but essential for maximizing LLM effectiveness and reliability at scale. As LLMs improve, coupling their power with robust modular design can unlock orders-of-magnitude gains in development speed, code quality, and maintainability. Beyond research, this approach is already being taught at MIT and explored through hackathons, with future prospects including low-code SaaS tools and AI-augmented workflows that generate and manage concept specifications—signaling a new era where intelligent tooling and sound software design evolve hand-in-hand.