The essay observes that AI-assisted coding is already yielding concrete improvements in software quality, such as fewer bugs and better performance in existing programs, and is enabling teams to add more and better features to the products they maintain. It frames these observations as conjectural answers to the question of why we do not see a surge of new, good software despite advances in AI‑driven coding.
However, these gains often remain invisible to most users because the benefits are unevenly distributed. Many organizations still rely on legacy decision‑making layers—senior engineers and product managers who must sign off on plans—that have not been accelerated by AI, so the ‘we’ that experiences improvement stays small. At the same time, AI is shifting software toward more ephemeral or private forms (e.g., answering a tax question via a chatbot instead of installing tax‑preparation software) and toward bespoke internal tools that never reach a broad audience, while domain experts and executives build their own private dashboards.
Furthermore, the essay notes that development teams sometimes choose to generate more software rather than refine what exists, and they may capture AI‑driven speed‑ups as extra leisure time—offering 2× speedups to their companies and then using tools that let them work 10× as efficiently before taking long coffee breaks. In other cases, AI helps practitioners realize a project is a bad idea or encourages analog activities like reading books, meaning that the net effect on visible software can be neutral or even negative.
The author also points out that much of the improvement goes beyond the software itself, affecting larger systems that blend code with non‑software elements; for example, when a recommendation service like Letterboxd works perfectly for a user, the result is a happier video‑consumption experience that may crowd out the use of other software, showing that gains are measured in overall satisfaction rather than in new software releases.