The solution might be cancelling my AI subscription

Recently, prominent developer Simon Willison shared a reflective piece by fellow engineer David Wilson, sparking immediate resonance across the developer community. The core pain point it highlights is strikingly real: since fully integrating AI coding agents into our daily workflows, we seem to have collectively fallen into a strange trap. What starts as a simple request to write a quick utility script often spirals. An hour later, the AI has generated a polished-looking repository complete with unit tests, comprehensive documentation, and an architecture that appears meticulously refined over weeks. The irony, however, is that these projects are almost instantly abandoned the moment they compile successfully. The original itch that triggered the session is rarely solved, leaving behind a graveyard of half-finished digital artifacts.

This points to a severely underestimated phenomenon: the emergence of the AI Project Bubble. In traditional software development, the physical friction of writing code acted as a natural, necessary filter. It forced engineers to pause, prototype mentally, and ask before typing a single line: Is this requirement actually worth the investment? Today, generative AI has reduced that friction to absolute zero. You can see a fully functional prototype in ten minutes. Your brain floods with dopamine, creating a powerful illusion of progress and completion. But seasoned engineers know that software was never just about writing code. Deployment pipelines, edge-case handling, security audits, and long-term maintenance are where the true costs accumulate. When output becomes incredibly cheap and instantaneous, these abandoned prototypes transform into cognitive liabilities that silently drain our attention and engineering capacity.

What makes this dynamic even more fascinating is the deeply polarized impact AI has on human attention spans. For the majority of developers, AI tools function as a thermonuclear ADHD amplifier. They lower the barrier to starting new things so drastically that we end up juggling multiple screens, tinkering with unrelated side projects in a false sense of hyper-productivity. Yet, paradoxically, for individuals who actually live with ADHD, these same AI agents serve as a rare focus exoskeleton. They act as a tireless pair programmer who maintains momentum, helping neurodivergent creators bridge the painful gap between a fleeting spark of inspiration and a working prototype before boredom inevitably sets in. This dichotomy reveals a profound shift in our cognitive workflows: as the technical barrier to implementation completely vanishes, the relationship between humans and AI is no longer about who can prompt a model faster. It is about who can better regulate their own impulses and resist the seduction of instant, low-effort gratification.

While much of the current industry discourse is still locked in a technical arms race to optimize context windows, fine-tune retrieval pipelines, and architect complex multi-agent swarms, Willison offers a strikingly counterintuitive prescription: sometimes, the most effective AI workflow is simply to step back, or even cancel your subscription temporarily. This is not a call for technological regression. It is a necessary reminder to return to first principles in software engineering and product thinking. For IT professionals navigating this new era, the real competitive advantage is rapidly shifting from raw code generation capability to requirement filtering and project restraint. We must establish strict, almost bureaucratic, AI Project Onboarding Criteria. Before letting an agent spin up a new repository, force yourself to answer three questions: If AI did not exist, would I still dedicate my personal time to this? Who will realistically maintain this code six months from now? Does it solve a validated, high-impact problem, or is it just a dopamine-driven distraction? If the answers are vague or lean toward the latter, closing the chat window and returning to deep, friction-filled work might just be the most efficient engineering decision you can make. The future of software development will not be won by those who generate the most code with AI, but by those who possess the discipline to know exactly when to stop. This shift demands a fundamental rewrite of our engineering habits. Instead of treating AI as an infinite idea factory, we should treat it as a strict implementation partner. Successful teams are already experimenting with AI cooling-off periods, mandating a twenty-four hour review window before any AI-generated code merges into the main branch. They are also redefining done not as code that runs, but as code that is deployed, monitored, and actively used. For individual contributors, this means cultivating a new kind of digital minimalism. The most valuable skill in the coming years is not knowing how to chain ten tools together; it is the ability to look at a glowing prompt interface, feel the urge to build something trivial, and consciously choose to step away. In a world where creation is frictionless, restraint becomes the ultimate superpower.