Martin Fowler's latest "Fragments" post is a candid reflection on the state of AI in software development. He covers a range of topics, from the AI bubble to the risks of LLM-generated code, and even the aesthetic distaste for AI-written prose.
The AI Bubble and Its Echoes
Fowler notes that we're clearly in an AI bubble, much like the dot-com era. The difference, he says, is that during the dot-com bubble, there was genuine excitement about building new things. Today, the focus seems to be on cost-cutting and efficiency, which feels less inspiring. He wonders if the eventual crash will leave behind anything as transformative as Amazon—or just a pile of hype.
The Board vs. Engineer Divide
A recurring theme is the mismatch between management's enthusiasm for AI and engineers' skepticism. Boards see LLMs as a way to boost productivity, but engineers worry about security, reliability, and the integrity of the codebase. Fowler shares an anecdote about a company that saved 100 million due to a mismatch in environmental conditions. The lesson: context is everything.
The Rise of Citizen Developers and Vibe Coding
Fowler highlights the growing trend of "vibe coding"—non-developers using LLMs to generate applications. While this democratizes development, it also creates shadow IT and governance nightmares. Companies are scrambling to put controls in place, such as separate infrastructure and deterministic data access, to tame the chaos.
Agents: Overhyped, But Useful
On the topic of AI agents, Fowler argues that many overestimate their capabilities. Incident resolution is rarely a linear process; it involves surprises and adaptation that LLMs struggle with. However, agents can be helpful for analyzing event streams and assisting with repeated incidents, as long as they document their actions and provide feedback to human teams.
The LLM-Speak Problem
Perhaps the most personal section is Fowler's rant against LLM-generated prose. He describes an almost visceral reaction to the telltale signs of AI writing—smooth, generic, and devoid of genuine voice. He worries that relying on LLMs for first drafts or even polish will strip writing of its individuality. His advice: read your writing aloud. Your spoken voice is closer to your true self, and that's what readers crave.
A Glimmer of Hope
Despite the criticism, Fowler acknowledges that LLMs can be valuable when used as evidence-based copilots—like in operations or when paired with domain-specific languages (DSLs). He points to examples where LLMs help translate high-level intent into deterministic code, ensuring safety and efficiency.
Ultimately, Fowler's post is a call to retain human agency. In a world flooded with AI-generated content, the distinctive human voice will stand out. And that's worth fighting for.