According to the blog post, the software industry is experiencing a troubling shift: some developers report they haven’t written code since 2025, code reviews are declared dead, and people no longer read code. While the industry is transforming, the author warns that organizations abandoning hands-on coding are taking on significant risk.
The core problem, according to the post, is that code maintainability and good architecture lack clear, immediate measurements. Bad effects from poor design can take months or years to surface. Experienced engineers develop an intuition for detecting code smells early—a skill built through extensive debugging and learning from past mistakes. This expertise is context-dependent and cannot be reduced to rigid rules.
AI systems, however, face a fundamental limitation: they cannot be trained on maintainability. Reinforcement learning requires immediate reward signals, not outcomes that appear years later. Current AI models learn from patterns in existing code—much of which is poor—and from beginner-focused rulebooks. The author notes that AI is notably bad at simplifying code, often splitting functions into smaller ones that aren’t actually reusable. Properly defining clarifying and reusable functions is an art requiring mastery, which most developers and current AI systems lack.
The author argues that no fitness function for maintainable code has been successfully defined, otherwise it would already be embedded in code linters. This gap between AI capability and the true requirements of software engineering is crucial.
The real danger, according to the post, emerges when developers stop making choices and taking responsibility for coding mistakes. When people rely entirely on AI to write and read code, they lose the opportunity to learn from errors. Since the AI doesn’t learn from its mistakes either, both the system and the developer stagnate.
While the author acknowledges LLMs as useful tools for handling tedious work and sees efficiency benefits, they view AI as just another tool with a limited lifespan. The author predicts that in the future, companies will market “NO-AI” policies as competitive advantages—and will be justified in doing so. Without responsible use, there will be consequences for those who outsource critical thinking in software development.
Key facts
- The author argues AI cannot be trained on code maintainability because it requires immediate reward signals, not outcomes that appear years later.
- Bad code architecture effects take months or years to become apparent, a timeline incompatible with how AI systems learn.
- The author warns that developers who stop writing and reading code lose the opportunity to learn from mistakes and reach mastery.
- Current AI models struggle with code simplification and function extraction—skills requiring expert judgment.
- The author predicts companies will market ‘NO-AI’ policies as competitive advantages in the future.