"Do You Still Read the Code?" • zanlib

Open original

The author argues that developers using AI fall into two distinct camps—those who read and understand generated code ("accelerators") and those who delegate implementation entirely to AI ("vibecoders")—and that these aren't points on a spectrum but fundamentally different commitments with different long-term costs.

  • Reading code matters because programming is ultimately about building a mental model of the domain, not just producing working software; without understanding the implementation, you lose the ability to explain decisions, anticipate consequences, and adapt to change.
  • Vibecoders risk accumulating "intent debt" as context drifts between sessions and requirements get lost, while accelerators risk drifting into vibecoding if they stop reading diffs—there's no safe middle ground.
  • The choice between these approaches isn't about how much the AI writes, but about your relationship to the output and whether you're committed to maintaining ownership of the reasoning behind the code.

The author challenges the assumption that reading code is becoming obsolete with AI. They distinguish between two camps: "accelerators" who use AI as a tool while maintaining deep understanding of what gets generated, and "vibecoders" who delegate implementation to AI and focus on specifying behavior and evaluating results. The key difference isn't how much code the AI writes—it's the developer's relationship to that code. An accelerator might generate most of a feature but takes ownership of understanding the reasoning behind it. A vibecoder might spend time refining specifications while treating the implementation as disposable. The author notes this split reflects fundamentally different bets about what matters in software development, and we're making long-term commitments without knowing which approach will actually hold up under the weight of maintenance, changing requirements, and team turnover. The author grounds their preference in Peter Naur's 1985 essay arguing that programming is theory-building, not code-production. The real asset isn't the source code itself—it's the mental model, the understanding of why each part exists and how it connects to the actual world it's meant to serve. A programmer with this theory can explain solutions, justify design choices, and adapt when circumstances change. Without it, you have brittle code that breaks when requirements shift. The code is like steam from a cooling tower—visible but not the actual product. What matters is whether the programmer understands the domain well enough to know if their sorting of reality actually matches reality. The author emphasizes that naming and structuring business logic through code serves a purpose beyond execution. It forces implicit knowledge into explicit form, which can help the people whose work is being modeled understand their own processes differently. But this only works if someone actually understands what they're modeling. Testing can only verify that code does what you told it to do, not that what you told it corresponds to reality. A vibecoder approach risks accumulating "intent debt"—forgotten requirements, drifting context, growing dependence on AI systems controlled by other companies. The real question isn't whether reading code will survive, but whether teams will maintain the understanding necessary to adapt when things inevitably need to change.

Ask about this article