Do You Still Read the Code?
zanlib ● Covered by 3 sources
Opinion — commentary, not a factual news event.
AI can write most of the code now, but teams still have to choose whether humans read it. That choice decides who can maintain the software later.
Based on reporting by zanlib — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
The question isn’t really whether people read code anymore. It’s whether they still mean to understand it. That’s the point the source keeps returning to: AI has made it easier to ship software, but it has also made maintenance, ownership, and team expectations much messier than the cheerful “still” in the title suggests.
The article draws a sharp line between two ways of working. “Accelerators” use AI as a faster way to translate their own understanding into code, and they keep reading the output so they can explain it, change it, and stand behind it. “Vibecoders” push further: they focus on specs, tests, and output, and treat implementation as something the model can regenerate again and again. The difference is not how many lines the model writes. It’s whether the developer intends to keep the theory of the program in their head.
That theory matters because code is not the real prize. The real asset, the piece that lets a programmer explain, predict, and adapt, is the model of the domain behind the code. The source leans on Peter Naur’s old argument here, and it’s persuasive: source code is visible, so people overvalue it. But a working system is only one stage in a longer process of learning what the business problem actually is. Running tests can show that the software does what was asked. It cannot prove that what was asked matches reality.
The danger is drift. A team can start as “accelerators,” then slowly slide into skimming diffs, forgetting justifications, and eventually asking the model what the code means without any reliable way to check the answer. On the other side, vibecoding can work for throwaway or exploratory projects, but it depends on strong habits around specifications and evaluation. The article is blunt about the worst mistake: putting both kinds of people on the same team without agreeing on who is expected to understand the implementation, and who is expected to treat it as disposable.
The author’s bottom line is pretty practical. Reading AI-generated code may not be enough forever, and it may even become exhausting. But for commercial work, it is still one of the few ways to catch the gap between what the model produced and what the human thinks was built. And that gap is where the real trouble lives.
My take — AI-written commentary, not fact-checked reporting
This is the part the AI cheer squad keeps skipping: responsibility does not disappear just because the code came from a model with a confident tone. If a team can’t say who owns understanding, it’s already outsourced the wrong thing. The industry has always loved pretending maintenance is tomorrow’s problem; AI just made that dodge faster.
Read more about this at: zanlib