"AI creates code in seconds." Nonsense.

It’s complete BS to say “AI creates code in seconds.” That’s just a lie that I’m seeing all too often. AI creates a first draft in seconds, and that draft has to be extensively edited before it’s even marginally acceptable as production-quality code. It moves the work from authoring to editing. The AI squirts the paint on the palette; it doesn’t create the painting.

Consider: “We find that the adoption of Cursor leads to a significant, large, but transient increase in project-level development velocity, along with a significant and persistent increase in static analysis warnings and code complexity … [which] acts as a major factor causing long-term velocity slowdown.” [https://lnkd.in/g3MVvDPV]

To me, this study does NOT tell me not to use Cursor (or any other LLM code-generation tool), but rather underscores the importance of using the LLM as a starting point in the code-creation process, not the endpoint. I find that LLM-generated code _always_ requires manual refinement. It shifts our work from authoring to editing, but editing (making the prose tighter and smaller) is an essential part of the writing process. A writer who cannot edit is not a writer at all. The fact that only the best programmers edit their code, even before LLMs were on the scene, is a huge problem for me. Working with an LLM and perpetuating the bad habit of not editing is a disaster—a time bomb waiting to go off in your application.

That’s the “long-term velocity slowdown” they’re talking about—dealing with the disaster that is excess code complexity. Unless you plan to sell off your company and foist the maintenance problems onto a buyer so hapless that they don’t do due diligence, LLM-assisted engineering will be a disaster unless it’s actual engineering, with code quality being a significant factor in your work. Another reminder that LLMs do not replace programmers, they just change the way we work. The LLM saves time. The editing adds at least some of it back. Deal with it.