Last modified: 2025-08-04
I've been building software for over a decade, and lately I've been having this recurring conversation with fellow developers. It usually starts with someone complaining about AI-generated code – "It's not efficient," they'll say, or "I spent more time fixing Copilot's suggestions than it would have taken to write it myself." And I get it. I really do. But I think we're looking at this all wrong.
We're trying to use a chisel like a screwdriver, and then getting frustrated when it doesn't work the way we expect.
For decades, we've approached software development like master watchmakers. Every component had to be precisely crafted, each piece fitting perfectly with the next. We'd spend hours – sometimes days – architecting the perfect system before writing a single line of code. The mental model was clear: design thoroughly, build carefully, scale methodically.
This made sense when every line of code was expensive to write and even more expensive to change. I remember my first job out of college where we'd have architecture review meetings that lasted entire days, debating the merits of different design patterns. We treated code like precious metal – once forged, it was costly to reshape.
The statistics back this up. Traditional development has always emphasized careful planning because refactoring was expensive. In 2021, about 25% of changed lines were refactored code – developers carefully maintaining and improving their carefully crafted systems. We were artisans, and our code was our craft.
But here's the thing: AI has fundamentally changed the economics of code creation. GitHub Copilot users complete tasks 55% faster. The cost per line of code has dropped from $12 for a human developer to $0.002 with GPT-3. We're not in the watchmaker's workshop anymore – we're in the sculptor's studio.
Think about how a sculptor works. They don't carefully assemble individual pieces. They start with a massive block of marble – more material than they could ever need – and then chip away until the form emerges. They might rough out the general shape quickly, then refine, adjust, and polish. Sometimes they'll discover the grain of the marble suggests a different approach than they originally planned, and they'll adapt.
This is exactly what's happening with AI-assisted development. Instead of carefully designing every function and class, we can now generate entire systems in minutes – our block of marble. Then we refine, reshape, and polish until we have what we need.
The numbers tell this story: code churn (the percentage of code that gets rewritten shortly after being committed) has doubled since 2021, projected to exceed 7% in 2024. Copy-pasted code has risen from 8.3% to 12.3%. To the watchmaker mindset, this looks like chaos. To the sculptor, it's just part of the process.
Here's where the friction comes in. According to recent research, 66% of developers cite "AI solutions that are almost right, but not quite" as their biggest frustration. Trust in AI accuracy has dropped from 43% to 33% in just one year. And paradoxically, AI tools can actually slow seasoned developers by 19%.
But I think this frustration comes from expecting AI to be a better watchmaker, when it's actually teaching us to be sculptors.
When you ask ChatGPT or Copilot to generate code, you're not commissioning a Swiss watch. You're ordering a block of marble. It might have the general shape you want, but it's going to need refinement. The frustration comes when developers expect the AI to deliver a perfectly crafted component on the first try – like expecting Michelangelo's David to emerge fully formed from the quarry.
One developer I talked to put it perfectly: "I kept getting annoyed that Copilot would suggest these verbose solutions with redundant code. Then I realized – it doesn't matter. I can generate ten versions, pick the best parts, and refactor aggressively. The cost of iteration is basically zero."
The most successful AI-native companies are already embracing this sculptor's mindset. They're using what researchers call the "V-Bounce Model" – minimal time spent in implementation (the bounce at the bottom of the V), with emphasis shifting to requirements gathering, architecture design, and continuous validation.
This doesn't mean we abandon craftsmanship. Sculptors are still artisans – they just work with different tools and methods. In fact, 73% of developers report staying in flow state when using Copilot, and 87% say it preserves mental effort during repetitive tasks. We're not becoming less skilled; we're focusing our skills differently.
The key is to accumulate material first, then sculpt. Instead of trying to write the perfect authentication system from scratch, generate five different versions. Mix and match the best parts. Refactor aggressively. Delete entire chunks and regenerate them with different prompts. The goal isn't to get it right the first time – it's to explore the solution space rapidly until the right pattern emerges.
This is already happening at scale. Shopify developers are accepting 24,000+ lines of AI-generated code daily. Accenture expanded their AI coding initiative to 50,000 developers after seeing 96% success rates. These organizations aren't succeeding because the AI writes perfect code – they're succeeding because they've adapted their development process to match the new paradigm.
I'll admit, there's something romantically appealing about the watchmaker approach. The careful craftsmanship, the precision, the pride in every line of code. But there's also beauty in the sculptor's approach – the emergence of form from chaos, the discovery of unexpected solutions, the freedom to experiment without fear.
The developers who are thriving with AI aren't the ones trying to make it write perfect code. They're the ones who've learned to work with rough drafts, who can see the potential in imperfect solutions, who understand that when the cost of refactoring approaches zero, the entire calculus of software development changes.
We're not abandoning the artisan approach – we're evolving it. The watchmaker assembles; the sculptor reveals. Both create beauty, just through different means.
The next time you're frustrated with AI-generated code, try shifting your perspective. You're not looking at a failed watch – you're looking at a block of marble. The question isn't "Why isn't this perfect?" but rather "What can I sculpt from this?"
Because in the end, the goal hasn't changed. We're still trying to build beautiful, functional software that solves real problems. We just have new tools that require a new mindset. And once you make that mental shift – from assembling to sculpting – the frustration transforms into possibility.
The marble is cheap now. So grab your chisel and start carving.