The 180,000-Line Leap of Faith

Software development is traditionally a game of patience. It took Rick Brewster more than two decades to build Paint.NET, a popular image-editing program, carefully crafting its 700,000 lines of code. But recently, his codebase grew by a staggering 180,000 lines almost overnight. Brewster didn't write them, and by his own admission, he hasn't fully read them either.

The author of this massive update was Claude, an AI model. Brewster enlisted the AI to solve a long-standing headache: getting Paint.NET to run smoothly on Linux via the WINE compatibility layer. The bottleneck was Direct2D, a Windows graphics API that WINE couldn't fully support. Since he couldn't simply disable the graphics engine, Brewster asked Claude to perform a clean-room, from-scratch rewrite of Direct2D.

What followed was a fascinating glimpse into the reality of AI-assisted software engineering, highlighting both its breathtaking potential and its frustrating limitations.

Brewster described Claude’s performance as wildly inconsistent. At times, the AI worked with the intensity and brilliance of "10 freshly unshackled Einstein genius-level 10x coders." It tirelessly reverse-engineered complex mathematical formulas required to replicate Direct2D’s built-in visual effects library—a task that would have taken a human weeks of tedious labor.

Yet, for all its genius, the AI also required intense babysitting. It frequently stumbled over fundamental computer science concepts. For instance, Claude repeatedly failed to handle resource management correctly, forgetting to include basic reference counting (like the COM AddRef() method) which is crucial for preventing memory leaks and crashes. Brewster also had to intervene when the AI made fundamentally flawed architectural decisions, correcting its course before the foundation of the code was compromised.

Because it is humanly impossible to thoroughly review 180,000 lines of AI-generated code, Brewster embraced a controversial new paradigm known as "vibe coding." Instead of meticulously checking every syntax choice, he evaluated the code based on whether the overall software functioned correctly and felt right in practice. It is an approach to software engineering that relies heavily on testing the final output rather than micromanaging the input.

This case study perfectly encapsulates the shifting role of the human developer. As AI models become capable of churning out hundreds of thousands of lines of code, the job of a programmer is evolving. They are no longer just typists or syntax experts; they are becoming project managers, architects, and editors. The AI can lay the bricks at lightning speed, but it still takes a human to ensure the house is actually safe to live in.

Key Points

• An AI model wrote 180,000 lines of code to help Paint.NET run on Linux via WINE.
• The developer used 'vibe coding,' evaluating the code by its overall function rather than reviewing every line.
• The AI excelled at complex tasks like reverse-engineering math formulas for visual effects.
• Despite its brilliance, the AI required constant supervision to fix basic resource management and architectural errors.
• The role of human developers is shifting from writing code to managing and architecting AI output.

Why It Matters

This real-world example demystifies how AI is currently used in complex software engineering, proving that while AI offers immense productivity, human oversight remains the critical safety net.

Sources:

• Quoting Rick Brewster — Simon Willison's Weblog