An Unfair Review of AI Agents: Who Truly Achieved Depixelizing Pixel Art?

As summer approaches and vacation plans are made, I felt inspired to write an engaging article while experimenting with new ideas. AI has various benchmarks, such as HumanEval, where models are tasked with writing short functions, or simpler challenges like creating a to-do list in React. But what happens when we present modern AI agents with a genuinely complex challenge? The task at hand is to implement an algorithm from a SIGGRAPH paper using Swift, without relying on any external libraries. The rules are straightforward: one prompt equals one implementation, with no clarifications, no feedback on errors, and no iterative corrections. The goal is to see what the AI can produce under these conditions.