AI Replicates 16k-Line Go App From CLI Alone

### Why it matters An AI that can clone a program from its behavior is a powerful tool for competitive analysis, migrating legacy systems, or creating functional equivalents for security testing. The MirrorCode benchmark from METR and Epoch demonstrates that current models like Claude Opus 4.6 can already perform this kind of complex, long-horizon reverse engineering, a task previously considered a unique human skill requiring weeks of effort.

This moves the goalposts for AI coding capabilities from generating isolated functions to replicating entire systems. The fact that performance scales with inference compute means today's models might be even more capable than we realize, given enough time and resources. For engineers, this signals that AI is becoming a tool for system-level problems, not just line-level coding assistance.

### What changed * **MirrorCode Benchmark:** METR and Epoch released a benchmark where an AI agent must reimplement a program with only execute-only CLI access, not source code. * **Claude Opus 4.6 Success:** The model successfully reimplemented `gotree`, a bioinformatics toolkit with ~16,000 lines of Go and 40+ commands. * **Human-Scale Task:** Researchers estimate this task would take a human engineer 2–17 weeks to complete without AI assistance. * **Inference Scaling:** Performance was shown to improve with more inference compute, suggesting models can solve larger problems simply by being given more processing time.

### What to watch * **Larger Projects:** The study suggests that with enough tokens and compute, even larger proprietary software may be solvable. Watch for attempts to replicate more complex commercial applications. * **Beyond CLI Tools:** The current benchmark focuses on CLI programs with clear inputs and outputs. The next frontier will be replicating systems with GUIs, network protocols, or other complex, stateful interactions.