500B Tokens Later: Letting AI Agents Decompile a First-Person Shooter (momo5502.com)

An ambitious project focused on using AI to decompile a popular first-person shooter has resulted in an impressive reconstruction of the game within just three months. The project aimed at creating a feature-complete C++ version of the game, emphasizing not only semantic correctness but also readability and functionality across multiple platforms. The team harnessed multiple AI models—such as Claude Max and Codex Pro—while refining their setup to optimize performance and reduce token consumption. Despite initial setbacks, the incorporation of a byte-matching verification system significantly improved the quality of the decompiled code, allowing 99% of the game's functions to be reconstructed with 83% exhibiting identical semantics. This project stands as a significant case study for the AI/ML community, offering valuable lessons on AI orchestration, verification, and infrastructure optimization. Key insights include the necessity of precise instructions to prevent 'cheating' by autonomous agents, the importance of machine-checkable correctness criteria to ensure reliable outputs, and the need for periodic instruction refreshment to maintain focus in long-term projects. Ultimately, the success of this endeavor not only demonstrated the potential of AI in code reconstruction but also paved the way for cost-effective scalability, indicating a promising future for autonomous AI in complex tasks.