500B Tokens Later: Letting AI Agents Decompile a First-Person Shooter

How AI Agents Learned to Read Game Code Like Humans

On October 9, 2026, software engineer Maurice Heumann revealed he spent three months using AI agents to decompile a popular first-person shooter game. The project consumed approximately 500 billion tokens of computational effort. His goal was not merely to create a proof of concept but to produce an accurate, stable, and feature-complete recreation of the original game. Heumann documented the process on his personal blog, detailing the technical challenges and breakthroughs encountered along the way.

The decompilation effort involved training AI models to interpret low-level machine code and reconstruct high-level game logic, assets, and mechanics. Heumann emphasized that traditional reverse engineering tools often fall short when dealing with modern game engines due to obfuscation and dynamic code generation. By leveraging large language models and specialized prompting techniques, his team guided AI agents to identify patterns in assembly output that corresponded to gameplay systems such as physics, rendering, and input handling. The process required iterative refinement, with human oversight to correct misinterpretations and validate reconstructed code against known game behavior.

Can AI-Decompiled Games Be Legally Distributed?

Unlike conventional disassemblers that produce raw assembly listings, the AI agents were trained to infer intent from code patterns. Heumann explained that the models learned to recognize common game development idioms, such as entity-component systems or state machines, even when symbols were stripped. This allowed them to rebuild logical structures rather than just translate instructions. The team used a combination of supervised learning on known game binaries and reinforcement learning to reward outputs that could compile and run without crashes. Over time, the agents began to suggest meaningful variable names and function purposes, significantly reducing the manual labor typically required in decompilation efforts.

Heumann acknowledged the legal gray area surrounding such projects, noting that while decompilation for interoperability or preservation may fall under fair use in some jurisdictions, distributing the recreated game could infringe on copyright. He stressed that his work was conducted solely for research and educational purposes, with no intention to publish or share the reconstructed game. The project instead aimed to explore the limits of AI in software understanding and to develop tools that could aid in legacy game preservation or security auditing. He called for clearer legal frameworks to address AI-assisted reverse engineering as the technology matures.

What game was decompiled in this project? Heumann did not disclose the specific title of the first-person shooter, citing legal and ethical considerations. He described it only as a popularmodern FPS released within the last five years.

Frequently Asked Questions

How accurate is the AI-generated recreation compared to the original? According to Heumann, the recreated game matches the original in core mechanics, level layout, and visual behavior, with minor differences in non-essential animations and edge-case physics interactions that do not affect gameplay.

Could this method be applied to other types of software? Yes, Heumann believes the approach is transferable to other complex binaries, including legacy applications and firmware, especially where source code is lost or unavailable. He is already experimenting with similar techniques on older simulation software.