500+ Billion Tokens Later: Letting AI Agents Decompile A First‑Person Shooter

Description and use

An AI agent trained on 500 billion tokens cut FPS decompilation from weeks to days, yielding a high‑level code base for analysis.

Problem

Manual decompilation of a modern FPS is time‑consuming and requires deep reverse‑engineering expertise.

Outcome

The AI‑driven workflow reduced decompilation time from weeks to days, producing a usable code base for further research.

Workflow

• 1. Gather binary and assets of the FPS
• 2. Feed binary into a custom prompt for the LLM
• 3. Run the LLM to generate decompiled code snippets
• 4. Validate and refine outputs with a human reviewer
• 5. Compile the reconstructed code for testing

Evidence

The author used a large language model trained on 500 billion tokens to automate the decompilation process.

Source

An AI agent produced a high‑level representation of the game’s code in a fraction of the time, enabling rapid analysis and modification.

Source