He Pointed 16 AI Agents at a Game Binary. 83% of Functions Came Back Byte-Identical
Developer Maurice Heumann has detailed how he spent about three months using a range of AI models to analyze a game executable and turn it back into C++. The reconstructed code remains private, so there is no way for outsiders to verify it directly.
On October 9, developer Maurice Heumann published an account of how he used a fleet of AI models to bring a first-person shooter’s source code back to life. Over roughly three months, he analyzed the finished game executable and worked it back into human-readable C++. According to Heumann, 99% of the game’s functions are now covered by the reconstructed code, and 83% of all functions matched the original byte for byte once recompiled.
How he confirmed the code matched the original
Whenever an AI analyzed the executable and produced code, that code was compiled with the same compiler used to build the original game and then compared against the shipped binary. Several autonomous AI agents were assigned to the work. In the final weeks, Heumann says he was mostly running 16 of them at once.
Within the first month, about 80% had been reconstructed, far enough along that the game would launch, show its menus, and load maps. But Heumann explains that the code at that stage was riddled with errors in function signatures, data structures, and more. The game rendered and some features worked, yet problems remained buried in the code. That is when he introduced byte-level comparison against the original.
The checks were run function by function. Machine-code bytes were compared one at a time, with address-dependent values verified separately. Even functions that had already passed were re-examined by automated checks, a way of catching code that had once matched but drifted out of alignment after later edits.
What happened to the remaining 17%
Heumann says some of the remaining functions are simply hard to reproduce identically at the compiler-output level. Even where the bytes do not match, he believes the behavior is still correct.
The volume of work the AI chewed through is only an estimate. The title of his post cites more than 500 billion, while Heumann’s own estimate of the volume actually processed is 600 billion to 700 billion tokens, the small text fragments models process. Part of the logs was lost, so there is no exact tally, and he did not disclose what the project cost.
Can anyone else try it?
Heumann has not named the game or released the reconstructed code, and he intends to keep the code private. There is no way for anyone outside the project, in Korea or elsewhere, to download the result and check it for themselves.
The most striking detail is that even after the game launched and loaded its menus, he kept finding errors and changed how he verified the work. Whether the functions that do not match byte for byte actually behave correctly, how much compute and money the effort consumed, and whether any artifact is ever published for outside verification are the next things to watch in judging this result.