The DOOM map serves as the computational canvas for the DOOMFORMER transformer.
DOOMFORMER: AI TRANSFORMER BUILT FROM DOOM GAME ENGINE
*A GitHub project repurposes the classic shooter into a machine‑learning engine. *It claims to classify MNIST digits, generate speech‑like output, and even control a DOOM marine’s decisions. *The creator seeks collaborators to scale the experiment.
A developer has turned the iconic first‑person shooter DOOM into a machine‑learning transformer. The project, posted on Show HN, swaps NumPy’s linear algebra for the game’s own mechanics, leveraging DOOM’s Turing‑complete nature. The result is a proof‑of‑concept that can recognize handwritten digits, emit speech‑like sounds, and even steer a marine’s actions inside the game. The creator frames the work as an experimental alternative, acknowledging a massive speed penalty but emphasizing the novelty of using a game engine as an AI substrate. The code is publicly available on GitHub, and the author is actively seeking partners to push the experiment further.
What is DOOMFORMER?
DOOMFORMER is a self‑described transformer built entirely from DOOM game mechanics. The author notes that DOOM is Turing complete, allowing the game’s logic to replace conventional numerical libraries. The code lives on GitHub under the repository mottopanikeiku/doomformer. The project emerged from a Show HN post that frames the effort as an alternative to NumPy, albeit dramatically slower.
How does performance compare to traditional tools?
The creator admits the implementation runs roughly 65000x slower than NumPy. Despite the speed gap, the author argues the proof‑of‑concept demonstrates that a game engine can serve as a computational substrate. The project mentions that larger‑scale training could run on a measly T4, hinting at modest hardware requirements for future experiments.
What capabilities does DOOMFORMER claim?
According to the source, DOOMFORMER can classify MNIST digits, even when the digits are drawn on a DOOM map. It can produce rudimentary speech‑like output, described as “talk(kinda).” The system can also make decisions, enabling it to play DOOM itself. The author envisions controlling a marine’s mind on a map generated by the transformer, and notes that thousands of DOOM games could run simultaneously under this framework.
What are the next steps for the project?
The author explicitly invites collaborators to help scale the system and explore “crazier things.” The call to action centers on expanding training runs and refining the transformer’s capabilities. No formal roadmap is provided, but the tone suggests an open‑source, community‑driven evolution.
Frequently asked questions
Can DOOMFORMER classify handwritten digits?
Yes, the project can classify MNIST digits even when they are drawn on a DOOM map.
Is DOOM considered suitable for AI computation?
The creator points out that DOOM is Turing complete and uses its mechanics as a substitute for NumPy.
How can others help develop DOOMFORMER?
The author invites collaborators to scale the system, mentioning that larger training could run on a T4.
DOOMFORMER blurs the line between entertainment and artificial intelligence, showing that even a legacy game can host modern models. While the speed trade‑off is stark, the experiment opens a dialogue about unconventional computing platforms. If the community rallies, the next iteration could shrink the performance gap and unlock new ways to train AI on off‑beat hardware.
Sources: Hacker News