Show HN: I Turned DOOM into a Transformer
Repository · https://github.com/mottopanikeiku/doomformer
First seen by Oriel on .
AI summary
Doomformer compiles integer neural networks into ZDoom maps, using voodoo dolls and sector geometry for arithmetic, then verifies every exposed value against NumPy with zero tolerance.
- The dolls perform multiply-adds, activations, comparisons, and argmax operations, while light levels and plane heights store intermediate sums and outputs.
- The MNIST demonstration computes 784 pixels through 64 hidden units, 45 comparisons, and 10 win counts, checking 248 exposed values exactly.
- Optimization reduces the digit-reading schedule from 5,705 game tics at --opt 0 to 273 tics at --opt 3, or eight seconds of real-time DOOM.
- The integer network reads 99 of the first 100 MNIST test digits correctly and achieves 97.16% accuracy across all 10,000 examples.
- The project uses ViZDoom with Freedoom assets and remains slow, while exactness validates executed values rather than untested inputs or overall classification correctness.
#zdoom #vizdoom #neural networks #mnist
Generated by gpt-5.6-sol. Check the original for details.
Metric trends
Last 30 days · Up to 120 recorded samples per source and metric, with reposts kept separate · First and latest retained
Points
Score reported by the source.
- Hacker News: 129, latest observation ; +125 over observed period since
Comments
Comment count reported by the source.
- Hacker News: 35, latest observation ; +35 over observed period since
Aliases
| Source | External ID | URL |
|---|---|---|
| Hacker News | 50187309 | https://github.com/mottopanikeiku/doomformer Read discussion |
| Hacker News | 50297697 | https://github.com/mottopanikeiku/doomformer Read discussion |