Tool directory · Gaming · Featured in Issue #2
DOOMFORMER
New this weekLast checked
Compiles an integer neural network into a DOOM map, where voodoo dolls do the arithmetic and light levels and floor heights hold the values.
Our Take
DOOMFORMER flips the old joke: instead of running DOOM on things, it runs things on DOOM. It's not a practical tool, and it doesn't pretend to be. The author reports that on all 10,000 MNIST test digits DOOM produced exactly the same numbers as the integer network (9,716 of them, 97.16%, classified correctly), with zero mismatches across 20,000 fuzzed networks. It's also very slow: the Show HN post says "65000x slower" than NumPy, and the README puts a small MLP at about 23,000 times slower. Worth a look for anyone who enjoys seeing how computation works when you strip it down to something absurd. Source
