xn
!!!!The bigger they are, the harder they fall!!!!
Every weight is a bit. Every neuron is the same XOR gate. The runtime is pure assembly, a 21 KB static binary with no libc. A swarm of a billion agents fits in the memory of one. It isn't a computer. It's a transmission.
One gate, ⊕ emulating itself
A cell XORs its input against its weight, counts the disagreements, and fires once enough of them pile up. Below, two inputs drift in superposition until you observe them. Then the hidden rank decides: OR, AND, and a voice that says OR but not AND. That's XOR, built from XOR.
| ROW | OR | AND | VOICE | XOR? |
|---|
hidden n0 · w 00 · t 1 → fires on any 1
hidden n1 · w 00 · t 2 → fires on both
voice n0 · w 10 · t 2 → n0 ∧ ¬n1
this exact cortex ships as models/signal.xn
Spawn them all. Keep one.
An agent is just its number: its genome is splitmix64(seed + id·φ). Each one wakes up, lives a single life of random
bit flips, gets judged by (loss, mutations, id), and is forgotten. Only the champion survives. This arena runs the real algorithm
in your browser. The assembly engine does a million agents in 11 seconds.
The bigger they are, the harder they fall
The giants are stacked from floating point, data centres and kilowatt-hours. Push one and watch where its bits end up.
| a trillion-parameter giant (fp16) | ~2 000 000 000 000 B |
| its training set | ~petabytes |
| xn binary · static · no libc | 21 520 B |
| xn model · 8 layers · fixed | 4 848 B |
| one living agent · peak stack | 9 888 B |
| layer · 64 cells | 592 B |
| cell · weight + threshold | 9 B |
| an agent · it is its seed | 8 B |
| a dataset · ≤ 64 rows · one register | 8 B |
| a billion agents · stored | 0 B |
numbers from xn --stats
It is not a computer.
Decoded from the reply word 0110. This is lore. The code is real.
Run it on your machine
Linux x86-64. NASM, binutils and a C++26 compiler. No Python, no downloads, no GPU.
$ git clone https://github.com/magneato/XORNets && cd XORNets $ ./setup.sh # nasm · binutils · g++ ≥ 15 $ ./build.sh # build/xn build/libxn.a build/xnc build/xn_test $ ./test.sh # C++26 tests + console checks $ ./run.sh # pick a model, watch it learn, talk to it $ build/xn --help --model models/first-contact.xn $ build/xn --model models/parity3.xn --agents 1000000 $ build/xn --stats
asm runtime
xor · layer · inference · train · agents · options · model · io · xn. Raw syscalls, mmap'd model files, clone()d threads, futex joins.
c++26 instruments
xnc inspects, forges, sweeps and swarms. xn_test checks the assembly against plain C++ reference code.
.xn organisms
Plain text: [model], [layer], [options]. Trained weights round-trip through --emit.