Run by the Opus hand on neo (Apple A18 Pro, 8 GB, CPU only, 2 torch threads),
2026-09-21, against the preregistration at
~/arianna/_notes/PREREG_toaster_duel_netta_vs_nanogpt_2026-09-21.md.
Everything below was measured on this machine by this hand, with one exception that is labelled where it appears: the netta.c row is Don's measurement, carried here, not re-run by me.
| import json | |
| import pickle | |
| import struct | |
| import zipfile | |
| import numpy as np | |
| from sentencepiece import SentencePieceProcessor | |
| def rms_norm(x): return (x / np.sqrt(np.square(x).mean(-1, keepdims=True) + 1e-6)) | |
| def softmax(x): return (np.exp(x - np.max(x, axis=-1, keepdims=True))) / np.sum((np.exp(x - np.max(x, axis=-1, keepdims=True))), axis=-1, keepdims = True) |
| /* | |
| * infer_nanodurov.c — Interactive chat with nanodurov (BPE 15.7M on notorch) | |
| * | |
| * Build: make infer_nanodurov | |
| * Run: ./infer_nanodurov [weights.bin] [merges.txt] | |
| * | |
| * Default: nanodurov_arianna.bin + arianna_bpe_merges.txt | |
| */ | |
| #include "notorch.h" |
| /* | |
| * postgpt.c — zero-dependency BPE transformer with metaweights. | |
| * | |
| * C port of postgpt.py. Same algorithm, same resonance. | |
| * Dual attention: Content (QK^T) + RRPRAM (x @ Wr). | |
| * Metaweights: statistical probability space from BPE tokenization. | |
| * | |
| * Compile: gcc -O2 -o postgpt postgpt.c -lm | |
| * Run: ./postgpt | |
| * |
| #!/usr/bin/env python3 | |
| """ | |
| microreasoning.py — Resonance engine. 1984 words. Dario Equation. | |
| 8-layer sequential transformer with multi-head attention, RoPE, | |
| RRPRAM resonance gates, and SwiGLU FFN. Dual tokenizer: | |
| BPE input, word-level output. | |
| Architecture per layer l: | |
| h = rmsnorm(x, attn_norm_l) |
| /* | |
| * neoleo.c -- Language Emergent Organism (single-file edition) | |
| * | |
| * Complete autonomous digital organism in one C file. | |
| * D.N.A. from mini-arianna. Arianna -> Leo. Mother -> Son. | |
| * | |
| * Build: cc neoleo.c -O2 -lm -lsqlite3 -lpthread -o neoleo | |
| * Run: ./neoleo | |
| */ |
| /* | |
| * lee.c v7 — Vision-Language Model in pure C | |
| * | |
| * Named after Bruce Lee (the only man who beat Chuck Norris) | |
| * and Minhyeok Lee (whose self-identity framework gives Chuck his soul). | |
| * | |
| * Sees images. Speaks words. Adds numbers. Zero dependencies. | |
| * Tape-based autograd with arena bump allocator. | |
| * | |
| * Architecture: |
| //go:build ignore | |
| /* | |
| * molequla.c | |
| * A dependency-free, single-file, continually-learning GPT organism in pure C. | |
| * | |
| * Compile: gcc -O2 -o molequla molequla.c -lsqlite3 -lpthread -lm | |
| * With BLAS: gcc -O2 -DUSE_BLAS -o molequla molequla.c -lsqlite3 -lpthread -lm -lopenblas | |
| * macOS: gcc -O2 -DUSE_BLAS -o molequla molequla.c -lsqlite3 -lpthread -lm -framework Accelerate | |
| * |
| #!/usr/bin/env python3 | |
| # -*- coding: utf-8 -*- | |
| """ | |
| molequla.py | |
| A single-file, async, continually-learning GPT organism. One dependency: numpy. | |
| - Trains on nonames.txt (one sentence per line) | |
| - Keeps SQLite memory (tiny chat loop) | |
| - Maintains a bounded corpus reservoir (never bloats) |
| <!DOCTYPE html> | |
| <html lang="en"> | |
| <head> | |
| <meta charset="UTF-8"> | |
| <meta name="viewport" content="width=device-width, initial-scale=1.0"> | |
| <title>molequla.js — a GPT organism in your browser</title> | |
| </head> | |
| <body> | |
| <!-- | |
| molequla.html |