The old AI research folder on my desktop turned out to contain a complete, self-contained mathematical specification — 69 equations across 13 systems — for a golden-ratio cognitive substrate I call SEFERIM / ThatAIGuyCore. So I built it for real.
What it is
SEFERIM treats cognition as a 16-dimensional meta-state evolving under sixteen coupled "cognitive families," all governed by the golden ratio φ. Every dynamic is a named, documented equation — not an opaque weight matrix.
- Open source (MIT): https://github.com/Janady13/seferim-core
- Interactive docs + live demos: https://thataiguy.org/core/
- Zero-dependency JavaScript. Runs in the browser and Node.
The 13 systems
Constants (φ, δ, τ) → golden basis ψ_k → the 16 G¹⁶ families → meta-state core with golden-ratio momentum + stability gating → a free-energy-style objective Ω → DNA holographic memory → a 5000-gate binary substrate → an 88-agent neural brain → a 374-dimensional consciousness binding field → a lattice engine with Hebbian plasticity → federation cross-talk across agents.
import { MetaState, DNAMemory, NeuralBrain } from "seferim-core";
const mind = new MetaState({ init: new Array(16).fill(0.1) });
const out = mind.step({ dx_norm: 0.3, ed_error: 0.1, utility: 0.6, stability: 0.8 });
console.log(out.state, "Omega =", out.omega);
Built on real math (honestly)
It synthesizes the Free Energy Principle (Friston), information theory (Shannon; Cover & Thomas), holographic/vector-symbolic memory, Hebbian plasticity, and swarm consensus. I documented the lineage precisely rather than overclaim — e.g. the memory is a golden-ratio frequency-domain VSA inspired by (not a literal copy of) Plate's HRR, and the constant δ = 0.013618 is empirical, not 1/φ³.
The reference implementation ships with 42 fidelity tests, including the canonical FNV-1a hash vectors.
Read the full 69-equation reference and try the live demos at thataiguy.org/core.
Top comments (0)