laserbrain

Gated Laplace diffusion kernel for token sequences.

Each position absorbs from its neighbours via exponential decay — close positions pull more than distant ones. A sigmoid gate controls how much blending occurs.

const { quadrize } = require('laserbrain')

const x = new Float32Array([1, 2, 3, 4, 5])
const out = quadrize(x)

API

quadrize(x, k, gate, d)

| param | type | default | description | |-------|------|---------|-------------| | x | Float32Array | — | input sequence, length ctx or ctx*d | | k | number | 1.0 | decay rate — higher k = sharper, more local | | gate | number | -1.4 | pre-sigmoid blend strength (~0.2 at default) | | d | number | 1 | embedding dimension (for 2D input, row-major) |

Returns a Float32Array of the same length.

quadrize.kernel

{ name: 'phronesis', k: 1.0 }

intuition

k high  →  sharp kernel, each token stays close to itself
k low   →  flat kernel, all tokens collapse toward the mean
gate    →  how strongly the blended signal overwrites the original

At default settings each position blends ~20% toward its neighbourhood.

2D (embeddings)

Pass flattened ctx × d row-major data with d set:

// 4 tokens, 3 dims each
const x = new Float32Array([
  1, 0, 0,
  0, 1, 0,
  0, 0, 1,
  1, 1, 0,
])
const out = quadrize(x, 1.0, -1.4, 3)

install

npm i laserbrain