weightspace / labRead the research ↗
AN EXPERIMENTAL PLAYGROUND

Small edits.
Different intelligence.

A neural network is a program written in numbers.
Step inside, change its weights, and watch its world change.

Open model experiments ↗
NO TRAINING LOOP. JUST WEIGHTS.
LIVE WORKSPACE / 001
Runs locally in your browser
01 / CHOOSE A BEHAVIOR

Program a boundary

A tiny network learns nothing here. You decide what its parameters compute.

02 / INTERVENE

The starting network contains a faulty output connection. Apply the hand-designed repair, or explore individual parameters below.

● Baseline loaded

Decision landscape

LIVE INFERENCE
INPUT YINPUT X
Class 0 Class 1Hover to inspect →
BEHAVIOR CHECK
Target agreement—
On a fixed 41 × 41 input grid
Weights changed0 / 17Compared with the loaded baseline
NETWORK
2 → 4 → 1
Inputs · ReLU units · Sigmoid
↳ Every pixel is an actual forward pass through these weights.
03 / READ & WRITE

The program is the parameters.

FORWARD PASSJavaScript
// No optimizer. No backpropagation.
const hidden = W.map((row, i) =>
  Math.max(0, dot(row, [x, y]) + b[i])
);

const probability = sigmoid(
  dot(v, hidden) + bias
);
Move a slider to edit a connection.
The landscape updates with each change.
REASON ABOUT THE WEIGHTS

Read. Repair. Write.

YOUR MODEL · INDEPENDENT EVALUATION

Run small-scale weight-space experiments using your own model. No optimizer, preset repair, or simulated AI response is used in this section.

Repair/write are scored against the target pattern selected above. Custom instructions do not change the scoring target. This architecture cannot express every possible behavior.

Model connection
EXACT MODEL INPUT

Read mode sends architecture and shuffled weights, without the answer. Repair sends a snapshot of your current weights. Write starts with all zeros. Test coordinates and expected answers are not included.

Prepare an experiment to begin.
EVIDENCE, NOT A DEMO SCORE

Results appear after evaluation. A proposal can fail, regress, or be rejected as invalid.

Scores use 1,024 pseudorandom inputs on [−1,1]², separate from the visible grid. A seed is saved in the run export for reproducibility. Reusing feedback is iterative testing, not an untouched benchmark.

Scope and limitations

These experiments use a 17-parameter classifier, not a language model checkpoint. Reading tests distinguish only the three built-in pattern families. No Qwen editing, game-playing environments, MATH-500 or IFEval benchmark, or replication of the paper’s numerical results is implemented. Models may produce incorrect explanations or parameters. Read the measured results, inspect the proposed weights, and compare multiple runs.

THE IDEA BEHIND THE LAB

What if weights were
something we could write?

Weight-space programming asks whether a language model can inspect, repair, or construct a neural network directly through its parameters.

This lab is an independent educational illustration. The original playground uses hand-crafted weights. The experiment workspace accepts real model proposals, but does not reproduce the article’s large-model experiments or benchmark results.

Explore the article by Rikiya Takehi & collaborators ↗