Small edits.
Visible intelligence.
Open the model. Turn a weight. Watch its behavior change.
Then prove the change before making it active.
A neural network you can
inspect, change, and undo.
Same world. Different decisions.
Better on which inputs?
All cases are public and repeatable. This is finite domain evidence, not proof of generalization or a private production gate.
See the whole sensor grid.
Each square is a different input pair. Lime means close to the reference. Coral reveals errors.
Closer FurtherNo candidate has been evaluated. Your session runs locally.
Checkpoint loading…Inspect the actual WASM activation trace +
Move either sensor to inspect this prediction.
Same model. Two WASM backends.
Scalar versus SIMD, measured here. A tiny network may spend more time crossing the JavaScript boundary than doing arithmetic.
Ready to measure.
256 predictions per batch · 3,000 measured batches
No assumed speedup.
Both backends must match the arithmetic reference.
Measurements include output allocation and the JavaScript/WASM call. They exclude model initialization and network transfer. Results vary by device and browser.
You edited the model itself.
Weights are numbers that shape a neural network’s decisions. Here you change two numbers directly, compare the candidate against defined expectations, and explicitly choose a revision.
Trace the prediction through real WASM.
Keep the candidate separate from active weights.
Reject shortcuts on separate cases.
Promote explicitly. Roll back exactly.