—
- System
- Tools
tools
- User
- Expected
Loading
data/circuit_traces.json…Recorded outputs, side by side
Greedy, 128 tokens, identical prompts. Copied verbatim from the campaign records inruns/.The circuit, live
Exact closed-form readout of the edited layer-23 SwiGLU → final RMSNorm → tied unembedding at the decision position. Recomputed on every slider move.First-token distribution at the decision position
| token | logit | Δ vs original | p |
|---|
Neuron activity along the original model's generation
Layer-23 MLP input at every generated position of the original model, re-read through the two edited neurons with the current sliders. Bars: act3468 (teal) and act3421 (amber); line: the shared gate.Run both models in your browser
Real token generation with Transformers.js + ONNX Runtime Web on WebGPU, in a Web Worker. One download (fp16 decoder ≈1.5 GB plus 4-bit embeddings ≈150 MB, cached by your browser) serves both models: the edited model is produced in memory by writing the 6,142 edited values into the layer-23 MLP tensors of the downloaded original, and the SHA-256 of the result is checked against the published edited file.
Why fp16 and not a ~650 MB 4-bit build: the edit reads directions in the layer-23 residual stream that were fitted on the BF16 model. With a q4 backbone that input drifts (cosine ≈0.89 to the BF16 activations on the same prompts) and the abstain gate stays closed on explanation prompts, so the two neurons never fire; measured by scripts/export_webgpu_onnx.py --drift (report: webgpu_export_report.json).
Checking WebGPU…
The live runner uses a 4-bit (q4f16) ONNX export, not the campaign's BF16 MLX weights. Treat its outputs as a demonstration; the token-identical claims in the blog post refer to the recorded runs in Mode 1.