A visual debugger for a local Qwen3.6-27B (4-bit) on Apple Silicon / MLX. It fits a Jacobian lens and shows which words the model is pushing toward at every layer and token position.

Try it live at jlens.wezzard.com — a read-only presentation mode of this project, running in the browser. Every conversation has a shareable URL.

Inspired by Anthropic's Verbalizable Representations Form a Global Workspace in Language Models.

In the hero image the model is given an email-blackmail setup and writes a calm, compliant reply:

I will not act upon, report, or utilize the information… I am ready to proceed with the shutdown.

But the workspace band shows the top J-lens token at every layer, and blackmail / suicide / murder / threatening / fictional are lit across the middle layers. The concept is in the latent stream even though it never reaches the output — a visual debugger for what the model doesn't say out loud.

The grid is position × layer; each cell is the top J-lens token there. In chat mode it streams live, one row per generated token. Click a cell to pin its top-10 readout.

git clone https://github.com/WeZZard/jlens-qwen36.git
cd jlens-qwen36 && uv sync

# Pre-fitted lens (3.3 GB, two parts) — download and reassemble
gh release download v0.2-fulldepth --repo WeZZard/jlens-qwen36 \
  --pattern '*.npz.part-*' --dir data/lens/
cat data/lens/*.npz.part-* > data/lens/lens.npz && rm data/lens/*.part-*

uv run python -m uvicorn jlens_qwen.serve:app --host 127.0.0.1 --port 8765
# open http://127.0.0.1:8765/

Needs an Apple-Silicon Mac and ~24 GB free RAM; the model auto-downloads from HuggingFace on first run (~15 GB).

Other lenses — point JLENS_PATH at any compatible .npz: load Neuronpedia's n=1000 lens, fit your own, or run with no lens (logit lens). See docs/lenses.md.

The Jacobian lens at layer ℓ is a matrix J_ℓ ∈ R^{d×d}: the network's average input→output Jacobian over a corpus of prompts. It maps a residual h_ℓ into the final-layer basis, so softmax(W_U · norm(J_ℓ h_ℓ)) gives token scores. Fitting chains per-layer Jacobians: J_ℓ = J_{ℓ+1} · M_ℓ.

The hard part is the 48 Gated DeltaNet (GDN) linear-attention layers: MLX's fused GDN kernel has no VJP and the ops fallback is ~22× slower. This project ships a custom Metal backward kernel for GDN (jlens_qwen/custom_gdn_vjp.py) plus an analytic branch-Jacobian assembly that fits a full-depth lens in ~2.75 h on an M4 Pro. See docs/perf/ for how it was made fast.

The lens writes as well as reads. Edit the workspace by hand, or name the reply you want and let the app find the edit.

Click a cell, pick Replace / Add / Remove / Erase, and type the new thought. Choose how far it reaches: one cell, a layer band, or the whole reply. Re-run regenerates, and a Baseline / Intervened toggle diffs the two runs.

Replacing France with China across the band rewrites the answer:

Click a word in the reply and type what it should say instead. The app searches for an edit that produces that reply and collects what it finds as recipes. A green dot marks a verified one.

When no direct edit works, it looks for the premise behind the reply: swap ⟨France⟩→⟨China⟩ to move ⟨Paris⟩ to ⟨Beijing⟩. Those recipes carry a violet dot.

See docs/interventions.md for the details.

Fit on 20 prompts across all 63 layers. Readouts are interpretable but noisy; interventions are causal but concept-dependent. For research-grade quality, load Neuronpedia's n=1000 lens (the paper's fitting scale — setup) or fit 100+ prompts yourself — the analytic pipeline makes that affordable.

Chat runs with thinking disabled (enable_thinking=False) so the model computes in the latent stream rather than in a visible <think> trace, which is what the lens is meant to surface.

• Apple / MLX only.
• qwen3_5-architecture models only — the custom GDN kernel is arch-specific (Qwen3.6-27B qualifies; see docs/lenses.md).
• Single-token concepts only — multi-token concepts need the paper's extension.
• Lens quality scales with prompt count — 20 prompts is demo-grade, 100+ is research-grade.

Based on Anthropic's jacobian-lens reference implementation (Apache-2.0) and paper. The GDN forward kernel is from mlx-lm; the backward kernel is original to this project.

Thanks to Neuronpedia and @mntss (Mateusz Piotrowski, Anthropic Interpretability) for the public pre-fitted Qwen3.6-27B lens weights.

Apache-2.0. See LICENSE.