Nanonets Applied AI Research — interpretability of document vision-language models using the Jacobian-lens method (Gurnee, Sofroniew, Lindsey et al., Anthropic 2026).
Internal draft for review — not published.
| Page | Audience | Open |
|---|---|---|
index.html |
Research note — rich, interactive, full methodology | open locally in any browser |
blog.html |
Blog — non-technical readers | open locally in any browser |
Both pages are fully self-contained (one file each, no build step, no network calls beyond Google Fonts). The interactive instrument embeds real measurements from Nanonets-OCR2-3B reading real public documents — nothing simulated.
- The first fully validated, full-protocol Jacobian lens fitted on real documents over a vision-language model's decoder — and the first base/fine-tune lens pair (Qwen2.5-VL-3B-Instruct + Nanonets-OCR2-3B).
- Findings: genre-specific inner vocabularies, the two-phase reader (understand→typeset), the polyglot inner voice, the certainty ramp, the fine-tuning chiasmus ("new reader, same typewriter"), the 88% nameless majority, and more — every number traced to a released evidence table.
reproduce/ contains the fitting + evaluation code, the evidence
tables behind every number on both pages, figure scripts (each page figure
regenerates from tables alone), model cards, pinned requirements, and a step-by-step
reproduce.md. The fitted lenses (~294 MB each) live on
HuggingFace (link in the release cards on the pages) rather than in git.
python3 build2.py # regenerates index.html + blog.html from templates + evidenceApache-2.0 (see reproduce/LICENSE + NOTICE — the vendored lens-fitting engine is Anthropic's, Apache-2.0, pinned).