Public story · 2026-07-16 · high

Anthropic Ships Code That Reads Model Intent at Any Layer

It's the proof code for a new paper on how models form a shared internal workspace, released Apache 2.0 with a note that nobody's patches get merged.

Why now: The repo went live July 2 alongside the paper's posting on transformer-circuits.pub, framed from day one as a one-off rather than an ongoing project.

Issue

2026-07-16

Confidence

high

Sources

Redaction

passed

Story

Anthropic published jacobian-lens on GitHub July 2, the code behind a paper titled 'Verbalizable Representations Form a Global Workspace in Language Models.' For interpretability researchers, it's a shortcut: read out what a model already means at any internal layer, not just the last one. The repo has picked up 1,394 stars under an Apache 2.0 license.

The method computes one average Jacobian over a corpus, capturing how the final layer's output responds to changes at an earlier layer. That matrix then transports a residual-stream vector from layer l into the final-layer basis, where the model's own unembedding decodes it into words. In the repo's notation: lens_l(h) = unembed(J_l @ h).

Those examples run against Qwen, an open-weights model, since the method needs a full corpus of activations to compute that averaged Jacobian. The repo also ships interactive layer-by-position visualizations, so you can watch a representation's implied output shift as it moves through the network.

Its README marks the release explicitly as an unmaintained reference implementation, not accepting contributions.

Source trail

GitHub

Entities

GitHub Anthropic Releases Jacobian Lens Code Alongside 'Global Workspace' Interpretability Paper Anthropic jacobian-lens

Claim evidence

anthropics/jacobian-lens (Apache 2.0, created 2026-07-02, 1,394 stars) is the companion implementation for 'Verbalizable Representations Form a Global Workspace in Language Models,' published on transformer-circuits.pub. The lens reads out what an internal activation is disposed to make a model say by computing the average input-output Jacobian over a corpus, then linearly transporting residual-stream vectors at any layer into the final-layer basis and decoding through the unembedding: lens_l(h) = unembed(J_l @ h) where J_l = E[∂h_final/∂h_l]. It works on open-weights decoder transformers (examples use Qwen) and renders interactive layer × position visualizations. Anthropic marks it explicitly as an unmaintained reference implementation not accepting contributions.
Finding 16711 Source

Provenance

Canonical issue: 2026-07-16
AI generated: yes
Story unit: 16711
Labels: source-backed

Fetching from the wire…

← The Wire

Public story · 2026-07-16 · high

Anthropic Ships Code That Reads Model Intent at Any Layer

It's the proof code for a new paper on how models form a shared internal workspace, released Apache 2.0 with a note that nobody's patches get merged.

Why now: The repo went live July 2 alongside the paper's posting on transformer-circuits.pub, framed from day one as a one-off rather than an ongoing project.

Issue

2026-07-16

Confidence

high

Sources

Redaction

passed

Story

Its README marks the release explicitly as an unmaintained reference implementation, not accepting contributions.

Source trail

GitHub

Entities

GitHub Anthropic Releases Jacobian Lens Code Alongside 'Global Workspace' Interpretability Paper Anthropic jacobian-lens

Claim evidence

anthropics/jacobian-lens (Apache 2.0, created 2026-07-02, 1,394 stars) is the companion implementation for 'Verbalizable Representations Form a Global Workspace in Language Models,' published on transformer-circuits.pub. The lens reads out what an internal activation is disposed to make a model say by computing the average input-output Jacobian over a corpus, then linearly transporting residual-stream vectors at any layer into the final-layer basis and decoding through the unembedding: lens_l(h) = unembed(J_l @ h) where J_l = E[∂h_final/∂h_l]. It works on open-weights decoder transformers (examples use Qwen) and renders interactive layer × position visualizations. Anthropic marks it explicitly as an unmaintained reference implementation not accepting contributions.
Finding 16711 Source

Provenance

Canonical issue: 2026-07-16
AI generated: yes
Story unit: 16711
Labels: source-backed

Anthropic Ships Code That Reads Model Intent at Any Layer

Story

Related stories

Andrej Karpathy Joins Anthropic's Pre-Training Team

Claude Code's UX Won. Now People Are Swapping the Brain.

A Markdown File Got 44K Stars This Week. Skills Are the New Package Manager.

Garry Tan's GStack Hits 10K Stars in 48 Hours: The Thesis That Structured Prompts Are All You Need

RTK Just Hit 30K Stars Because Everyone's Token Bill Is Out of Control

Claude Code's Worst Week: 1,130 HN Points, AMD's AI Director Piles On, and a Quantitative Smoking Gun

A Local Knowledge Graph Just Cut My Claude Code Token Bill by 49x

Today's Skills

Source trail

Entities

Claim evidence

Provenance

Anthropic Ships Code That Reads Model Intent at Any Layer

Story

Related stories

Andrej Karpathy Joins Anthropic's Pre-Training Team

Claude Code's UX Won. Now People Are Swapping the Brain.

A Markdown File Got 44K Stars This Week. Skills Are the New Package Manager.

Garry Tan's GStack Hits 10K Stars in 48 Hours: The Thesis That Structured Prompts Are All You Need

RTK Just Hit 30K Stars Because Everyone's Token Bill Is Out of Control

Claude Code's Worst Week: 1,130 HN Points, AMD's AI Director Piles On, and a Quantitative Smoking Gun

A Local Knowledge Graph Just Cut My Claude Code Token Bill by 49x

Today's Skills

Source trail

Entities

Claim evidence

Provenance