Mind, AI & Consciousness · Deep Dive #014

Hacking the Room that Claude Built

J-space, and what a language model keeps where it can reach it

On July 6, 2026, Anthropic published evidence that Claude maintains a small privileged set of internal representations sitting on top of a much larger volume of automatic processing. Nobody wired it in. It organized itself during training. The researchers call it the J-space. MAC spends two hours walking through it.

When
Thursday, September 24, 2026
6:30 PM to 8:30 PM PDT
Where
Downtown Vancouver. Exact address goes out to approved registrants.
Room size
20 people, prepared

Registration is approval-based and you need to be on the Mind, AI & Consciousness discussion list to attend. Not on it yet? Start at the MAC community page.

Cover artwork by Loki, redrawn here for the web.

The question on the table

If a model keeps a small set of thoughts where it can report on them, reason with them, and act on them, what have we actually found?

What they actually found

Three moves, and only the third one makes it a result.

  1. 01

    The J-lens

    For every word in the model’s vocabulary, the J-lens finds the internal activity pattern that pushes the model toward saying it. Point that lens at a live activation and you get a readout of what the model is leaning toward, including words it never says out loud.

  2. 02

    The J-space

    The readable directions are not scattered across the network. They cluster into one small region in the middle layers. The paper argues this region carries the representations available for report, for deliberate reasoning, and for flexible reuse, while grammar and fluent recall run automatically elsewhere.

  3. 03

    The edit

    This is the part that makes it a finding rather than a picture. The team removed one concept vector, added another of equal strength, and the model’s answer changed to match. The region is load-bearing, not decorative.

The intervention

Swap one vector. Watch the sentence move.

The paper removed a concept from the workspace and added another of equal strength. The model finished the sentence in the new frame. That is the difference between reading a model and steering it.

Workspace vector

The model is mid-sentence. The J-lens reads the workspace before a single word is emitted.

J-lens readout

  • pitch1.00
  • striker0.82
  • penalty0.71
  • offside0.58
  • keeper0.44

Outputthe ball is played forward to the striker at the edge of the box.

An illustration of the intervention described in the paper, using the soccer and rugby example the authors report. It is not a live model and the numbers are indicative, not measured. The real readouts are in the paper itself.

Pre-reading · 2.4 hr required

Come having read the two on the left.

MAC is a reading group before it is anything else. The room works because everyone arrives with the same text in their head. If you can only do one thing, do the Anthropic post.

Context, if you want the longer argument

  1. 03paper40 min

    Global Workspace Dynamics: Cortical “Binding and Propagation” Enables Conscious Contents

    Bernard J. Baars, Stan Franklin, Thomas Zoëga Ramsøy

    Baars originated global workspace theory in 1988. This is him stating the mature version. Worth reading so the room can judge how much of the theory the Anthropic result actually inherits.

    Frontiers in Psychology · 2013

  2. 04paper50 min

    Conscious Processing and the Global Neuronal Workspace Hypothesis

    George A. Mashour, Pieter Roelfsema, Jean-Pierre Changeux, Stanislas Dehaene

    The neuroscience version, with the neural evidence and the ignition dynamics. This is the standard the language-model claim is implicitly being measured against.

    Neuron · 2020

  3. 05paper1 hr

    Consciousness in Artificial Intelligence: Insights from the Science of Consciousness

    Patrick Butlin, Robert Long, and seventeen co-authors

    The report that turned theories of consciousness into a checklist of indicator properties for AI systems. A global workspace is on that list, which is exactly why the July result landed the way it did.

    arXiv · 2023

  4. 06paper1 hr

    On the Biology of a Large Language Model

    Anthropic Interpretability

    Method lineage. Same team, earlier tooling. Useful if you want to see how the interpretability programme got from tracing circuits to claiming a workspace.

    Transformer Circuits · 2025

Every link above was resolved and title-checked on 2026-08-19. This is the list for one room on one night. The wider trail, thirteen Deep Dives deep, is in the MAC Library and the community source index.

What the room will argue about

Nobody has to agree. Everybody has to have read.

  1. 01

    A global workspace was one of the named indicators in the 2023 “Consciousness in AI” report. One indicator is now evidenced. What does that change, and what does it plainly not change?

  2. 02

    Human global workspace theory assumes recurrence: contents are broadcast and held over time. A forward pass has one shot. Is J-space a workspace, or a still frame of one?

  3. 03

    The J-lens can only surface concepts that map to a single token. How much of the model’s inner life is invisible to the instrument by construction?

  4. 04

    If a system can report on its own privileged representations, is self-report now evidence of anything, or is it still just fluent text?

  5. 05

    Editing the workspace steers the model. That is an interpretability win and an alignment tool. Whose hand is on the lens?

How this works

Two hours, twenty people, one hard text.

Mind, AI & Consciousness started as an organic conversation in the Vancouver AI community and turned into a monthly habit. We use what is happening in AI as a lens on an older question: what consciousness actually is, and whether anything separates a human mind from a machine that acts like one. Thirteen Deep Dives are already in the archive. This is the fourteenth.