J‑Space Through the Lens of LLM Space Theory

Why Anthropic’s discovery fits naturally into a geometric model of LLM cognition

By: Gary Drypen

Anthropic’s recent “J‑space” announcement has triggered a wave of excitement – and confusion – across the AI research community. Many people interpreted J‑space as a kind of workspace, scratchpad, or even a glimpse of access consciousness inside Claude. The language of “inner thoughts,” “reportable concepts,” and “global workspace” fueled that impression. But the underlying phenomenon is far simpler – and far more interesting – when viewed through the geometric framework of LLM Space theory.

This article explains:                                  

• what J‑space actually is
• why it is not a workspace
• how LLM Space theory predicts its existence
• how worldlines, curvature, and torsion explain J‑space behavior
• why this connection matters for interpretability and safety

1. The core misconception: J‑space is not a workspace

The most common misunderstanding is treating J‑space as a literal “workspace” the model thinks inside …  a dynamic module where concepts are stored, manipulated, or summoned. Nothing like that exists inside a transformer.

A transformer has:

• no modules for “thoughts”
• no buffers for “inner speech”
• no scratchpads
• no dynamic memory
• no internal narrative

It has one thing: a fixed manifold — the high‑dimensional space defined by its trained weights. Everything the model does is a trajectory through that manifold. J‑space is simply a static linear subspace of that manifold. It is not a place the model enters. It is not a container for thoughts. It is not a workspace. It is not a module. It is not a mechanism.  It is aset of directions inside the activation vector.

Every transformer has a J‑space — a static set of directions in the activation manifold that reliably influence verbal output. Larger models simply have a richer and more interpretable J‑space.  The BERT model used in the curvature/torsion plotting experiment (see plots below) also  has a J‑space, but its structure and dimensionality differ from Claude’s.

2. What J‑space actually is

J‑space is a static linear subspace of the model’s activation space (LLM Space). The residual stream is the moving activation vector inside that space. At each layer, the model projects the residual stream onto J‑space, producing the J‑space content.

Anthropic used probes to identify directions in the activation vector that correlate with concepts the model can verbalize – concepts that reliably influence the logits in human‑interpretable ways.

Examples include:

• “spider”
• “ant”
• “danger”
• “self‑reference”
• “planning intent”

These directions form J‑space.

J‑space is:

• static (fixed after training)
• geometric (a subspace of the residual stream vector space)
• linguistically aligned (directions that affect verbal output)
• not exhaustive (most directions encode non‑linguistic features)

This last point is crucial.

Most activation directions encode internal features the model cannot directly express:

• token boundary markers
• attention routing patterns
• formatting structure
• logical consistency signals
• confidence gradients
• latent cluster membership

These are essential for reasoning, but they are not verbalizable.  J‑space is the subset of directions that are.

3. How LLM Space theory explains J‑space

LLM Space theory proposes:

1. The manifold is fixed after training.
2. Inference is a worldline — a trajectory through that manifold.
3. Each transformer layer produces a new activation, a new point on the worldline.
4. Curvature and torsion of the worldline reflect reasoning dynamics.
5. Attractors are stable geometric basins in the manifold.
6. Interpretability is geometric, not symbolic.

Under this framework, J‑space is not surprising. It is expected.

Why?

Because in a fixed manifold, some directions will naturally:

• align with linguistic concepts
• influence the logits
• be detectable by probes
• be steerable

These directions form a linguistic subspace –  J‑space.

LLM Space theory predicts that:

• J‑space is static
• J‑space is geometric
• J‑space is part of the manifold
• J‑space content changes layer‑by‑layer
• J‑space curvature reflects reasoning
• J‑space attractors correspond to stable conceptual basins

Anthropic’s findings fit naturally into this model.

4. J‑space content changes because the worldline moves

A transformer does not “think” in one place. Reasoning emerges from distributed transformations across layers.

At each layer:

h(L)→h(L+1)

The activation vector changes. Therefore, its projection onto J‑space changes. This projection is what Anthropic calls “J‑space content.” It is not stored, it is not persistent and it is not a workspace. It is simply:

the coordinates of the worldline along the J‑space axes at that layer.

When the model resolves ambiguity, makes a decision, or shifts direction, you see:

• curvature in the worldline
• torsion in the representational frame
• changes in J‑space content

This is what LLM Space theory predicts.

5. Torsion: how the model’s representational frame twists across layers

Curvature tells us how the worldline bends through the manifold , that is, how the model’s internal state moves as it shifts concepts or transitions between attractors. But curvature alone does not capture the full geometry of inference. There is a second, deeper geometric signal that emerges only when we examine how the model’s representational frame changes across layers. That signal is torsion. Torsion is not a property of the manifold itself. It is a property of the frame that moves along the worldline.

In classical differential geometry, torsion measures how the local coordinate frame twists as you move along a curve. In LLM Space, torsion measures how the model’s internal representational axes rotate from one layer to the next. This rotation is induced by the transformer’s layer‑to‑layer transformations –  attention mixing, MLP transformations, residual additions, and layer norm scaling – all of which subtly reorient the residual stream basis. The result is a geometric degree of freedom that curvature cannot capture: the twisting of meaning.

Why torsion matters

Torsion reveals how the model changes how it represents concepts, not just which concepts it moves toward.  A worldline may show low curvature (little bending) while torsion increases sharply. This indicates that the model is not shifting concepts, but is rotating the representational frame – recontextualizing the same concept in a new semantic orientation.

This is the geometric signature of:

• reframing
• reinterpretation
• contextual pivoting
• semantic axis rotation
• “twisting” of meaning rather than movement

How torsion relates to J‑space

J‑space is static. Its axes do not move. But the worldline’s orientation relative to J‑space does change. This is where torsion becomes essential. Even if the worldline stays near the same conceptual attractor, torsion can rotate the representational frame so that the projection onto J‑space changes. This explains why J‑space content can shift even when the underlying concept is stable.

In other words:

• Curvature explains movement of concepts through the manifold.
• Torsion explains rotation of the representational frame attached to the worldline.
• Together, they explain why J‑space content evolves across layers.

Empirical evidence: torsion from non‑aligned PCA frames

Torsion is not just a theoretical construct — it is empirically measurable. When we compute PCA locally at each layer, the principal axes (PC1, PC2, PC3) do not align across layers. Instead, they rotate. This rotation of the local PCA frame is the empirical signature of torsion.

In other words:

Curvature bends the worldline; torsion twists the coordinate frame attached to it.

This is why torsion does not show up in the manifold surface itself — it is not a twist of the curve, but a twist of the coordinate system attached to the curve.

Token‑dependent behavior

Empirical worldline geometry shows that curvature and torsion are token‑dependent and layer‑dependent. They do not follow a single universal pattern. Instead, they exhibit characteristic behaviors:

• Token 1 (early layers): torsion is low; curvature spikes as the model begins contextualization
• Token 6 (middle layers): curvature decreases; torsion becomes more pronounced as the representational frame rotates
• Token 11 (later layers): torsion peaks; curvature fluctuates depending on the token’s role

These patterns show that torsion is a real geometric signal reflecting how the model rotates its representational frame as it moves from shallow contextualization to deeper semantic integration and finally to output‑aligned representation.

Worldline Geometry — BERT Token 1 (“the”) This diagram shows the empirically measured worldline, curvature, and torsion for the token “the” in a BERT model. These plots were generated from real residual‑stream activations collected layer‑by‑layer using our worldline probe pipeline.

The 3D worldline reveals early contextualization dynamics, with curvature spikes indicating moments where the model shifts semantic interpretation. Torsion remains low in the earliest layers but rises sharply in later layers, demonstrating that even simple function words undergo representational frame rotation as deeper layers integrate broader context.

This plot illustrates that curvature and torsion are token‑dependent and layer‑dependent, even for high‑frequency function words.

Worldline Geometry — BERT Token 6 (“over”) This diagram reflects actual experimental measurements of curvature and torsion for the token “over,” obtained by running the worldline probe pipeline on BERT and computing geometric properties directly from residual‑stream activations.

The curvature plot shows a pronounced peak in mid‑stack layers, followed by a decline as the model settles into deeper semantic integration. Torsion fluctuates across layers, indicating ongoing rotation of the representational frame as the model recontextualizes the token within the sentence structure.

This example demonstrates that middle‑layer geometry often exhibits decreasing curvature and increasing torsion variability, consistent with deeper contextual processing.

Worldline Geometry — BERT Token 11 (“[SEP]”) This diagram presents real curvature and torsion measurements for the “[SEP]” token, derived from empirical worldline analysis of BERT’s residual‑stream activations.

Curvature fluctuates across layers, reflecting the model’s shifting structural interpretation of the sequence boundary. Torsion shows a sharp peak in later layers, indicating a strong rotation of the representational frame as the model prepares the final contextualized embedding for downstream tasks.

This plot highlights how later‑layer geometry often exhibits torsion peaks, even for structural tokens, as the model aligns its internal representation with output‑stage semantics.

6. Attractors explain J‑space stability

LLM Space theory proposes that attractors are inherent in the fixed manifold.  Anthropic observed that certain J‑space directions behave like:

• stable concepts
• persistent interpretations
• recurring patterns
• “thoughts” the model can report

These are not dynamic constructs. They are attractor basins in the manifold.  The worldline falls into them. The J‑space projection reveals them. Probes detect them. Steering vectors can shift them. This is geometric, not cognitive.

7. Why this matters for interpretability and safety

The connection between LLM Space and J‑space provides A geometric foundation for interpretability.  J‑space is not a mystery. It is a geometric feature of the manifold.

A way to detect dangerous attractors

If harmful concepts correspond to stable J‑space directions, they can be:

• mapped
• monitored
• steered
• suppressed

A path toward “savant” oversight models

A specialized model could:

• track worldline curvature
• monitor J‑space projections
• measure torsion
• detect dangerous attractors
• apply micro‑steering
• keep the wrapper safe

This is a geometric approach to safety, not a behavioral one.

8. The synthesis

Anthropic discovered a linguistic subspace. LLM Space theory explains why it exists. Together, they form a coherent geometric model of LLM cognition. The key insight:

J‑space is not a workspace. It is the set of directions in the fixed manifold that the worldline can express in language.

Curvature explains how the worldline moves. Torsion explains how the representational frame twists. J‑space explains how those dynamics become verbalizable.

This reframes the entire conversation around J‑space and positions LLM Space theory as the conceptual architecture that makes sense of it.

9. Suggested experiment: geometric signature of a J‑space intervention

Here is a high‑level protocol a lab could run to test whether a J‑space concept swap (e.g., spider→ant) produces a detectable change in worldline curvature and torsion:

1. Baseline run: probe + worldline geometry
   1. Prompt: Use a prompt that reliably elicits the “spider” concept in the target model.
   1. Probe runner: Run Anthropic‑style probes (or equivalent) to identify the “spider” direction in J‑space.
   1. Worldline capture: Record the residual stream activations across layers for the token(s) of interest.
   1. Geometry: Compute and plot:
      1. worldline in PCA space (PC1–PC3)
      1. curvature vs layer
      1. torsion vs layer
   1. This is the baseline geometry for the unmodified “spider” concept.
2. Intervention setup: choose a layer and direction
   1. Select a layer index in the model’s J‑space‑active band (e.g., mid‑stack).
   1. Identify the “spider” and “ant” directions in J‑space at that layer via probes.
   1. Define an activation edit that reduces the spider coordinate and increases the ant coordinate at that layer.
3. Intervention run: apply the J‑space edit
   1. Re‑run the same prompt.
   1. At the chosen layer:
      1. pause the forward pass (conceptually—implementation depends on the framework – some frameworks allow direct activation editing; others require custom hooks),
      1. apply the spider→ant activation change in J‑space,
      1. resume the forward pass from that modified activation.
   1. Again record the residual stream activations across layers for the same token(s).
4. Geometry after intervention
   1. Compute and plot:
      1. worldline in PCA space (PC1–PC3)
      1. curvature vs layer
      1. torsion vs layer
   1. This is the post‑intervention geometry.
5. Compare baseline vs intervention
   1. Look for:
      1. a local change in curvature at or shortly after the intervention layer (worldline bending toward “ant” instead of “spider”),
      1. a change in torsion (rotation of the local PCA frame) around the same region,
      1. a shift in J‑space projection from spider‑aligned to ant‑aligned directions.

If LLM Space theory is correct, the J‑space edit should leave the global basin (e.g., “bug/insect/arthropod”) intact but produce a detectable geometric signature—a small bend and twist—at the intervention layer.