Context Jamming · Research instrument

The Sliver and the Subspace

J-Space meets the Universal Weight Subspace

Static parameter compression (Kaushik’s Universal Weight Subspace) and dynamic functional workspace (Anthropic’s J-Space / Global Workspace) keep rediscovering the same geometric fact: the load-bearing structure of an intelligent system lives in a low-dimensional linear subspace vastly smaller than the ambient dimension. This page is not a victory lap. It is a falsifiable instrumentfor asking whether those two objects share a basis — evidence, if they do, for Architectural Determinism; a clean fracture, if they don’t.

Run the instrument Read the question

01 · The core question

Do the dynamic broadcast directions live inside the static weight subspace?

The Jacobian lens isolates dispositionally verbalizable representations — corpus-averaged linearized effects of activations on future token likelihoods. Their span is the J-space: selective, capacity-limited, reportable, used for multi-hop reasoning. Anthropic frames it as a global workspace. Most of the residual stream lives outside it.

Independently, spectral analysis of weight deviations across hundreds of finetunes recovers a shared low-rank basis — the Universal Weight Subspace. Task adaptation appears to play out in ~16 principal directions, largely independent of task, initialization, and data.

Neither paper projected one onto the other. That cross-space test is the payload of this instrument.

If J-space directions lie disproportionately inside the top-k universal weight directions, activation bottleneck and weight bottleneck are two faces of one geometric object. If orthogonal, the “all one manifold” story fractures at exactly the flagged seam. Either result is publishable — the mark of a real claim.

02 · The convergence

Five nodes, one bottleneck

Architectural Determinism does not claim gravity is cortex, or cortex is a transformer. It claims a compression theorem — multiply instantiated under different boundary conditions — and asks how far the isomorphism runs before it breaks.

01
Physicsprior
Holographic boundary
AdS/CFT · effective DOF scale with the boundary, not the volume
02
Biologyisomorphism
The Watt & the Manifold
Sparse coding + low-dim cortical manifolds under energy constraint
03
AI · Weightmeasured
Universal Weight Subspace
Kaushik et al. · ~k=16 shared directions across 1,100+ models
04
AI · Activationmeasured
Linear reps & manifolds
Park linear representations · Gurnee linebreaking / manifold geometry
05
AI · Workspacemeasured
J-Space (Global Workspace)
Anthropic Jacobian lens · ~25 active vectors · capacity-limited broadcast

Five nodes. One geometric fact: effective dimensionality ≪ ambient dimensionality. The open question is whether nodes 3 and 5 share a basis — not just a rhyme.

Schematic · ambient space → effective subspace

03 · Key correlations

What survives a hostile referee

Most false excitement comes from sliding between spaces. Name them. Keep them separate. Then test the bridge.

Static vs Dynamic

Distinct objects

Universal Weight Subspace
J-Space (Jacobian lens)

Kaushik measures frozen parametric geometry after training. Anthropic measures live, output-facing broadcast directions during inference. Same low-rank motif; different tensor domains.

Output-facing bridge

Load-bearing bridge

Unembedding / output map
J-lens + Park causal metric

Both the workspace construction and the privileged concept metric drink from output geometry. If the causal inner product is induced by unembedding covariance, “right metric for concepts” and “broadcast format” may be two views of one object.

Isomorphism ≠ identity

Epistemic flag

Structural rhyme
Causal identity

k≈16 (weights) and ~25 active vectors (workspace) are in the same ballpark. Matching integers across incommensurable spaces is numerology until directions are shown to correspond. Own the caveat.

Dimension	UWS (static)	J-Space (dynamic)
Domain	Weight / parameter	Activation / Jacobian
Method	HOSVD / stacked LoRAs	J-lens corpus average
Rank signal	~16 principal dirs	~25 active vectors
Capacity claim	PEFT works because of it	Workspace <10% residual var
Status	Empirically strong	Empirically strong
Cross-test	—	Unrun until this instrument

04 · The falsifiable instrument

Principal angles against a permutation null

Extract an orthonormal basis S for the Universal Weight Subspace (top-k left singular vectors of centered, stacked weight deviations). Extract concept / broadcast directions J via the Jacobian lens. Measure principal angles between the subspaces and the fraction of variance of each J-vector explained by S.

Then scramble: draw thousands of random orthonormal bases of the same rank, recompute overlap, and build a null distribution. A z-score above ~5σ under planted alignment means the instrument can detect structure when structure is there. On real models, the same machinery issues a hold / fracture verdict without assuming the answer.

Caveat: weight and activation directions may not share a coordinate frame. The browser mock plants a shared manifold by construction. Real runs need an explicit transport operator.

Live instrument · browser mock

Dial the alignment. Watch the verdict.

α = planted cosine · FVE ≈ α² · principal angles + permutation null

Alignment strength α0.70

0 · orthogonal1 · coplanar

σ confidence (z-score)16.5σ

marker = 5σ strong-evidence threshold (matches Python POC)

Principal angle spectrum (0°–90°)

0° alignedmean 47° · null ~83°90° orth.

HYPOTHESIS HOLDS

Strong geometric overlap. Dynamic broadcast directions map significantly into the static universal weight subspace under this planted alignment.

Readouts · synthetic d=512 · k=16

Mean principal angle: 47.0°; 0.820 rad
Mean FVE: 46.6%; J · proj onto UWS
Null mean angle: 82.8°; ±0.038 rad
p-value (approx): < 1e-6; one-sided vs null

FVE curve · expected ≈ α²

Python POC example · signal_strength=0.6

d=512 · k=16 · N_perm=1000
observed angle ≈ 0.412 rad · null ≈ 1.448
z ≈ 28.8σ → holds under planted manifold

Illustrative mock output shape — not a real-model measurement.

05 · Why this matters

Architectural Determinism, sharpened

Demis Hassabis has said the quiet part out loud: a lower-dimensional manifold may be true of most of reality. Wolfram talks of pockets of reducibility. Biology has been running a twenty-watt sparse coder for half a billion years — cortical manifolds that refuse to fill ambient dimension. Kaushik shows trained nets converge to a shared sliver of weight space. Anthropic shows reportable computation collapses into a capacity-limited workspace.

If those two AI results are faces of one geometry, the Architectural Determinism thesis gains a measurable joint — not an analogy between physics, brains, and models, but a shared rank structure with an experiment attached. If they are orthogonal, we learn something sharper: low rank is real, but the static and dynamic bottlenecks are different objects, and the unification story needs to be rewritten at the seam.

Either way, the science improves. The point of a falsifiable instrument is that you do not get to choose which result you like.

Hassabis

Lower-dimensional manifold — maybe true of most of reality.

Wolfram

Pockets of reducibility in a computationally irreducible world.

Biology

500 Myr of sparse coding under a hard watt budget.

Architectural Determinism →·arXiv Alpha: UWS →·TUCG →

06 · Citations & sources

Primary literature

[1]
The Universal Weight Subspace Hypothesis
Kaushik, Chaudhari, Vaidya, Chellappa, Yuille · arXiv 2512.05117 · Dec 2025
HOSVD across 1,100+ models; ~16 shared directions; PEFT compression.
[2]
Verbalizable Representations Form a Global Workspace in Language Models
Gurnee, Lindsey, et al. · Anthropic · Transformer Circuits · July 2026
Jacobian lens; J-space as capacity-limited broadcast workspace; ignition; ~25 active vectors.
[3]
The Linear Representation Hypothesis and the Geometry of LLMs
Park, Choe & Veitch · ICML 2024 · arXiv 2311.03658
Concepts as directions; causal inner product; embedding/unembedding unification.
[4]
When Models Manipulate Manifolds: The Geometry of a Counting Task
Gurnee, Ameisen, Kauvar, Tarng, Pearce, Olah, Batson · Transformer Circuits · 2025
Dynamic manifold geometry of computation in activation space.
[5]
Generative Models for Discovering Sparse Distributed Representations
Hinton & Ghahramani · 1997 · RGBN
Sparse distributed reps; explaining-away; generative MoE genealogy.

Epistemic note: quantitative details of the July 2026 workspace paper and related synthesis notes should be verified against primary sources before citation in formal work. This page is a research landing instrument, not a peer-reviewed claim that the cross-space overlap has been measured.

07 · Next experiments

From mock manifold to open models

The browser dial proves the instrument can see planted structure. The science starts when the manifold is no longer planted.

Transport operator

Define the map from weight tangent space → activation tangent space (layer Jacobian / NTK linearization / local pushforward). The mock plants a shared basis; real models need an explicit transport.

Project J onto UWS

Extract Kaushik-style UWS from open models (LoRA collections or multi-seed finetunes). Extract J-lens vectors. Measure FVE + principal angles vs permutation null.

Causal whitening

Optional Park-style whitening via unembedding precision matrix before overlap. Tests whether output anisotropy is the whole story.

Sweep k

Prior materials disagree on rank (~16 vs ~32). Treat k as a first-class parameter. Report the curve, not a single integer.

Run the real instrument on open models

If you have LoRA collections, multi-seed finetunes, or J-lens infrastructure — and want to co-run the cross-space test — open a line.

Contact Substack