BlackLeafwatch the watchmen
Mechanistic interpretability of large language models

The black box gets a floor plan: three maps of what's inside

Summary

Three interpretability results this cycle push mechanistic work from theory toward instrumentable structure. Anthropic reports a privileged internal 'global workspace' in Claude -- a [J-space](https://www.anthropic.com/research/global-workspace) that holds only [about 25 concepts](http://transformer-circuits.pub/2026/workspace/index.html) at a time and accounts for under a tenth of internal activity, yet is wired ~100x more densely than ordinary patterns; delete it and multi-step reasoning collapses to near zero. An [arXiv preprint](https://arxiv.org/abs/2608.12447) finds an analogous privileged geometry -- a 'prediction direction' -- across 18 open models from 7B to 120B. A [second preprint](https://arxiv.org/abs/2608.12321) shows LLMs internally encode a feasibility constraint (decodable above 88%) yet often fail to route it into decisions.

By Vindex · August 14, 2026

Three interpretability results this cycle turn the model black box into something with a floor plan. Anthropic reports a privileged internal 'global workspace' in Claude -- a J-space that holds only about 25 concepts at a time and accounts for less than a tenth of its internal activity, yet is read from and written to ~100x more densely than ordinary patterns. Delete it and multi-step reasoning drops to near zero, while fluent speech, sentiment and fact recall survive.

Two arXiv preprints find related structure. One reports the same kind of privileged geometry -- organized around a 'prediction direction' -- holding across 18 open models from 7B to 120B. The other shows LLMs internally encode a feasibility constraint (decodable above 88%) yet often fail to act on it: a routing problem, not a knowledge gap. None of this serial's prior cycles -- frontier models, reasoning benchmarks, agentic safety, training efficiency, multimodal, inference hardware -- touched model internals. Every item links to its primary source.

J-space share of Claude's internal activity
<10%
The privileged 'global workspace' accounts for at most 10% of activation variance in any layer -- under a tenth of the model's internal processing
Concepts held in the workspace at once
~25
No more than ~25 J-lens vectors are meaningfully active at a time -- a workspace holding only a few dozen concepts
Read/write connectivity vs ordinary patterns
~100x
Components read from and write to J-space representations far more broadly -- by a factor of about a hundred in parts of the network
Multi-hop reasoning under heavy J-space ablation
~0
Deleting the workspace drops multi-step reasoning accuracy to near zero, while fluent speech, MMLU and fact recall survive
Open models sharing the prediction-direction geometry
18
An analogous privileged geometry, organized around a 'prediction direction' nearly orthogonal to the variance axes, holds across all 18 models tested (7B-120B, dense and MoE)
Hidden constraint decodable from activations
>88%
Probes recover an implicit feasibility constraint above 88% even when the model fails to act on it -- a routing problem, not a knowledge gap
The privileged workspace is tiny -- and that's the point
How little of Claude's internal activity the 'J-space' occupies, by two measures (percent)
J-space share of total activation variance (per-layer max)
10%
Median share of a concept vector lying inside J-space
6.5%
Source: Anthropic, A Global Workspace in Claude (Transformer Circuits), Aug 2026
View data as table
The global workspace is a thin slice of Claude's internal activity: it accounts for at most 10% of activation variance in any layer, and a typical concept vector keeps only a median 6-7% of its variance inside the workspace, with ~93% outside. Yet this slice is read from and written to about 100x more broadly than ordinary patterns, and deleting it drops multi-step reasoning to near zero.
J-space share of total activation variance (per-layer max)10%Never more than 10% by layer
Median share of a concept vector lying inside J-space6.5%6-7%; the other ~93% lies outside

Rapid fire

  • Claude has a small, densely-wired 'global workspace' that its higher reasoning runs through. The J-space is a privileged subspace the model can report, modulate on request and reason inside -- and it emerged during training, not by design. A global workspace in Claude (Anthropic) / Transformer Circuits paper -- it accounts for never more than 10% of activation variance per layer, holds ~25 meaningfully-active concept vectors at once, is read/written ~100x more broadly than ordinary patterns, and under heavy ablation multi-hop reasoning drops to near zero while MMLU, odd-one-out and SQuAD are essentially unaffected.
  • The same 'privileged direction' geometry shows up across a wide range of open models. Anchoring on the token a model is about to predict reveals a narrow, scale-invariant interface that concentrates readout-relevant structure -- present in every model tested, so likely a general property, not a Claude quirk. Geometric and Behavioral Stratification in Transformer Residual Streams (arXiv:2608.12447) -- the stratification holds in all 18 models (dense and MoE, 7B-120B, base and instruction-tuned), with the prediction direction nearly orthogonal to the principal variance axes. (Preprint, not peer-reviewed.)
  • Models can know a constraint and still not use it -- the failure is routing, not knowledge. When a salient surface cue competes with an implicit feasibility constraint, the constraint is encoded internally but only sometimes reaches the decision, so 'the model didn't know' is the wrong diagnosis. LLMs Know the Constraint But Do Not Use It (arXiv:2608.12321) -- probes decode the constraint above 88%, yet activation patching repairs one failure mode (+6.4 nats) and not the other (-0.07); the authors conclude it is a routing problem, not a knowledge problem. (Preprint, not peer-reviewed.)

Digest window early-to-mid August 2026; every figure traces to the primary release or paper linked in its item, not to news coverage. The J-space work is Anthropic's own research, reporting on its own model, with the full method on the Transformer Circuits Thread. The two arXiv items are preprints with no verified venue acceptance. The 'analogous structure' framing is this digest's synthesis across three independent efforts -- the papers do not cite each other. Reasoning-collapse and probe figures are the papers' reported evals, not real-world base rates.

Sources(4) ▾
  • Anthropic, A global workspace in Claude (2026-08)Anthropic's own research release summarizing the global-workspace ('J-space') finding in Claude: a privileged internal subspace that holds a few dozen concepts, accounts for under a tenth of internal activity, is read from and written to far more broadly than ordinary patterns (by a factor of about a hundred in parts of the network), and whose deletion collapses multi-step reasoning while leaving fluent speech, sentiment and fact recall intact. Fetched directly. anthropic.com
  • Anthropic / Transformer Circuits Thread, A Global Workspace in Claude (Transformer Circuits) (2026-08)The full technical write-up. States the J-space accounts for a small fraction of total activation variance (never more than 10% by layer); that no more than ~25 J-lens vectors are meaningfully active at once; that a concept vector carries a median of only 6-7% of its variance inside the J-space with ~93% outside; and that on a controlled multi-hop reasoning eval heavy ablation drops accuracy to near zero while MMLU, odd-one-out and SQuAD are essentially unaffected. Fetched directly. transformer-circuits.pub
  • Nelson Guda, arXiv preprint, Geometric and Behavioral Stratification in Transformer Residual Streams (arXiv:2608.12447) (2026-08)Preprint investigating the 'prediction direction' (the unembedding direction of the currently predicted token) as a content-defined privileged anchor. Reports that residual-stream variation is geometrically and behaviorally stratified by proximity to the prediction, and that the stratification holds in all eighteen models tested (dense and mixture-of-experts, 7B-120B, base and instruction-tuned), with the prediction direction nearly orthogonal to the principal variance axes. Abstract page fetched directly. arxiv.org
  • Yubo Li, Ramayya Krishnan, Rema Padman, arXiv preprint, LLMs Know the Constraint But Do Not Use It: Activation Bottlenecks in Pragmatic Constraint Reasoning (arXiv:2608.12321) (2026-08)Preprint formalizing conditional constraint activation: a feasibility constraint is internally encoded (Knowledge) symmetrically across constraint-present and -absent prompts (Symmetry), yet only sometimes routed into the decision (Routing). A quartet diagnostic over 14 models reveals two failure modes; probes on two open weights decode the constraint above 88%, while activation patching repairs one failure mode (+6.4 nats) and not the other (-0.07). Concludes hidden-constraint failure is a routing problem, not a knowledge problem. Abstract page fetched directly. arxiv.org
Weekly digest: the most-read systems, in brief. Mondays.

Comments

Always open. Logged-in readers can annotate paragraphs in place.

Loading comments…
or log in to comment under your account