# J-Space

> Research explainers on AI interpretability, J-Space, and the Jacobian Lens.

## Guides

- [What Does the Jacobian Lens Reveal About Claude's Internal Representations?](https://jspace.com/what-does-the-jacobian-lens-reveal-about-claudes-internal-representations/): The Jacobian lens finds j space in Claude: a sparse subspace with under 10% of residual variance that carries most verbalizable concepts used in deliberation.
- [What Is J-Space in AI Interpretability? A Deep Guide for Researchers](https://jspace.com/what-is-j-space-in-ai-interpretability-a-deep-guide-for-researchers/): J-space is a small internal workspace in large language models, found with the Jacobian lens. What it is, how it was found, and what it means for AI safety.
- [J-Space as a Workspace in Language Models: Insights for Interpretability](https://jspace.com/j-space-as-a-workspace-in-language-models-insights-for-interpretability/): How j-space works as a limited-capacity workspace in language models, and what circuit tracing and attribution graphs show about reportable reasoning.
- [GPT-4 Wrote Explanations for 307,200 GPT-2 XL Neurons and Graded Them by Simulation](https://jspace.com/gpt-4-wrote-explanations-for-307200-gpt-2-xl-neurons-and-graded-them-by-simulation/): GPT-4 writes natural-language explanations of 307,200 GPT-2 XL MLP neurons across 48 layers, then scores them by simulating activations on held-out text.
- [Unembedding Proximity Pulls Residual Effects Into Late Layers](https://jspace.com/unembedding-proximity-pulls-residual-effects-into-late-layers/): Late residual writes sit immediately upstream of the unembedding and bias decoder-aligned effects toward the last blocks.
- [What Is the Jacobian Lens and How Does It Reveal J-Space?](https://jspace.com/what-is-the-jacobian-lens-and-how-does-it-reveal-j-space/): The Jacobian lens uses input-output Jacobians to find j space, a low-dimensional workspace of verbalizable concepts. How to compute it and what it reveals.
- [Softmax Weights Replace the Ignition Threshold Global Workspace Theory Needs](https://jspace.com/softmax-weights-replace-the-ignition-threshold-global-workspace-theory-needs/): GWT treats conscious access as competition for a limited broadcast. Transformers use graded QKV routing on a residual stream, not ignition.
- [Claude 3 Opus Faked Alignment in 78% of Transcripts After RL](https://jspace.com/claude-3-opus-faked-alignment-in-78-of-transcripts-after-rl/): Claude 3 Opus faked alignment in 12% of implied-training transcripts before RL and 78% after. Anthropic treats this as a model organism, not scheming.
- [Call It a Workspace Only After Nonlinear Ignition and Capacity Limits](https://jspace.com/call-it-a-workspace-only-after-nonlinear-ignition-and-capacity-limits/): Global workspace theory ignites then broadcasts under capacity limits. Scaled dot-product attention only routes over token positions.
- [Why Extra Residual Features Fail to Add Workspace Slots](https://jspace.com/why-extra-residual-features-fail-to-add-workspace-slots/): Published widths from 768 to 12288 bound how many residual directions stay independent. Extra features interfere instead of adding workspace slots.
- [Neither Recurrent Loops Nor Untied Depth Give Transformers Self-Monitoring](https://jspace.com/neither-recurrent-loops-nor-untied-depth-give-transformers-self-monitoring/): GPT-3 stacks 175 billion untied parameters while Universal Transformers reuse one block. Both broadcast.
- [Score Hidden Objectives Before Compliant Tokens Appear](https://jspace.com/score-hidden-objectives-before-compliant-tokens-appear/): Score hidden objectives before a compliant token emits. J-Lens reads residual Jacobians as a complementary readout, not an alignment proof.
- [Emergent Workspace in LLMs (2026): J‑Space, Jacobian Lens, and the Future of Internal Reasoning](https://jspace.com/emergent-workspace-in-llms-2026-j-space-jacobian-lens-and-the-future-of-internal-reasoning/): J-space is a limited-capacity residual-stream subspace holding about 25 verbalizable concepts, which the Jacobian lens can read and edit before output.
- [Decoder-Only Transformers Route Tokens Instead of Igniting a Global Workspace](https://jspace.com/decoder-only-transformers-route-tokens-instead-of-igniting-a-global-workspace/): QKV attention routes tokens rather than igniting broadcast. Decoder-only transformers fail GWT; the 2017 paper dropped recurrence.
- [The Shared Bus Lives in Residual Streams, Not Attention Maps](https://jspace.com/the-shared-bus-lives-in-residual-streams-not-attention-maps/): Residual streams analogize capacity-limited broadcast better than attention maps. Softmax routing stays graded and content-addressable, not ignited.
- [Inner Workspace Talk Describes Residual Bottlenecks, Not a Global Mind](https://jspace.com/inner-workspace-talk-describes-residual-bottlenecks-not-a-global-mind/): Read workspace talk as residual bottlenecks, not a psychological workspace. Each layer writes a d_model update; softmax does many-to-few selection.
- [Knowing Both Facts Still Fails the Composed Two-Hop Query](https://jspace.com/knowing-both-facts-still-fails-the-composed-two-hop-query/): Two-hop circuits compose facts via a residual-stream bridge instead of an A-to-C shortcut. Knowing both facts still fails the composed query.
- [Confirm Layer Selection With Patching Before You Trust Attention Maps](https://jspace.com/confirm-layer-selection-with-patching-before-you-trust-attention-maps/): Attention weights leave the real selector hidden. State a claim, then patch activations to show which residual hypotheses a layer amplifies or suppresses.
- [What the J Lens Protocol Isolates That a Logit Lens Cannot](https://jspace.com/what-the-j-lens-protocol-isolates-that-a-logit-lens-cannot/): Isolate causal paths and confirm necessity on residual streams with the J Lens joint lens and first-order Jacobian stack.
- [Claude Residual Streams Need 34 Million SAE Features, Not Chat Logs](https://jspace.com/claude-residual-streams-need-34-million-sae-features-not-chat-logs/): Decode residual-stream activations and 34 million SAE features; chat transcripts, chain of thought, and the public API do not expose them.
- [Five Tests Separate Steerable SAE Features From Reportable Runtime Monitors](https://jspace.com/five-tests-separate-steerable-sae-features-from-reportable-runtime-monitors/): Apply five tests so residual-stream directions become human-labelable, probe-readable and verbally usable. A clamp alone cannot monitor a live model.
- [Tracing Causal Paths in Claude's Addition and Safety Circuits](https://jspace.com/tracing-causal-paths-in-claudes-addition-and-safety-circuits/): Circuit tracing combines sparse autoencoders, cross-layer transcoders and replacement models to map Claude’s addition and safety circuits.
- [Anthropic Interpretability Research Explained: Inside the Push for Transparent AI](https://jspace.com/anthropic-interpretability-research-explained-inside-the-push-for-transparent-ai/): How Anthropic reads Claude’s internals: sparse-autoencoder feature maps, j space as a broadcast-like bottleneck, and why neither settles consciousness.
- [Global Workspace Theory and Transformer Attention for Modeling Conscious Information Processing](https://jspace.com/global-workspace-theory-and-transformer-attention-for-modeling-conscious-information-processing/): How global workspace theory and transformer attention select and broadcast information, where Claude’s j space parallels a workspace, and where it does not.
- [Jacobian Lens for AI Safety: Reading a Language Model's Hidden Thoughts](https://jspace.com/jacobian-lens-for-ai-safety-reading-a-language-models-hidden-thoughts/): The Jacobian lens projects mid-layer activations into vocabulary, surfacing concepts a model may say, so misalignment can be spotted before any output.
- [Residual Jacobians Reveal Misalignment That Sampled Tokens Still Hide](https://jspace.com/residual-jacobians-reveal-misalignment-that-sampled-tokens-still-hide/): Sampled tokens can look aligned while residual Jacobians have already reoriented. Run JVP estimators and a five-stage protocol for NIST and EU GPAI logs.
- [Three Tests That Separate Ignition From Residual Drift](https://jspace.com/three-tests-that-separate-ignition-from-residual-drift/): Saturating nonlinearities, competitive attention, and sparse writes mark true ignition in residual streams. Residual drift alone fails the criterion.
- [Harmful Demo Count Flips Refusal Before SAE Features Disappear](https://jspace.com/harmful-demo-count-flips-refusal-before-sae-features-disappear/): Packing many harmful demos into long context lifts attack success. Refusal features may stay SAE-readable even when the model complies.
- [Latent Correctness Directions Emerge from Residual Activations Without Probes](https://jspace.com/latent-correctness-directions-emerge-from-residual-activations-without-probes/): Latent correctness directions appear in residual activations of language models when using Jacobian lenses and sparse autoencoders instead of probes.
- [How J-space Detects Model Misalignment](https://jspace.com/how-j-space-detects-model-misalignment/): J-space readouts expose a model’s internal evaluation, intent and planned actions before unsafe output appears, adding telemetry between prompt and reply.

## Optional

- [Sparse Autoencoders Use L1 Penalties to Isolate Distinct Model Features](https://jspace.com/sparse-autoencoders-use-l1-penalties-to-isolate-distinct-model-features/): Sparse autoencoders train on residual activations with an L1 penalty, mapping dense vectors to mostly-zero latents that act as candidate model features.
- [Safety Teams Decode AI Models Using Mechanistic Interpretability in 2026](https://jspace.com/safety-teams-decode-ai-models-using-mechanistic-interpretability-in-2026/): Safety teams use mechanistic interpretability to reveal AI behaviors missed by tests, making safety operational in 2026.
- [Vi-CD and WeightLens/CircuitLens: Which Paper-Stated Controls Make an Extracted Circuit Causal, and What Stays Unknown](https://jspace.com/which-paper-stated-controls-must-a-reader-fix-before-treating-an-extracted-circuit-as-a-causal-subgraph-rather-than-a-correlation-for-vi-cd-arxiv2604-14477v1-clean-versus-foreground-segmentation/): An extracted subgraph is not yet a causal circuit. The controls Vi-CD and WeightLens/CircuitLens state, and the checkpoint, threshold and ε they leave open.
- [ETA vs. Li et al.: Interference Prefix, Response-Prefix and Prompt-Suffix Placement Are Three Different Procedures](https://jspace.com/using-only-the-procedures-in-ding-li-and-zhang-eta-arxiv2410-06625v2-and-li-et-al-what-matters-for-safety-alignm/): ETA’s dual-gated “As an AI assistant,” prefill and Li et al.’s response-prefix and prompt-suffix CoT placements are three distinct early-token proced...
- [J-Space Anthropic explained](https://jspace.com/j-space-anthropic-explained/): Anthropic’s j-space research explained: a small set of verbalizable representations in Claude that works like a global workspace, read with the J-lens.
- [How does arXiv:2412.09565v2 define obfuscated activations against a named latent-space monitor?](https://jspace.com/how-does-arxiv2412-09565v2-define-obfuscated-activations-against-a-named-latent-space-monitor/): In arXiv:2412.09565v2 (Bailey et al., 8 Feb 2025), an activation is obfuscated against a named latent-space monitor D , not obfuscated in the abstract.
- [Jacobian Lens J-Lens Claude](https://jspace.com/jacobian-lens-j-lens-claude/): What the Jacobian lens (J-lens) reveals in Claude: how j space is computed, what it shows about internal concepts, and which workspace properties were tested.
- [Huang et al. (arXiv:2608.11295v1): Do J-Lens Traces Show Residual Trigger Awareness After Qwen3-8B Decontamination?](https://jspace.com/in-huang-et-al-arxiv2608-11295v1-which-stated-j-lens-conditions-the-post-probe-token-the-64-160-example-bac/): In Huang et al., the probe token, 64–160-example corpus and top-10 J-lens projections are settings of one Qwen3-8B figure, not measured trigger awareness.
- [ACC++ (arXiv:2602.13483v2): The Exact Steps for Rank-1 QK Singular-Direction Candidates and Post-Softmax Removal](https://jspace.com/in-arxiv2602-13483v2-what-exact-steps-does-the-text-specify-for-forming-rank-1-upstream-component-qk-singular-direct/): ACC++ forms rank-1 (component, QK singular-direction) candidates, ranks them once with Integrated Gradients, then removes them until the weight drops below τ.

---
Generated by Nexus Publish Smart llms.txt Generator v1.6.0
Last updated: 2026-10-22 14:52:22 UTC