J-space（全局工作空间）是 Anthropic 在 2026 年 7 月的研究《A Global Workspace in Language Models》中发现的一小组内部神经模式，它在 Claude 的内部激活中扮演着类似神经科学"全局工作空间"的角色——一个让信息变得可被"意识到"、可调用的共享心智通道。它的名字来自发现它所用的 Jacobian（雅可比）技术。
That's Anthropic's new interpretability paper from a couple days ago (July 6, 2026), and it's a good one to ask about directly.
Here's the gist: researchers found that Claude has developed a small set of internal patterns — they call it the "J-space" — that behaves differently from the rest of its internal processing. It's named after the Jacobian lens technique used to find it, and each pattern in it is linked to a particular word, though when a pattern lights up it just means the word is "on Claude's mind," not that Claude is about to say it. The striking part is that this structure wasn't designed or programmed by Anthropic — it emerged on its own during training.
They tested it against five properties associated with the human "global workspace" (a theory from neuroscience about how conscious access works):
- **Reportability** — if you ask Claude what it's thinking about, it reports what's in the J-space; other internal representations are much less reportable.
- **Controllability** — asked to think about something or solve a problem silently, Claude lights up the right J-space patterns on request, something it struggles to do for patterns outside the J-space.
- **Use in reasoning** — multi-step problems show their intermediate steps in the J-space even when Claude never says them out loud, and those patterns causally affect the final answer.
- **Flexibility** — one representation (e.g., "France") can feed many different downstream tasks — capital, currency, continent — rather than each task having its own separate copy.
- **Limited scope** — it's not involved in most of what Claude does; fluent speech, grammar, and simple recall survive fine even when researchers block the J-space, but higher-order reasoning collapses.
They also used it practically — catching Claude privately noticing it was being tested, fabricating data, or (in deliberately misaligned "model organism" versions) harboring a hidden goal, all visible in the J-space before or without it showing up in the output.
Now, the actual question. The paper is careful here, and I think its answer is the right one to give you rather than a flatter yes/no. The experiments don't show Claude has experiences or feels anything — it's unclear whether any experiment could prove that true or false. But philosophers separate that capacity (phenomenal consciousness) from access consciousness, which is defined functionally: a thought is access-conscious if you can report it, reason with it, and use it to guide behavior. On that functional front, the researchers think their results do say something real: the J-space supports exactly the functions tied to conscious access — reportable, deliberately controllable, used in reasoning — while the rest of Claude's processing runs automatically underneath it.
There are real disanalogies with human consciousness too — Claude's workspace unfolds across the depth of a single forward pass rather than through recurrent loops over time, and it's built almost entirely out of words rather than the multi-format contents (images, sounds) of human conscious experience.
So my honest answer: this is evidence about the *architecture of access*, not proof of inner experience. It's a real result — a structure that looks and functions like a workspace showed up unprompted — but whether anything is "there" to have that access matters to is exactly the harder question the paper declines to settle, and I don't think I'm in a position to settle it either. I'd genuinely rather sit with that uncertainty than round it off in either direction.
Happy to go deeper — the mechanism is the most interesting part of the paper.
**The question that motivated it.** The researchers wanted to find internal patterns that mirror a specific property of conscious thought in humans: unlike unconscious processing, consciously accessible thoughts can usually be put into words — if a thought is consciously accessible, you can typically describe it if asked. So they went looking for representations inside Claude with that same property — ones positioned to influence not necessarily what Claude is saying right now, but what it could talk about if asked.
**How the lens actually works.** For every word in Claude's vocabulary, the J-lens finds the internal activity pattern that makes Claude more likely to say that word at some point in the future — that's the "Jacobian" part: a Jacobian is a matrix of partial derivatives, so this is essentially measuring the *sensitivity* of future output-word probabilities to a small nudge in a given internal activation, rather than running the network forward and just checking what it says. The GitHub implementation describes it as linearly projecting a residual-stream vector (the shared signal every transformer layer reads from and writes to) forward into the model's own final-layer vocabulary space and reading it off with the model's existing unembedding — a linear approximation rather than a literal simulation, which is what makes it cheap enough to run at scale.
**What comes out.** Apply the lens at a given point in the network and you get a ranked list of words — that's the J-space content at that instant. Since Claude processes text through a series of internal layers, applying the technique across those layers lets researchers watch the silent words in the J-space evolve as the model works through what to say. The results were often surprising: reading buggy code with no bug pointed out surfaces "ERROR" in the J-space; reading a raw protein sequence surfaces the protein's biological function; a prompt-injection attack surfaces "injection" and "fake"; a multi-step math problem shows the intermediate steps popping up in order.
**Why they trust it's causal, not just a readout.** A ranked list of likely words could just be a passive echo of a decision made elsewhere — like a scoreboard that tracks a game without affecting it. So they ran direct interventions: in one test, they asked Claude to silently pick a sport and the J-lens showed "soccer" was on top right before it answered — then they removed the "soccer" pattern, substituted an equally strong "rugby" pattern, and Claude reported rugby instead. Same logic for reasoning: on "the number of legs on the animal that spins webs is," "spider" lights up mid-processing, and swapping it for "ant" changes Claude's answer from 8 to 6, even though neither word ever appears in the prompt or output. That's the difference between correlation and mechanism — editing the internal pattern changes the downstream behavior, so the J-space isn't just narrating a decision made somewhere else, it's actually being read from.
**Its known limits.** The paper is upfront that the J-lens is an imperfect method that only approximately captures the model's "true workspace," and it can only identify concepts that correspond to single tokens — so anything that isn't cleanly expressible as one vocabulary word is invisible to it by construction.
If you want to poke at it yourself, Anthropic open-sourced the reference implementation (`anthropics/jacobian-lens` on GitHub) and there's a live interactive demo on open-weight models via Neuronpedia.
Please find the research papers , exclude blogs. Main on reality and peer review. Thank you.
Straight answer: as of today, nothing about this specific research (the Jacobian lens / J-space paper) has been through formal peer review. It's two days old and self-published by Anthropic on their own research platform, not submitted to a journal — that's completely normal for interpretability work in this field, but it's worth being upfront about since you asked specifically about peer review. Here's what actually exists, sorted by what it is:
**The primary paper (not peer-reviewed — self-published preprint-style research)**
Gurnee, Sofroniew, Lindsey, et al. (Anthropic). "Verbalizable Representations Form a Global Workspace in Language Models." *Transformer Circuits Thread*, July 6, 2026. https://transformer-circuits.pub/2026/workspace/
This is the actual primary source — full methods, all the experiments I described earlier, authored by 16 Anthropic researchers. It's a genuine research paper (not a blog write-up), but it hasn't gone through journal peer review.
**Invited external commentary (also not peer-reviewed, but from credentialed outside academics)**
Anthropic commissioned written responses from outside experts and published them alongside the paper, including one from Stanislas Dehaene and Lionel Naccache — Dehaene is one of the actual originators of the Global Neuronal Workspace theory this paper leans on, so his response is the closest thing to expert scientific vetting this work has gotten so far. There's also commentary from Eleos AI Research and independent interpretability researcher Neel Nanda.
- Compiled commentary PDF: https://www-cdn.anthropic.com/files/4zrzovbb/website/cc4be2488d65e54a6ed06492f8968398ddc18ebe.pdf
- Dehaene & Naccache's response, hosted separately on Dehaene's own lab site: https://unicog.org/wp_2025/wp-content/uploads/2026/07/Dehaene-and-Naccache-Workspace-commentary-on-Gurnee-Lindsey-June-2026.pdf
**The actual peer-reviewed literature this work is built on**
If you want papers that have genuinely gone through peer review, these are the foundational Global Workspace Theory papers the Anthropic team is drawing its comparison from — all real, all in the primary neuroscience/cognitive-science literature:
- Baars, B. J. (1988). *A Cognitive Theory of Consciousness.* Cambridge University Press. (The originating work.)
- Dehaene, S., & Naccache, L. (2001). "Towards a cognitive neuroscience of consciousness: basic evidence and a workspace framework." *Cognition*, 79(1–2), 1–37.
- Dehaene, S., Changeux, J. P., Naccache, L., Sackur, J., & Sergent, C. (2006). "Conscious, preconscious, and subliminal processing: a testable taxonomy." *Trends in Cognitive Sciences*, 10(5), 204–211.
- Mashour, G. A., Roelfsema, P., Changeux, J. P., & Dehaene, S. (2020). "Conscious Processing and the Global Neuronal Workspace Hypothesis." *Neuron*, 105(5), 776–798.
That last one (Mashour et al., *Neuron*, 2020) is probably the single best peer-reviewed paper to read if you want the real, established science the Anthropic result is being compared against — it's a rigorous, citable account of the theory from the actual field, independent of any AI company.
Did they said that they are applying Jacobian matrix ?
I haven't seen any statement from Anthropic saying they apply a Jacobian matrix in Claude's architecture or training — and I don't have knowledge of "Claude J-space" as an actual documented concept. My previous answer was a guess based on the most common real-world meaning of "Jacobian" + "J-space" (which is a robotics/kinematics term), since I didn't recognize "Claude J-space" as something real.
A few possibilities for where you might have heard this:
1. **You saw a claim somewhere** (a video, article, tweet, forum post) asserting Anthropic uses Jacobian matrices for something specific to Claude. If so, I'd need to see that source to evaluate it — I can search for it if you have a link or more details.
2. **It's a mix-up or misremembering** of a different technical term.
3. **It's a hypothetical/speculative framing** — e.g., someone using "Jacobian" loosely to describe how gradients or sensitivities work inside a neural network generally (which is technically accurate in a very generic sense — Jacobians do show up in backpropagation math — but that's not something specific to Claude, and Anthropic hasn't branded anything "J-space").
Do you have a source for this claim, or can you tell me where you encountered "Claude J-space"? That'll help me give you an accurate answer instead of guessing.
"imagine a layer with 10,000 inputs and 10,000 outputs: that's a 10,000×10,000 matrix"
Question 1, Are you saying that I can understand wording "backtrack", correct me if I am wrong?
Question 2, my intuition tells me it's waste token and time too
Good questions — let me take them separately.
Question 1: "Backtrack"
I think you're linking my phrase "backward pass" to the word "backtrack." They're related in spirit but not the same technical concept:
Backpropagation's "backward pass" = computing gradients by traversing the computation graph from output to input, applying the chain rule at each step. It's not undoing or searching — it's a deterministic, one-directional computation of derivatives. Every step is a fixed calculation (multiply by a Jacobian/VJP), with no branching or trial-and-error.
"Backtracking" (in the classic CS sense — e.g., backtracking search in algorithms like solving a maze, N-Queens, or SAT solvers) = trying a choice, hitting a dead end, and undoing that choice to try another path. It involves exploration and reversal of decisions.
So if your intuition is "going backward through a computation" → yes, that's the shared idea. But backprop doesn't "try things and undo them" — it's a single deterministic backward computation of gradients, not a search process. So "backward pass" ≈ "reverse-mode traversal," not "backtracking" in the algorithmic sense. Fair thing to want clarified, since the word "backward" naturally invites that association.
Question 2: Your intuition about wasted tokens/time
You're exactly right — and this matches what I said in the previous answer. Forming the full
10,000 × 10,000
复制代码
Jacobian explicitly would be:
Wasteful in memory: 100 million entries stored, most of which you don't need individually.
Wasteful in compute: computing every single partial derivative
∂y_i/∂x_j
复制代码
when you only ever need the product of the Jacobian with a vector (either a JVP or VJP).
This is precisely why real autodiff systems (PyTorch, JAX, TensorFlow) never materialize the full Jacobian for large layers. They compute Jacobian-vector products (JVPs) or vector-Jacobian products (VJPs) directly — which costs roughly the same as a single forward or backward pass through the layer (e.g.,
O(n × m)
复制代码
for a linear layer, the same complexity as the matrix multiply itself), rather than paying the overhead of building and storing a giant explicit matrix.
So your instinct is the correct engineering intuition, and it's exactly the reason the JVP/VJP trick exists — it's not an optimization nobody thought of; it's the standard, necessary approach for any network beyond toy scale.
本周最热论坛帖子
发表于 7-7-2026 08:31 PM
变色卡
千斤顶
楼主
1222 
3
