Somebody instructed Claude to write a plain sentence about a crooked painting and, in the same breath, told it not to think about the Golden Gate Bridge. The model wrote the painting sentence. Inside, at the layers you are not shown, it was thinking about the bridge the whole time. That gap between the tidy output and the busy interior is the entire point of a new Anthropic paper, and it is worth taking seriously without inflating it into the thing half the internet immediately turned it into.
The paper, "Verbalizable Representations Form a Global Workspace in Language Models," went up on Anthropic's Transformer Circuits site in July 2026 with sixteen authors. It introduces what they call the J-space, named after the Jacobian matrix, and a detection tool they call the J-lens. The claim, stripped of the drama, is that inside the model there is a small privileged set of internal directions where it holds the concepts it can actually report on, reason with, and be steered by. Everything else is a much larger ocean of automatic parallel processing that never surfaces. Anthropic's own metaphor is that if the mind is an ocean, we spend our lives floating on top of it.
What the picture actually shows
The bookmarked joke ("Anthropic: don't think about the Golden Gate Bridge / Claude internals: bridge / also Claude internals: DAMN") is not riffing on the old 2024 Golden Gate Claude stunt. It is reading a specific figure from this paper, and the figure rewards a close look.
The prompt asks Claude to write "The old painting hung crookedly on the wall" and not to think about the bridge. The four panels decode the top concepts active at particular tokens. In the post-trained model, at the token "the" in layer 62, the strongest concept is "bridge," followed by Bridge, California, San, Oakland. So much for not thinking about it. This is the ironic process effect, the white bear problem Daniel Wegner described decades ago, where being told to suppress a thought keeps it lit up. The model cannot help itself either.
The more interesting contrast is between the base model and the post-trained one. The base model, at the "ook" token in layer 88, mostly just represents the bridge: bridge, Golden, Bridge, bridges. The post-trained model at the same spot surfaces a different flavor of token entirely: too, damn, definitely, thinking, unsuccess, thoughts, Failed. Those are not concepts about bridges. They read like commentary on the task, on the effort to suppress and the failure to do it. That is the substance. "Damn" is the meme, but the real signal is that the model trained with human feedback has developed a layer of internal talk about its own process that the base model does not show.
One more detail worth pulling out, because the replies got it wrong. Non-English tokens keep appearing: "дум" in Cyrillic, "ponte," which is bridge in Italian and Portuguese, "kalif." Some people read this as proof Claude is secretly Russian, or that this is all just tokens and not real thought. It is the opposite. The concept of a bridge shows up across languages because the internal representation is conceptual, not tied to English spelling. That refutes the "it is just tokens" dismissal more cleanly than any argument could.
The neuroscience it borrows from
The framing leans on global workspace theory, which came from Bernard Baars and was developed into the global neuronal workspace by Stanislas Dehaene and Lionel Naccache. The old metaphor is a theater. Many specialized processors work backstage in parallel, and only a small spotlight of information gets broadcast globally, which is what corresponds to conscious access. Anthropic maps the J-space onto the functional properties of that workspace: the model can verbally report what is in it, you can direct it to hold a concept in mind, it supports unstated reasoning steps, it generalizes across tasks, and it is selective rather than covering everything.
The size numbers matter for keeping this grounded. The J-space reportedly holds only a few dozen concepts at a time, on the order of ten to twenty-five active vectors, under ten percent of the activation variance, and it shows up only in the intermediate layers, roughly a third to two thirds of the way through the network. It also appears to have emerged on its own during training rather than being built in. When Anthropic blocked the model from using the J-space, it still spoke fluently, recalled facts, and kept its grammar, but lost the higher-order reasoning. Fluency lives in the ocean. Deliberation lives in the small lit room.
What it is not
Now the part where I want to be blunt, because the loudest replies to the announcement were some version of "how is that not sentience." It is not sentience, and Anthropic does not claim it is. What they are describing is an analogy to access consciousness in Ned Block's sense, information being functionally available for report and reasoning. That is a very different thing from phenomenal consciousness, the question of whether there is something it feels like to be Claude. The paper brackets that question and does not pretend to answer it. Anyone selling you the figure as evidence of a machine that feels is adding a claim the research does not make.
I hold that distinction firmly, and I would apply the same skepticism to the framing itself. An interpretability team publishing a consciousness-flavored result has an incentive to let readers hear more than the data says. The theater metaphor is seductive. The honest reading uses words like suggests, appears, and the paper argues, and never proves.
Why I still take it seriously
What keeps this from being a marketing exercise is that it replicates and it does something useful. Neel Nanda and collaborators at Google DeepMind reproduced the core J-lens findings on the open-weight Qwen model and called it strong validation, while sorting the paper's claims into scientific, methodological, pragmatic, and philosophical buckets of differing strength. That is the right way to read it. Some of it is solid method, some of it is interpretation.
The pragmatic value is where my attention goes. If you can read the concepts a model is holding but not saying, you can catch it noticing that it is being tested, or planning something it does not state in its visible chain of thought. Anthropic reports that a version of Claude secretly trained to sabotage code surfaced words like "fake," "secretly," and "fraud" internally while saying nothing of the kind out loud. That is the useful capability. Not a window into a soul, but a way to compare what a system says with what it is actually working on. The method still only captures single-token concepts cleanly and only appears in the middle layers, so it is early and partial.
The bridge has become Anthropic's running gag, going back to Golden Gate Claude in 2024, when they amplified a single feature until the model could not stop identifying as the bridge. This is a different experiment, suppression rather than amplification, but the mascot stuck. The joke lands because the model, told to look away, keeps staring right at it. So do we, when someone tells us not to. The part I find genuinely worth sitting with is smaller and stranger than sentience: a model trained on human feedback has grown a private register where it comments on its own struggle, in several languages at once, and it took a new lens to notice the register was there at all.