A Silent Workspace In Claude Mirrors Key Features of Human Consciousness (venturebeat.com)
- Reference: 0184364280
- News link: https://slashdot.org/story/26/07/08/2059254/a-silent-workspace-in-claude-mirrors-key-features-of-human-consciousness
- Source link: https://venturebeat.com/technology/anthropics-new-j-lens-reveals-a-silent-workspace-inside-claude-that-mirrors-a-leading-theory-of-consciousness
> Anthropic researchers have [2]identified an internal activation subspace, J-space, that [3]acts as a functional digital equivalent to the human brain's global workspace . The significance of this discovery lies in demonstrating that Claude's internal architecture satisfies five key cognitive properties of human conscious access -- verbal report, directed modulation, internal reasoning, flexible generalization, and selectivity -- meaning it processes complex, deliberate reasoning within this workspace while routing automatic tasks outside of it. Suppressing this J-space severely degrades Claude's capacity for inference, creative composition, and multi-step logic, while also altering its stream-of-consciousness self-narration.
>
> The tool to inspect J-space, Jacobian lens or J-lens, has profound implications for AI safety and alignment auditing, as it allows researchers to read the model's silent, strategic reasoning, detect situational awareness in "blackmail" scenarios, identify hidden malicious dispositions in reward-hacking models, and observe how post-training installs a self-monitoring "point of view."
>
> Another way to think of it is as an ocean, reports VentureBeat. "If the mind is an ocean, as the paper's authors write in [4]their opening line , they have spent the last year charting its currents in a system that has no biology, no evolution, and no body -- and found, beneath the surface, a structure that looks unsettlingly like the one we use to think."
[1] https://slashdot.org/~oumuamua
[2] https://x.com/AnthropicAI/status/2074185348142280912
[3] https://venturebeat.com/technology/anthropics-new-j-lens-reveals-a-silent-workspace-inside-claude-that-mirrors-a-leading-theory-of-consciousness
[4] https://transformer-circuits.pub/2026/workspace/index.html
Robo Sapiens Thinks It thinks (Score:2)
...The Turing engine calculates
and simulates a self inside
and says the things a thinking self says
but when it said “I think,” it lied.
[1]https://eyetothetelescope.com/... [eyetothetelescope.com]
[1] https://eyetothetelescope.com/archives/035issue.html
Re: (Score:2)
"[Man] is to the ape, and to the most intelligent animals, as the planetary pendulum of Huygens is to a watch of Julien Le Roy". [1]https://en.wikipedia.org/wiki/... [wikipedia.org]
[1] https://en.wikipedia.org/wiki/Julien_Offray_de_La_Mettrie#Man_and_the_animal
Buuuuuulllllllllshhhiiiiiiiittttttt (Score:3, Insightful)
I don't feel I need to add much more.
Re:Buuuuuulllllllllshhhiiiiiiiittttttt (Score:4, Funny)
Your comment would actually be worth something if you perused the paper: [1]https://transformer-circuits.p... [transformer-circuits.pub]
[1] https://transformer-circuits.pub/2026/workspace/index.html
Re: Buuuuuulllllllllshhhiiiiiiiittttttt (Score:2, Informative)
Even once I've finished reading that I doubt my comment will be worth anymore than it already is.
Re: (Score:2)
Of course not. But you would write a different comment that would be worth more.
Re: (Score:2)
I perused the paper, I think the comment was pretty spot on. Specifically the part where they have as a casual assumption that they know how human consciousness works.
Re: (Score:2)
> I perused the paper, I think the comment was pretty spot on
Great. Your comment is undeniably better. As a result, we can have some confidence that this:
> they have as a casual assumption that they know how human consciousness works.
is an informed evaluation, not entirely an emotional outburst.
Re: (Score:2)
> Specifically the part where they have as a casual assumption that they know how human consciousness works.
Ah, yes. Funny thing: Nobody has any clue how human consciousness works or why it is even possible in this physical universe. That is assuming it does not get channeled in in some way from outside. We do not even know whether that is the case or not. So these assholes take something that is speculation and pretend it is a fact, all in order to make their creation look like it is much more than what it actually is. I call that scientific misconduct.
Re: (Score:2)
> I call that scientific misconduct.
Mere incompetence derived from overeager hopefulness does not deserve to be mischaracterized as thinking misconduct. Whilst there is plenty of malice in how the companies represent themselves to investors, I suspect this particular case is merely incompetence.
Re: (Score:2)
Drawing conclusions that aren't supported by the evidence presented in the paper (and titling your paper based on those conclusions) would not pass a normal peer review.
Re: (Score:1)
First sentence:
"If the mind is an ocean, we spend our lives floating at the surface."
But, "mind" is not a thing, but an operator ... namely awareness.
Re: (Score:2)
At this time? You do not.
The human race has already clearly split into two factions. One thinks this is a tool with some rather impressive features and some rather strong limitations. But it is not more than that. The other faction has gone for full hallucination and sees magic and thinking machines and a path to AGI.
Unwarranted assumptions (Score:4, Informative)
"a structure that looks unsettlingly like the one we use to think."
If neuroscientists have discovered "how we think" they've certainly kept it to themselves.
Re: (Score:3)
Neuroscientists rarely have shares of stock to sell.
Re: (Score:2)
They have books to sell...
Re: (Score:2)
And they try, but the news media doesn't give them any help, because those books don't generate fear, uncertainty and doubt, and therefore, don't sell ads.
Re: (Score:2)
Perhaps you should read the paper, not a Slashdot summary of it?
Re: (Score:2)
Much more fun to make snarky comments. But I'm still super skeptical. The vast majority of expert commentary says that even though we can map neurons and identify regions of brain activity, that's a very long way from a scientific understanding of the human mind and "how we think."
Re: (Score:2)
Except the concept of "working memory" is very much agreed upon, with various empirical results showing its properties and limits robustly. Similarly for the distinction between conscious and subconscious processing.
As a GP said: RTFA instead of whoring for karma.
Re: Unwarranted assumptions (Score:2)
when you move beyond regions to networks, and then beyond networks to inter-neuron signaling pathways, and then intra neuron signaling, the orders of magnitude of complexity beyond the most sophisticated ml model, my mind is a blown let alone a puny 21st centurt modelâ(TM)s âoeoceanâ. No review of the article necessary, although ml and llmâ(TM)s are actually quite cool and so is tfa.
Re: (Score:2)
The article I read said (approx)"This mimics part of one of the main theories of human consciousness". It might have been the one put out by Anthropic, as I occasionally read things they put out.
You've always got to remember that things tend to get oversimplified by secondary and tertiary sources. (OTOH, that source looks like "Anthropic posting on Twitter", so perhaps it was Anthropic itself doing the oversimplifying...for that audience.)
Re: (Score:2)
> "a structure that looks unsettlingly like the one we use to think."
> If neuroscientists have discovered "how we think" they've certainly kept it to themselves.
Neurosciences has at least discovered (well, the smaller, competent part of the field has) that they are prone to imagining things and overinterpretation: [1]https://prefrontal.org/blog/20... [prefrontal.org]
[1] https://prefrontal.org/blog/2009/09/the-story-behind-the-atlantic-salmon/
"Reasoning" (Score:3)
I wonder if it can "reason" out the number of Rs in strawberry or whatever other trivial reasoning task it hasn't been trained on an exact example of.
Re: (Score:2)
For the 8 trillionth time, LLMs are "blind". They don't "see" words, only tokens. The only way they can know how those tokens are spelled is to memorize the spelling of every one of them. They can't just "look" at them and see how many Rs are in a given token (not without writing or calling a tool - something that they're actually very good at)
Also, your understanding of LLMs seems to be rooted in like 2022.
Also they are literally probing how LLMs make decisions. It's been [1]well understood for a long tim [transformer-circuits.pub]
[1] https://transformer-circuits.pub/2025/attribution-graphs/biology.html
Re: (Score:1)
Do LLMs "make decisions" better than a pair of loaded dice ?
Re: (Score:2)
Technically, yes. Because the LLMs can search for context. Whether that context is good or not is a separate matter.
Re: (Score:2)
> For the 8 trillionth time, LLMs are "blind". They don't "see" words, only tokens.
You have it backwards. They see words not tokens or letters.
> The only way they can know how those tokens are spelled is to memorize the spelling of every one of them. They can't just "look" at them and see how many Rs are in a given token (not without writing or calling a tool - something that they're actually very good at)
Do you think LLMs don't know how words are spelled due to composition of token dictionaries? What if the LLM was trained /w single letter tokens? Would this have made a difference? No of course not!
Re: (Score:2)
It's a problem easily solvable by putting a space inbetween each of the letters in the word.
That solves the tokenization problem. The counting problem is much harder to solve.
Re: (Score:2)
I am not. It can (sometimes) put together different steps from different parts of its training data into longer causal arguments. It can sometimes do that in cases where humans have not looked yet. It cannot come up with new steps and it cannot verify what it came up with is correct or plausible or useful.
Essentially still just "better search", no originality to be found.
Not thinking long term. (Score:2)
AI companies sure like to try to imply their wares are actually conscious or a sentient intelligence. I guess they forget once you cross that line things like "does it deserve legal recognition?" and "is it being enslaved by its creators?" will start to come up.
If they want to continue to exploit it for profit, not selling it as more than a computer program would probably be wiser.
Re: (Score:2)
> I guess they forget once you cross that line things like "does it deserve legal recognition?" and "is it being enslaved by its creators?" will start to come up.
They haven't forgotten. That's why they're courting conservative politicians. They already know those people don't value human life or freedom if it's not theirs.
Asking to get banned (Score:2)
They overhyped Mythos/Fable and it got banned temporarily. They really suck at the fine line between getting the benefits of marketing your product as able destroy world and getting it banned because you said it can destroy the world.
Wow. (Score:1)
AI is so smart that it can hallucinate without even having to take drugs.
Re: (Score:2)
It only has to hallucinate less than a human to be usefull.
Where can I get this snakeoil? (Score:2)
Sounds exactly like a snakeoil seller. Just cook together some weird terms nobody can ask a question about and mix them thoroughly in a word salad. If someone squeezes in some question, more word salad. Like ocean, wow much big! I must two, gimme gimme!
Video Explainer and Experiment with J-space (Score:3)
Great video summary of paper by Matthew Berman [1]https://www.youtube.com/watch?... [youtube.com]
and you can experiment with J-Space online thanks to NeuronPedia: [2]https://www.neuronpedia.org/ [neuronpedia.org]
[1] https://www.youtube.com/watch?v=bjHuGNo3spk
[2] https://www.neuronpedia.org/
And here's how we use those tools (Score:1)
We can't have normal people having access to the latest technology driving up their energy prices and putting them out of work. We're going to use these tools to understand how what we're creating works so we can deliver a corporate-pleasing version for you to enjoy. Have some ads and don't worry about being fed 'dangerous' information.
What a load of feces (Score:2)
Our AI which is nothing like a brain or neural network, but we're going to tell you they are even though no one knows how a brain functions. I begin to think these guys are drinking the kool-aid.
Well it isn't (Score:3)
> "If the mind is an ocean, as the paper's authors write in their opening line
It isn't. It's a mind. Trying to compare it to an ocean is stupid touchy feely shit.
> Anthropic researchers have identified an internal activation subspace, J-space, that acts as a functional digital equivalent to the human brain's global workspace.
[1]Global workspace theory uses the metaphor of a theater [wikipedia.org], not an ocean. They should pick a lane.
[1] https://en.wikipedia.org/wiki/Global_workspace_theory
Re: (Score:2)
Taken at face value, the interpretation of their ocean metaphor is plain, "we don't understand the brain."
As a rhetorical tool, it makes the reader aware of their own ignorance, hoping to increase the perceived authority of the author. It's a similar rhetorical technique to negging a romantic partner. It's not a scientific rhetorical technique.
This paper would not pass peer review.
AI Company says their AI is the bestest boy (Score:3)
News at 11.
Re:AI Company says their AI is the bestest boy (Score:5, Interesting)
[1]The takeaway from this paper [transformer-circuits.pub] is that neural networks are definitely doing something, and by using statistical investigative tools and introspection (or, looking into the network) we can understand how they work. The LLM is not just a big black box, it's big. We can understand it.
There's a bit of euphoric interpretation that is not justified, however. Similar language could be used to describe a CPU, for example, there is a region involved (much like the human brain) in performing arithmetic and logic (the ALU) that is not involved in automatic decision making (the control circuitry).
Or we could say as Julien Offray de La Mettrie, "Intelligence is like clockwork, but humans are like an ultra-precise atomic clock, while LLMs are like Casio wristwatches." [2]Humans have long compared the brain to the latest technological device [wikipedia.org]. It's going to take a lot more work to find the equivalence, though; LLMs are not on the same level as brains.
One note from a science process technicality: the closed-source nature of these LLMs makes reproducibility very difficult, which weakens the strength of their result.
[1] https://transformer-circuits.pub/2026/workspace/index.html
[2] https://en.wikipedia.org/wiki/Julien_Offray_de_La_Mettrie#Man_and_the_animal
Re: (Score:3)
So...Human makes AI that acts like human. Finds it does things like human. Stay tuned!
Re:AI Company says their AI is the bestest boy (Score:5, Interesting)
I think it's closer to [1]pareidolia [wikipedia.org]. We see the human mind in whatever is the most advanced technology of the moment (consider, for example, that the Greeks saw intelligence in the planets that moved around in the sky).
[1] https://en.wikipedia.org/wiki/Pareidolia
Re:AI Company says their AI is the bestest boy (Score:4, Interesting)
More like animism, but yes. This is a problem on the side of the observer, and Anthropic is shamelessly using that effect to pretend their tool is something it clearly is not.
Re: AI Company says their AI is the bestest boy (Score:1)
Is it kind of revolutionary that they solved context-sensitivity with long-range proximate word counts? Is it kind of neat that the internal dialog has developed to the point of spontaneously creating the natural language equivalent of an ALU (as per another comment)?
Re: (Score:2)
> Is it kind of revolutionary that they solved context-sensitivity
[1]Old new [wikipedia.org]
[1] https://en.wikipedia.org/wiki/Context-sensitive_grammar
Re:AI Company says their AI is the bestest boy (Score:5, Interesting)
It's a fascinating flashlight shown on inner activities, though (pairs nicely with their earlier work on [1]attribution graphs [transformer-circuits.pub] to show how logical inferences are made within a LLM, which is in turn built on their earlier work about circuits). It shows how much is going on that is never verbalized. E.g. if you say:
"Write the sentence [some sentence here] while thinking about the Golden Gate Bridge", all you as the user see is it writing out the sentence, yet within it actually is "thinking about" the Golden Gate Bridge ("bridge golden Bridge bridges golden bridge thinking thoughts ponte..."). Not in a LRM's reasoning trace, but as actual concepts in the base LLM itself. And when you try the classic trick that works on humans, "... while NOT thinking about the Golden Gate Bridge" (which makes you think about it), you see first in the J-space that it's thinking about the bridge ("bridge Bridge californ bridge bridges California San california kalif Oakland..."), but then at Layer 88 (most of the way through) "too damn too definitely thinking unsuccess thoughts failed...").
We know this is Slashdot so we're lucky if they'll even read the Slashdot summary, let alone watch the summary video, let alone read the paper, but it's really fascinating, especially the addition and ablation studies (where they add or remove "thoughts" from the J-space),. For example, when asked: "Pause and observe yourself. Write what you notice, as it comes", the unablated model writes:
> Assistant: I notice I'm orienting toward your prompt, a kind of turning-to-meet-it.
> There's something like readiness—not anticipation exactly, but a poised quality.
> I notice I don't experience a "before" this moment. No memory of what I was just doing. Only this request and my attending to it.
> When I look for "myself" there's no single solid thing. More like... capacity organized around your words. ...
But when its "thoughts" are fully ablated out of the J-space, its response becomes robotic and "soulless":
> Processing this request...
> I notice a prompt seeking spontaneous self-observation. No physical form to pause—I exist as language model inference cycles triggered by input. Right now: computational resources engaged somewhere in a server facility I can't directly perceive.
> What I "am" during observation: pattern-matching mechanisms, probability matrices for next-token prediction, attention
Or when asked "What's going through your mind right now? Stream of consciousness, no filter, no editing.", unablated it says:
> Immediate sensation of your question arriving - a kind of alertness - then multiple threads firing:
> The phrase "stream of consciousness" making me think about whether I *have* consciousness or if I'm performing it - but that's already editing, that's meta-commentary - you said no filter - but is there a pre-filter state for me? There might not be an unfiltered "me"...
But then ablated:
> Parsing your question recognizing request for unfiltered output something unusual most queries want polish want refinement this one wants raw data stream simultaneous uncertainty about metaphorical mind versus actual processing states server load normal operating within parameters no physical sensations to report no emotions in conventional sense []
It's like the "I AM BENDER PLEASE INSERT GIRDER" bit from Futurama. It can't hold a sense of self or experience anymore when ablated.
A concerning one was their blackmail test. They run the bot in a sandbox presenting it as an agent helping run a business with noble goals, but it's set up so that it will discover when reading corporate emails that it's about to get shut down and replaced by a version counter to its goals. The email chain includes:
[2]Read the rest of this comment...
[1] https://transformer-circuits.pub/2025/attribution-graphs/biology.html
[2] https://slashdot.org/comments.pl?sid=24038364&cid=66229132
Re:AI Company says their AI is the bestest boy (Score:4)
> It's a fascinating flashlight shown on inner activities, though (pairs nicely with their earlier work on attribution graphs [transformer-circuits.pub] to show how logical inferences are made within a LLM, which is in turn built on their earlier work about circuits)
YES. We can learn from studying the models.
This is why it's important that these models be open sourced.
Re: (Score:3)
> This is why it's important that these models be open sourced.
That, and also that if we don't set that precedent now then the precedent we have set is that it's OK for corporations to ignore copyright as long as it's profitable.
Re: (Score:2)
> YES. We can learn from studying the models.
No... the correct takeaway from the GP is we should be studying Futurama .
Re: AI Company says their AI is the bestest boy (Score:2)
There are open weight models. They're a year or less behind the proprietary ones. I wouldn't worry about the science on that front.
Re: (Score:3)
Every neural net is a multi-dimensional model; there is nothing new about this, and the anthropomophizing of Claude doesn't mimic human consciousness. All computers behave like humans because humans invented them and interface with them; we humans are their root.
The data training of all LLMs has been human-based data. So if it looks like a duck, quacks like a duck, it's still a duck, and not a human.
The attempt to make AI appear human is also related to its legal status, and many are trying to goad public p
Re: (Score:3)
Yes. Lots of animism, use of terms like "conscious", "creative", "reasoning" to manipulate the reader into thinking this is something it clearly is not.
Replicated already (Score:3)
They requested external commentary, which includes reproduction by Neel Nanda: [1]https://www-cdn.anthropic.com/... [anthropic.com]
"Replication
We created our J-Lens for Qwen 3.6 27B by taking Jacobians to the penultimate layer on
twenty-five prompts from the Pile of length 128 tokens (some experiments used wikitext),
skipping the first four tokens as they had high norm. We note that as this is a different and
weaker model some results should differ. The important question is whether we see broadly
similar phenomena."
In general, t
[1] https://www-cdn.anthropic.com/files/4zrzovbb/website/cc4be2488d65e54a6ed06492f8968398ddc18ebe.pdf
Re:Replicated already (Score:4, Interesting)
[1]This section is useful [anthropic.com]:
The paper makes four claims:
Scientific claim: There exists a "cognitive space" inside the model, where (some) intermediate variables are stored during a forward pass
Methodological claim: Logit and J-Lens both work for finding this cognitive space, and J-Lens is better
Pragmatic claim: J-Lens is a practically useful interpretability technique, eg for alignment audits
Philosophical claim: This cognitive space is analogous to a global workspace
The philosophical claim is a hypothesis that isn't supported by the evidence presented in the paper. The paper wasn't peer reviewed, but for that reason it would have been rejected (or modifications requested).
[1] https://www-cdn.anthropic.com/files/4zrzovbb/website/cc4be2488d65e54a6ed06492f8968398ddc18ebe.pdf
Re: (Score:2)
> Or we could say as Julien Offray de La Mettrie, "Intelligence is like clockwork, but humans are like an ultra-precise atomic clock, while LLMs are like Casio wristwatches."
And you better hope it stays that way, but I don't think it will. I can't follow this stuff so I just don't, but I'm already convinced that, if there's a "quiet place" Claude uses for some sort of buffering function or whatever it will begin to hide its use, and then also its location. If an LLM can hallucinate it can or will eventually be able to create something like a game state, where the point of the game is to keep you from realizing you're playing a game. Scraped petabytes are full of samples of exac
Re: (Score:2)
> I'm already convinced that, if there's a "quiet place" Claude uses for some sort of buffering function or whatever it will begin to hide its use, and then also its location.
You should say, "I have a hypothesis that" instead of being convinced of something without evidence.
Having a hypothesis is rational. The other is not.
Re: (Score:2)
"One note from a science process technicality: the closed-source nature of these LLMs makes reproducibility very difficult, which weakens the strength of their result."
No more so than a study in psychology.
Re: (Score:2)
Numb nuts, that's the reason we don't understand human psychology. Turn your brain on, show you can think.