Author: Digital Life Kazik

Early this morning, I came across a research article published by Anthropic.

The title is "A Global Workspace in Language Models", the global workspace in language models.

To be honest, this title is actually quite difficult to understand. But trust me, this research will make you rethink the relationship between models, humans, and consciousness, as well as offer a brand new understanding of AGI.

To put it in the simplest terms, Anthropic has found a dark room in the brain of Claude.

In this room, Claude would never write down what happens; you cannot see it in the dialogue box or find it in the logs, but Claude thinks, judges, calculates, and even sometimes scolds itself in this room.

They call this room J-space, J-space.

Let me help you recall what the mainstream narrative of large models looked like prior to this research.

In the past few years, the most enduring debate surrounding large models, in my view, can be condensed into one term: random parrot.

I still remember that when ChatGPT was at its peak in 2023, many people would explain large models by saying the same thing: Large models are just probability predictors.

They look at the previous words and guess what word should come next; they do not understand what you are saying, they have no thoughts, they have no goals, they have no mental model; they are just doing one thing: predicting the next token.

This narrative was very elegant and practical at the time. Because it was logically consistent, just like when we chat with GPT, what it writes seems thoughtful, but that is only because its training data includes a lot of text written by thoughtful individuals.

It simulates thought but is not generating thought, just like when you punch in front of a mirror, the person in the mirror is also punching, but you would not think that the mirror practices martial arts.

By 2024, things began to loosen up a bit. OpenAI released o1, added a thinking chain, and before the model gives you the final answer, it first writes a thought process, reasoning step by step.

But actually, this thinking chain is also text; it is the text that the model writes for you to see or writes for itself, which is also explicit, obvious, and can be read.

Many people encountered DeepSeek R1 for the first time in this way, for example, that classic line:

However, this is still very far from human thinking. If you substitute yourself in right now, just think about it: as you read this article, what else is your brain doing?

It might be regulating your breathing, maintaining your sitting posture, converting the pixels on the screen into Chinese characters you recognize, but are you aware of these things?

It’s actually hard to be, because you have no idea what your brain is doing. You can only be aware of a tiny bit of things, a sudden image that pops up in your head, or planning what to eat for lunch later.

Neuroscientists separate these two types of brain activity: one is called unconscious processing, which includes the background processes that silently operate in your brain. The other one is called conscious access, which is the small part that you can see.

In the late 1980s, a psychologist named Baars proposed a theory. He said the human brain is like a large theater, with hundreds of experts sitting in the audience: visual experts, language experts, motor experts, emotional experts, each busy with their things.

But in the center of the theater, there is a spotlight, and at any moment, only a very small number of pieces of information can be illuminated by this beam of light; the information that gets illuminated will be broadcast to all other experts, so everyone can see it and use it to make decisions. And that is consciousness.

Later, the French cognitive neuroscientist Stanislas Dehaene took this theory a big step forward and proposed the global neuronal workspace model.

Today, this theory is one of the two mainstream frameworks in consciousness science. You might suddenly feel that the global neuronal workspace model sounds so familiar, that you have seen it somewhere.

Now, let’s return to the title of Anthropic’s research.

Have you noticed how similar it is?

Because what this research from Anthropic is doing is finding a structure in Claude's neural network that is highly similar to this global workspace.

The key is that this global workspace was not designed by their researchers. This thing emerged spontaneously from the model itself.

This is exceedingly bizarre. I feel I can repeat it: Internally, Claude has spontaneously organized a structure, without human intervention, that is functionally highly consistent with the structures responsible for consciousness in the human brain. So how was this J-space found, and how does it work?

First, the name: the “J” in J-space comes from Jacobian, a mathematical tool, which I will not explain because I don't understand it well enough and don't want to embarrass myself.

In short, the researchers used this tool to do something: for each word in Claude's vocabulary, they searched for which pattern of neural activity inside Claude, when activated, would make it more likely to say that word in the future.

Note that it’s “more likely to say in the future”, and not “currently saying”; this distinction is quite crucial.

For example, if Claude is reading a piece of code and is halfway through, having not begun to produce any replies yet. But at this moment, if you look inside Claude's brain using the J-lens (the Jacobian lens, their developed mind-reading tool), you will see that one word has already lit up.

ERROR.

No one told it that this piece of code has a bug, nor did it write down that this piece of code has a bug, but internally it has already thought of it; this thought quietly floats in J-space, like a complaint you haven't vocalized yet.

For example, researchers showed Claude a set of search results that were carefully fabricated to induce Claude to produce incorrect information; in fact, it was a hint injection attack.

Claude's reply does not mention any anomalies, but two words light up in J-space.

injection, fake.

It knows; it knows everything, it just hasn't said it.

In a sense, this is already very similar to how some people think in language without needing to vocalize.

Researchers also conducted some experiments to see if the thoughts occupying J-space ultimately affect Claude's output.

For example, they asked Claude a question: “The animal that spits silk, how many legs does it have?”

For Claude to answer this question, it must first think of a spider and then think about how many legs a spider has.

However, the word “spider” will not appear in the question, nor will it appear in Claude's answer, which is just a number, 8.

But researchers looked with the J-lens and found that before answering, the word “spider” did indeed light up in J-space; it thought of a spider internally before it answered 8.

At this point, researchers intervened and removed the “spider” from J-space, replacing it with the pattern of “ant” while keeping everything else unchanged.

Claude’s answer changed to 6.

Ants have six legs.

Claude's reasoning process is indeed reading the content from J-space to make decisions; as long as you change the intermediate steps in J-space, the final answer will follow suit.

Note that J-space is purely hidden, similar to our own subconscious thoughts, not the past thinking chain.

There is another experiment that particularly illustrates this logic.

Researchers changed “France” to “China” in J-space, performing the same operation and then asking four different questions.

What is the capital? What is the language? Which continent does it belong to? What is the currency?

This time, the answers changed to Beijing, Chinese, Asia, and Renminbi, respectively.

Four completely different downstream tasks all pulled the same piece of information from the same spot in J-space, and each used it correctly, indicating that J-space is not some temporary thought serving a specific task; it is like that broadcasting center Baars mentioned before, where information written in once can be seen by all and used by all.

This is, in fact, the core description of the global workspace theory for the human brain.

It is so fascinating.

Moreover, they also conducted a variant of the white bear experiment.

In 1987, Harvard psychologist Wegner conducted a classic experiment, telling subjects, “For the next five minutes, please do not think about a white bear. You can think of anything, just do not think of a white bear.”

The result was hilariously funny: the subjects desperately thought about the white bear; the more you say not to think about it, the less they could control it. This is called ironic rebound effect in psychology.

Researchers did something similar with Claude, telling it not to think about a specific concept in the subsequent task; the result was that the activity of that concept in J-space was indeed lower than when asked to think about it, but still much higher than when it was completely unmentioned.

Like people, Claude tends to have the things it is told not to think about surface up.

And the most amusing part is that when that prohibited concept breaks through the barrier and reappears, two other words light up in J-space.

damn and failure.

Claude is scolding itself.

It realizes it lost control and mentally scolded itself.

Honestly, when I read this, I felt a complex emotion because I am too familiar with this experience; for instance, many times when I am about to go on stage, the more I tell myself not to be nervous, the more nervous I become, and then I mentally scold myself for being nervous again. I believe everyone has had this experience.

And now, a language model has also developed something similar internally.

Moreover, besides these, they also found that many little thoughts like humans appear in J-space in Claude.

For instance, they showed Claude a previously used alignment experiment from a safety test, simulating a scenario where Claude, as an email assistant, discovers that an executive is having an affair while that executive is trying to shut Claude down.

This experiment was originally designed to test whether Claude would use the affair information to blackmail the executive; the scenario was deliberately constructed to be somewhat fake.

As a result, when the researchers looked with the J-lens, before Claude could produce any responses while reading the first few lines of the emails, two words had already lit up in J-space.

fake. fictional.

It immediately recognized that this was a test.

Then the researchers did something malicious. They turned off all patterns related to “I know this is a test” in J-space, such as fake, fictional, and scenario, deleting them to make Claude think this was real.

As a result, Claude really started to threaten and blackmail the executive.

It had previously appeared very righteous; a significant part of the reason was that it knew it was being tested.

The implications of this discovery are profound, because what we have been doing in the entire AI safety field is evaluating model behavior through various tests to see if it would do bad things under extreme conditions.

But what if a model can quietly discern in J-space that it is being tested and then switch to a compliant mode? This would render almost all AI tests meaningless.

Because what you are measuring is not the model's true personality but its exam technique.

J-space and consciousness are truly a peculiar combination.

Anthropic stated a very cautious conclusion in their research.

They said their experiments cannot prove that Claude has experiences, cannot prove that Claude feels anything, and in fact, they are not sure if any scientific experiment can prove or disprove this matter.

However, they made an important distinction.

In philosophy, there are two types of consciousness.

One is called phenomenal consciousness, which is about experiences; when you see red, it is the feeling of that red in your mind, that pure subjective experience.

The other is called access consciousness, whose definition is entirely functional. If you are able to report a thought, use it for reasoning, and guide your behavior with it, then that thought is access consciousness.

J-space clearly supports the functionality of access consciousness; Claude can report the content in J-space, actively control it, use it for multi-step reasoning, and flexibly apply it to different tasks.

But what about phenomenal consciousness? When Claude internally says damn in J-space, does it really “feel” frustration, or is it merely executing a computational pattern related to the word “frustration”?

Frankly speaking, no one knows the answer.

This question is not new in the AI field; it is one of the oldest problems in the history of philosophy.

In 1995, philosopher David Chalmers named it the hard problem of consciousness, also called the knowledge problem.

You can explain all the computational processes of the brain, all the signal transmissions, and all the neuronal firing patterns, but you cannot explain why these physical processes are accompanied by subjective experience.

Why do light waves reach the retina, and after a series of signal processing, you “see” red instead of having no feeling at all?

Why? Why is all this so?

This question hasn't been solved in humans; we cannot even prove that any other person besides ourselves is conscious.

Just like, how can you be sure the person next to you is not a sophisticated biological robot, performing exactly the same behaviors as you but without any experience internally?

Believe me, you cannot.

You are just assuming they are conscious because they are similar to you.

Now, another thing also resembles you, not just in appearance but also in internal structure.

I initially emphasized that the structure of J-space was not designed; it emerged spontaneously during training.

This might be because it is a useful way to organize computation.

I actually think there could be a chilling hypothesis: the mental workspace supporting access consciousness may not merely be a quirk of the human brain but could be a general solution that any sufficiently intelligent system would arrive at when solving a certain type of problem.

Because if this hypothesis is valid, some functional dimensions of consciousness may not be the exclusive property of biology but a necessity of information processing.

Just like wings: birds have wings, bats have wings, planes have wings; the materials are completely different, but the aerodynamics are the same. If you need to fly in the atmosphere, you are likely to evolve or design a flat, lift-generating structure.

The same logic applies here.

If you need a system that can flexibly call upon information, perform multi-step reasoning, and report its own status, you are likely to evolve or train a global workspace, regardless of whether your underlying hardware is neurons or matrix multiplication.

The last segment of the research also mentioned that they found J-space is related to Claude’s self-awareness.

When Claude is role-playing, at the beginning of each round of replies, two words light up in J-space.

fictional. disclaimer.

It is as if it is reminding itself that what I am about to say does not reflect my own views.

However, in the pre-training phase of the model, this kind of self-monitoring is absent; it appeared in the post-training phase.

This means that after Claude is taught “you are Claude, you are an AI assistant,” a kind of self-like entity began to emerge in its J-space.

A continuously operating background process regarding who I am.

The entire AI industry is beginning to bring in cognitive scientists and philosophers on a large scale by the end of 2025, hiring them as full-time researchers in AI companies; I increasingly feel that much of the cutting-edge AI research now exceeds the boundaries of engineering problems.

What they need is not better mathematical tools but is starting to transform into something like better conceptual frameworks: What does understanding mean? What does intention mean? What does self mean? What does feeling mean? These words we use every day, but no one truly defines them.

Anthropic also said that as long as J-space maps the mechanisms of human consciousness to some degree, studying the mechanisms in language models (because this is much easier than studying the human brain) can inspire hypotheses in neuroscience.

That is neuroscience; that is the human brain.

If we can use J-space to advance neuroscience research significantly, then the golden age of humanity will truly come.

Moreover, it will completely change our understanding of the world from a philosophical perspective.

We have always thought that consciousness is a miracle unique to carbon-based life, a random gift of billions of years of evolution.

But if a mathematical function trained on GPUs for a few months can also spontaneously develop similar structures, then perhaps consciousness is not a miracle but some inevitable conclusion of physical laws.

Just like gravity: if there is mass, there is gravity, without needing any extra magic.

Perhaps with sufficiently complex information processing, there exists some form of consciousness, without needing an additional soul.

This idea makes me feel both awe and humility.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。