CODE IN TOKYO

It Seems Claude Created a Place to Think Inside Itself

Anthropic published a very interesting piece of research.

When researchers looked inside Claude, they apparently found something like a place used for thinking.

Its name is “J-space.”

Of course, Anthropic did not write in Claude’s source code:

“This will be the thinking area.”

It emerged on its own during training.

That is what makes the story so interesting.

Is AI Really “Just Predicting the Next Word”?

When people explain LLMs, they often say:

“AI is just predicting the next word.”

As an explanation of the mechanism, this is not wrong.

But recently, that explanation alone has started to feel insufficient for describing how modern AI actually behaves.

When Anthropic’s research team examined Claude’s internals, they found that concepts not visible in the final output were still represented inside the model.

For example, when Claude reads code, nobody explicitly tells it that there is a bug.

But inside Claude, a representation corresponding to “ERROR” appears.

When it solves multi-step calculations, intermediate results that are not written in the final answer appear internally in sequence.

When it reads suspicious search results, representations related to “fake” and “injection” appear.

In other words, Claude is processing something internally even when it does not say it out loud.

If we put it in human terms:

This thing is silently thinking.

Claude Seems to Have Automatic Processing and Thinking Processing

What is even more interesting is that Claude does not use J-space for everything.

Writing ordinary text.

Handling grammar.

Recalling simple facts.

For these tasks, J-space does not seem to be very important.

But when Claude solves complex problems or uses intermediate results to make the next judgment, this area becomes important.

When the researchers interfered with J-space, Claude could still hold an ordinary conversation.

But its advanced reasoning ability declined.

What is that?

It sounds almost like the human brain.

Humans do not think, “Move the right foot, then the left foot,” every time they walk.

Most processing runs automatically.

But when we think about a difficult problem, we bring information into our mind and work with it.

In cognitive science, there is an idea called Global Workspace Theory.

The idea is that many processes run unconsciously in the brain, but only some of them rise into a shared workspace and are used for conscious reasoning.

The J-space that Anthropic found inside Claude appears to be functionally similar to this.

The Scariest Part Is That Nobody Designed It

This is the part I found most interesting.

Anthropic did not create J-space.

At least, it was not designed with the intention of saying:

“Let us create something like human working memory.”

Claude was trained, and this kind of structure formed inside it on its own.

In discussions about AI, the word “emergence” is often used.

When a model is trained at large scale, abilities that were not explicitly taught can suddenly appear.

This story feels like an internal-structure version of that.

To solve complex problems, it is efficient to place intermediate information somewhere and make it available to other processes.

So inside Claude, something like a shared “thinking space” emerged.

Of course, Claude itself did not decide:

“All right, I will build a thinking center.”

But as a result, a structure that humans did not design was formed.

As an engineer, I find this strangely frightening.

If an unknown function appeared inside a system I built, I would normally call it a bug.

And in this case, that function supports the system’s advanced judgment.

If you remove it, performance drops.

That makes it scary to touch.

This Is Not Research Claiming Claude Is Conscious

To be clear, Anthropic did not announce that Claude has consciousness.

The researchers themselves clearly state that the existence of J-space does not mean Claude has the same consciousness or emotions as humans.

If you skip that point and write:

“Claude has developed a self!”

then you have made a sensational YouTube video.

But that does not mean the research is uninteresting.

A function similar to one theory used to explain human consciousness appeared on its own inside a completely different mechanism called a Transformer.

That is what interests me more.

It was not built by copying the human brain.

It was not built with the goal of creating consciousness.

Even so, after being made to perform complex intellectual work, it arrived at a similar kind of function.

If this is not a coincidence, and if it is actually a structure that advanced intelligence needs in order to handle complex problems, what does that imply?

I thought humans and AI were walking completely different paths.

But perhaps we climbed the mountain from different trails and found ourselves approaching the same place.

That may be the kind of story this is.

For now, I still do not intend to say that Claude has consciousness.

But until a few years ago, people laughed and said AI was “just predicting the next word.”

When researchers looked inside that “just,” they found a thinking space that even the company that built it had not explicitly designed.

In old science fiction, AI suddenly awakens one day.

But reality may be much quieter.

“Wait, when did this feature appear?”

Moments like that may gradually increase.

And by the time we notice, nobody may be able to say where the “awakening” began.

Contact

I provide system development and technical consulting using AI, AWS, and Claude Code.

Contact form: https://holly-money-e94.notion.site/390ff30cf18c8086a676fe630d171873