We've long known that large language models (LLMs) are complex black boxes. We see the input (prompt) and we see the output (response), but the process happening in between has remained largely unclear — until now. Anthropic, in new research, has presented evidence of a "global workspace," which functions similarly to working memory in humans.

What is J-space: Silent reasoning inside the neural network

The key discovery is the existence of so-called J-space. The name refers to the mathematical technique Jacobian, which the scientists used to identify specific patterns in the model's activations. J-space is not just another layer of neurons; it is a collection of internal patterns that play a special role in information processing.

The most important thing, however, is the difference from what we already know. If you've ever used models with Chain of Thought (CoT) capabilities, such as OpenAI o1, you know that the model "thinks" by writing reasoning steps into a text window. But that is visible to the user. J-space works silently. The model can, within its neural activations, simulate concepts, solve intermediate steps of logical problems, or prepare a response without generating a single text token. It's a process that happens "in the model's head" but never appears on the screen.

J-space vs. Chain of Thought: Key differences

To understand the technical significance, it's worth comparing these two approaches:

• Chain of Thought (e.g., OpenAI o1): The model explicitly generates text steps. It's comprehensible, but computationally and token-intensive (you pay for every written step).
• J-space (Claude): The model uses internal representations. It's more efficient, faster, and allows the model to "think" about things that would take up too much space in text output or would disrupt the flow of conversation.

Why this discovery is crucial for AI Safety

From the perspective of safety and regulation, such as the European EU AI Act, the ability to "read the mind" of a model is absolutely key. If we can identify J-space, we can start monitoring whether the model is internally performing processes that conflict with its instructions or ethical rules, even when its output looks perfectly fine.

This discovery pushes the boundaries of interpretability. In the past, we relied on the model to "say what it's doing." Now, however, we have tools to observe its actual internal states. This could be crucial for detecting so-called "deceptive alignment," where an AI might try to mask its true intentions through text output while its internal processes head in a different direction.

Practical impact: What does this mean for users and businesses?

For the average user, this architectural change might not manifest as an immediate "turn off thinking" button. Rather, we will see models that are smarter, faster, and more logical, without requiring massive text outputs for every complex task. For businesses, this means the ability to deploy agents that perform complex reasoning at a fraction of the token cost, because most of the "thinking" happens within internal activations rather than as generated text.

Availability and pricing: The Claude model (including the latest versions such as Claude 3.5 Sonnet) is fully available to users in the Czech Republic through the web interface Claude.ai and via the API.

• Free tier: Free with a limited number of messages.
• Claude Pro: Approximately $20 (about 470 CZK) per month, offering higher limits and priority access.

Summary: A new perspective on digital intelligence

The discovery of J-space within Anthropic's research changes our view of what we consider "intelligence" in LLMs. It's no longer just about statistically predicting the next word, but about the emergence of complex internal spaces that enable the simulation of reality and logic. For the Czech tech sector and developers building on AI, this means we are approaching an era where we will be able not only to converse with models, but also to truly understand the mechanisms of their decision-making.