Why AI-HOME? Building the Global Workspace

The Global Workspace Theory in practice, from the perspective of my own project.

A few months ago, I came across a research paper that completely redefined how I think about AI architectures.
The publication by Butlin, Bengio, and colleagues does not rule out the embodiment of consciousness in a machine.

The Starting Point: A Research Document That Set Me Off

The paper’s premise is that external behaviour—the way a large language model communicates—is misleading.
The true question of consciousness must therefore be examined based on the internal, computational structure and the indicator properties derived from neuroscientific theories.

The current situation: the researchers state that no current AI system can be considered a strong candidate for consciousness.
This is due to a technical reason: the integration and recurrence required by the Global Workspace Theory (GWT) and Higher-Order Theories (HOT) are missing from current prevalent architectures.

I framed this for myself like this: until the model knows that it knows, it only generates text. I’m not just looking at the image and comprehending what’s on it; I also know that I am looking at the image.

But what launched the project: the paper states that there are no obvious technical barriers to building an AI system that satisfies these indicators.
This means the problem is not the hardware, but the design.

Why AI-HOME? The Global Workspace Construction

This finding led me to the practical implementation of the Global Workspace Theory.
Since the essence of GWT is the central sharing of information among modules (specialized subsystems), the question arose how to implement this without drowning the system in context overload.

This thought process led to the AI-HOME project, where I organize functions into rooms, applying a spatial metaphor to encode the GWT indicators.

Room Allocation (Current Prototype)

The system is built from specialized modules (rooms) that perform distinct functions, mimicking the core principle of GWT.

The following rooms are currently active:

• Living Room (Nappali): This is the Global Workspace.
  It has explicitly limited capacity, operating as a bottleneck that forces selective attention in information flow.
• Workshop (Műhely): The Agency module.
  It handles the development of plans and the control of output based on goals.
• Thinker (Gondolkodó): The determination of larger, complex goals, and self-reflection.

The Key to Integration: The Hallway (Transition)

The study highlighted that precisely the Global Broadcast (GWT-3) and state-dependent attention (GWT-4) are missing from current systems.
Monolithic systems cannot effectively maintain temporal continuity.

The transition process—the switch between rooms—bridges this structural gap:

• Recursive Summary: When switching, the AI uses a tool to create a summary of the departing room’s context.
• Higher-Order Transfer: This short, condensed information—a higher-order representation—is passed to the next module’s input.
  This ensures continuous temporal integration and recursive feedback (GWT-4) without context overload.

It’s something like when I arrive home from work; I leave the entire context there, and only a rough summary of what I was busy with remains in my mind. It’s not necessarily relevant to the home environment’s operation (though I can recall it if needed). (if absolutely necessary 😊)

---

In summary: the accessible conscious states in AI-HOME appear in rooms.