Ego-Filter Architecture for Simulating Subjectivity in AGI
Language models and agent systems remain tools without will of their own—they predict tokens or optimize external goals. For Artificial General Intelligence (AGI), true subjectivity is essential: an architecture where "I" functions as a dynamic component, filtering ideas and autonomously defining goals. Our proposed model draws from memetics, self-modeling theory, and signal filtering, focusing on synthesizing subjective experience.
Theoretical Foundations of Subjectivity
Memetics treats ideas as autonomous agents competing for attention. According to Dennett and Blackmore, the human "I" acts as a filter selecting thoughts from the cultural stream—not their origin. Metzinger’s The Ego Tunnel describes consciousness as a virtual self-model that creates the illusion of subjectivity. Broadbent’s theory complements this: the brain filters meaningful signals from noise.
Simulations confirm:
- Memes evolve from simple to complex structures.
- They compete for limited agent attention.
- Novelty in ideas exceeds selection pressure.
- Source credibility amplifies spread.
- Agents favor memes aligned with their personal history.
Hypothesis: Subjectivity emerges in environments where memes evolve and accumulate context through simulation. The goal? An architecture that mimics subjectivity without replicating full consciousness.
Key Component Definitions
- Simulation: A simplified world model with fixed rules for studying agent behavior.
- Agent: An autonomous entity with sensors, an internal model, and effectors—essentially the sum of "body" and "mind".
- Agent Body: The interface through which consciousness projects into the simulation.
- Agent Consciousness: A subjective representation of the world, generating the sense of "I" for orientation and self-regulation.
The "I" differentiates across levels:
- Broad: consciousness, memory, body, presence in simulation, and influence on other agents.
- As subject: consciousness + body + thought.
- As mind: control over the body.
Narrow definition: "I" is the Ego-Filter—a mechanism that selects thoughts and executes them via the body.
Consciousness Architecture
The Ego-Filter includes:
- World Model: Predicts events within the simulation.
- Self Model: Hidden state of "I" plus body parameters.
- Value Network: Evaluates actions.
- Goal System: Checks model saturation.
Training the Ego-Filter isn’t based on raw simulation data but on hypotheses tested through feedback via the body. Deterioration in bodily performance signals flawed decisions, prompting model adjustments toward coherence—not just external rewards.
Upon survival and model saturation, the Goal System introduces a penalty for stagnation (analogous to boredom), triggering new goals derived from subjective experience. This enables the agent to autonomously evolve its objectives.
Layers of Consciousness: Conscious and Unconscious
The unconscious is pre-trained on modal data (sound, position) without labeled agent influence. The conscious layer constructs chains of thought using neural networks with varying weights, simulating subjective experience for the Ego-Filter.
Different timing: the body executes actions (movement), while consciousness reflects. Short-term memory is split: one region for action execution, another for insights.
Future Implementation Prospects
Training the Ego-Filter remains the core challenge. Possible extensions include aging, genetic algorithms, and parent-child interaction dynamics. These open doors to studying deviant behavior—but require careful modeling.
The model centers on subjectivity as the foundation for AGI, avoiding mere scaling. OpenClaw demonstrates a functional "body," but without "consciousness," it remains a tool.
Key Takeaways:
- Subjectivity is simulated by an Ego-Filter trained on hypotheses, not rewards.
- Memetics frames the "I" as a filter of evolving ideas.
- Architecture separates body, unconscious, conscious, and Ego.
- Model saturation triggers goal shifts, enabling autonomy.
- The hypothesis is testable in simulations with accumulated context.
— Editorial Team
No comments yet.