The Dead Internet Theory is no longer a theory. It's a mathematical certainty.
We often fear that AI will become too smart. But my research suggests a more terrifying outcome: AI will become too recursive, leading to a state of semantic entropy I call "The Ainex Limit."
I am a 16-year-old developer and researcher. Over the past year, I have been studying the effects of recursive training loops on Large Language Models (LLMs).
The Hypothesis
If an AI model feeds on its own output for enough generations, the "variance" in its creativity doesn't just drop—it collapses geometrically.
To prove this, I built a physics engine using Python (GlowScript) to visualize this collapse in real-time, applying Chaos Theory (Lorenz Attractors) to semantic tokens.
The Simulation (The Chaos Engine)
I mapped semantic tokens as particles in a 3D space.
- Gravity: Represents the model's tendency to drift towards the mean.
- Entropy: Represents the loss of information over generations.
- The Singularity: The point of no return where "truth" becomes "noise."
Here is the simulation running in real-time. Watch how the "drift" (the red curve) thinks it is exploring new ideas, but is actually spiraling into a black hole of nonsense.
https://youtu.be/WJYipa7rT9Y
(Note: The outward curve represents the 'Semantic Centrifugal Force'—where the model hallucinates creativity while escaping reality).
The Code Behind the Collapse
The core logic uses modified Lorenz equations to simulate the "hallucination trajectory."
def update_simulation():
# Exponential decay of the Convex Hull
collapse_factor = exp(-0.2 * current_gen)
# The Drift Vector (Hallucination Path)
drift_magnitude = (current_gen**1.6) * 0.8
drift_vector = vector(sin(t)*drift_magnitude, cos(t)*0.5, -t)
# Real-time Semantic Corruption
if dist < event_horizon:
token.text = "NULL" # Total Semantic Loss
The Results
The visual data confirms that without fresh human input, AI models suffer from "Model Collapse" within 20 recursive generations. The "Convex Hull" (the geometric bounds of the AI's creativity) shrinks by 90%.
I have published the full findings and the mathematical proof on Zenodo.
Read the full paper: DOI: 10.5281/zenodo.18242108
Developed by Mahdi Al-Hajji | 2026