Ainex-Limit-Experiment

The Ainex-Limit-Experiment provides a groundbreaking examination of AI model integrity and the phenomena of semantic collapse using GPT-2 Small (124M) through a series of recursive synthetic experiments. The project introduces the Ainex Integrity Score (denoted as $cA$), a novel geometric metric that evaluates the integrity of an AI’s generated output, going beyond traditional perplexity measures that fail to capture semantic fidelity.

Key Findings

• Iterations: 20
• Dataset: Recursive Synthetic (Self-Feeding)
• The experiment revealed a 66.86% loss of semantic accuracy by Generation 20, culminating in a model that exhibits permanent hallucinations accepting incorrect outputs as truth.

The Mathematical Proof

The study presents the Ainex Equation: $$ \mathcal{A}{gen} = \frac{V{hull}(\mathbb{R}^3)}{1 + \lambda \cdot || \Delta \mu ||^2} $$
Where:

• $V_{hull}(\mathbb{R}^3)$ measures the explorable semantic volume in 3D PCA space and indicates creativity.
• $||\Delta \mu ||$ quantifies the Euclidean drift from the human centroid, serving as an indicator of hallucination.

Observations

1. Phase 1 (Gen 0-5): Characterized by volumetric implosion; the model loses approximately 85% of variance, resulting in overly conservative and repetitive outputs.
2. Phase 2 (Gen 5-20): Displays a linear drift away from human construct norms, leading to a transition to a logic-less framework.

Case Study: The "Crocodile" Artifact

Analyzing a static prompt, "The fundamental laws of physics...", illustrates the semantic disintegration of the model:

This study conclusively demonstrates how synthetic data environment leads to the emergence of false axioms within model outputs.

Results Visualization

A 3D PCA projection visualization further emphasizes the transformation between outputs across generations, showcasing the striking contrast between human structure and the degenerated manifold.

Replication of Results

This repository offers a self-contained environment to enable replication of the experiments demonstrating model collapse.

Requirements:

• Hardware: Recommended T4 GPU or better, though CPU can be used with extended processing time.
• Key Dependencies: torch, transformers, scipy, sklearn, rich.

Usage

To replicate the findings:

# Clone the repository
git clone https://github.com/mhh1430hacker/Ainex-Limit-Experiment
cd Ainex-Limit-Experiment

# Install dependencies
pip install -r requirements.txt

# Ignite the collapse loop
python main.py

This research provides a foundational understanding of AI model behavior and the critical importance of semantic integrity, opening new avenues for future studies in artificial intelligence and machine learning.