Lab 02: The Flat Memory Illusion — Why LLMs Fail to Separate Instructions from Data
Attention hijacking simulation (QK^T), context saturation, and evaluating XML delimiter guardrails versus traditional x86 NX-bit architectures.
Key Lab Topics
- Matriz de Atención
- Delimitadores XML
- Dilución de Contexto
- Mitigación Determinista
Prerequisites
- Python 3.10+
- PyTorch / Transformers
Resources & Architecture Guide
This hands-on lab runs in a Jupyter Notebook (lab-02-memoria-plana.ipynb) and includes a comprehensive architectural guide: