Socratic Question-Driven Learning

Queries by Charan

Input Text: "The cat sleeps"
Tokenizer: Splits text into 3 tokens: ["The", "cat", "sleeps"]
Token IDs: Maps tokens to row addresses: [101, 3797, 6251]
Embedding Lookup: Selects Row 101, Row 3797, and Row 6251 from the learned embedding matrix.
Input Matrix Stacking: Stacks the 3 selected vectors in sequence order, forming a 3 × 1,536 matrix.
Positional Encoding: Injects token position information so the model knows "The" is first and "sleeps" is third.
Transformer Layers: The $3 \times 1,536$ tensor enters multi-head self-attention layers.

Understand the deep architecture of AI by asking the right conceptual questions. Each question breaks down intuition using Feynman-style analogies, diagrams, and direct links to the interactive AI roadmap.

— queries — topics 100% Feynman explanations ▶️ #charan Visual Masterclass ↗

🔍

🔭

No queries found matching your search.
Try searching for keywords like "dimension", "token", "attention", "RAG", or "matrix".