AI Wednesday 02 : Under the Hood of Modern LLMs

Overview

This week’s AI Wednesday session was about understanding how modern AI models like ChatGPT actually work behind the scenes. Instead of treating AI as something magical or human-like, we broke it down into simpler ideas and mechanisms.

We started by clearing a common misconception that these models “think” or “understand” things. The discussion focused on how LLMs are trained to predict the next piece of text, and how doing this at a very large scale creates the illusion of intelligence.

The session was mostly conceptual and discussion-driven. The goal was not to teach people how to build models, but to help everyone develop a clear mental model of what is happening inside these systems.

Topics

• Why LLMs do not think like humans and what they actually do instead
• What “next token prediction” means in simple terms
• How text is broken into tokens and converted into numbers
• What embeddings are and how they represent meaning
• A brief look at how language models evolved from simple word prediction to transformers
• What attention is and why transformers made modern AI possible
• Some important limitations of LLMs and why they sometimes fail

Photos

Highlights

• A major takeaway was realizing that modern AI feels intelligent not because it understands, but because it has learned patterns from huge amounts of human language.