Understanding Large Language Models

Lesson 2, AI Foundations Module

This session opened the hood on how ChatGPT, Claude, and similar tools actually work. We explored the three-stage process: training (learning patterns from billions of texts), processing (breaking your input into tokens), and generation (predicting the most likely next word, thousands of times). The crucial concept of tokens – roughly 0.75 words each – helps explain why AI tools have usage limits and costs.

We also tackled context windows: AI’s “working memory” that fills up during long conversations, causing it to forget earlier information. Perhaps most importantly, we confronted the hallucination problem – AI confidently generating false information because it predicts likely text patterns, not factual truth. Through exercises testing made-up product names and companies, participants learned to recognise when AI invents plausible-sounding but entirely fictional details. The lesson? LLMs are extraordinary pattern prediction machines, but they require human verification, especially for specific facts, dates, and citations.

This post is part of the Cambridge AI Lunch & Learn series, a 24-week upskilling programme designed to transform participants from AI curious to AI confident. Through bite-sized
30-minute weekly sessions, we’re building practical AI literacy across Cambridge University Press & Assessment, equipping our team with the skills to work effectively and ethically with artificial intelligence.

Each session follows a proven learn-by-doing approach: 10 minutes of instruction, 10 minutes of demonstration, and 10 minutes of hands-on practice. This isn’t about becoming AI experts but rather it’s about understanding how to augment our work, make better decisions, and navigate the evolving landscape of AI tools with confidence and responsibility.

Build Your AI Foundation

Leave a Reply