Unit 1: Key Concepts

This unit covers what AI systems are, how they learn from data, and why they behave the way they do. You’ll train a small model yourself with Teachable Machine, and try to talk a chatbot out of its secret password in Hack Merlin.

For this unit, you’ll need:

  • A computer with a modern web browser
  • A webcam or microphone for the Teachable Machine activity
  • Distinguish key categories: artificial intelligence, machine learning, large language models, and generative AI
  • Describe, at a high level, how neural networks and language models work (weights, tokens, embeddings, attention, transformers), and how an LLM generates text
  • Explain, at a high level, how models are trained, what training data is, and why it matters
  • Recognize the difference between supervised, unsupervised, and reinforcement learning
  • Apply basic evaluation concepts (training/test split, overfitting, accuracy, precision, recall, F1) in a research-appropriate way
  • Explain why LLM guardrails are brittle, and what prompt injection demonstrates about model behaviour
  • Explain why LLMs produce fluent but unreliable output, why their answers vary, and why benchmarks aren’t enough to judge them

Contents

  1. Reading: Basic Concepts
    • Terminology: AI, machine learning, deep learning, generative AI, and LLMs
    • From rules to examples: how machine learning differs from traditional software
    • Neural networks and weights
    • Types of machine learning: supervised, unsupervised, and reinforcement learning
  2. Activity: Teachable Machine: training a model from scratch
  3. Reading: Model Training
    • The training loop: loss, gradient descent, and epochs
    • Overfitting and generalization
    • Testing: training, development, and test sets
    • Measuring performance: accuracy, precision, recall, and F1 (with a worked example)
    • The black box problem, and how bias enters through data collection, labelling, and model design
  4. Reading: Language Models
    • Tokens, and how names, old spellings, and other languages are split
    • Word embeddings: turning tokens into numbers
    • From sequence to sequence: RNNs, LSTMs, and transformers
    • Text generation: next-token probabilities, sampling, and temperature
  5. Activity: Hack Merlin: prompt injection and LLM guardrails
  6. Reading: LLM Behaviour
    • Plausibility vs. truth: hallucination, confidence, and variation between answers
    • Learned patterns all the way down: why guardrails can be talked around
    • You can see what a model does, not why
    • The limits of benchmarks, and fluency without correctness