All courses
Online · Self-paced · IXILab

Human-AI Interaction and Collaboration

A virtual online course on how people and computers work together, played as short browser games. Walk a map, meet trainers, try live demos, and pass each leader’s quiz.

Course info
Format
Browser games, one per chapter. Nothing to install.
Length
20–45 minutes per chapter
Author
Dr. Yi-Chi Liao (yichi@csie.ntu.edu.tw)
Based on
The NTU course Human-AI Interaction & Collaboration (CSIE 5648)
Language
English
Audience
Anyone. No prior knowledge of HCI is assumed.
Cost
Free
How to play

Start at Chapter 01!

Walk with the arrow keys or WASD, talk with Space or Enter, and answer with 1–4. On a phone, use the on-screen pad. Your progress is saved in your browser.

Objectives

After the course, you should be able to…

  1. Define Human-Computer Interaction (HCI).
  2. Name the key elements in the model of human-computer interaction, from sensing input to how people perceive, decide and learn.
  3. Name the key phases of the design process.
  4. Run a short design sprint, from a design brief to testing a paper prototype.
  5. Define research, evaluation and controlled experiments in HCI.
  6. Plan and run a controlled experiment to evaluate a working system, and analyze it with ANOVA.
  7. Know different research methods, and when to use them.
  8. Explain how a neural network maps input to output, and how it learns with gradient descent and backpropagation.
  9. Explain how language becomes tokens and embeddings, how attention with queries, keys and values builds context, and how self-supervised learning trains large language models.
  10. Harness AI models without updating them, through prompt engineering, context engineering, inference-time strategies, and prompt evaluation.
Chapters

Five parts, twenty-two chapters, three areas each.

Every trainer teaches one idea with a live demo, and each area ends with a leader’s quiz. Select a chapter to play it. More parts will be added.

Part 1 · Introduction to Human-Computer Interaction

Part 2 · Research methods and evaluation

Part 3 · Deep learning

Chapter 08
~30 min

Neural networks

  • Pixel Path AI, ML and deep learning, a network as a function, and the input layer as 784 pixel activations
  • Feature Forest Hidden and output layers, learned features, weights and the weighted sum, and a line detector
  • Sigmoid Shore The sigmoid activation, the bias, and the matrix form σ(Wx + b)
Play chapter 08
Chapter 09
~30 min

Gradient descent

  • Label Lane How many parameters a network has, training data and labels, and accuracy during training
  • Cost Canyon The cost function, learning as minimizing it, and why we step against the gradient
  • Gradient Slope Gradient descent with a learning rate, reading the gradient vector, and what a trained network really learns
Play chapter 09
Chapter 10
~20 min

The idea of backpropagation

  • Nudge Meadow What backpropagation computes, and three ways to raise an output: bias, weights, or the previous layer
  • Wishing Wells Which weights to raise most, which way each activation should move, and summing every output’s wishes
  • Backflow River Wishes flowing back layer by layer, and full vs. stochastic (minibatch) gradient descent
Play chapter 10
Chapter 11
~30 min

Backpropagation in math

  • Sensitivity Steps The cost of one example, what depends on what, and how sensitive the cost is to a weight
  • Chain Link Bridge The chain rule, the derivatives for the bias and the previous activation, and passing them back a layer
  • Index Cliffs Many neurons (indices j, k), averaging over examples, the gradient vector, and the update w ← w − η∇C
Play chapter 11
Chapter 12
~40 min

Computer vision models

  • Keypoint Knoll Images as numbers, classic feature extraction and matching, and how CNNs and deep learning took over
  • Kernel Creek Fully connected vs. convolutional layers, a CNN layer by layer, and image classification
  • Bounding Box Bay Object detection (R-CNN, YOLO), hand-designed vs. learned features, and the family from CNNs to foundation models
Play chapter 12

Part 4 · Modern AI models

Chapter 13
~25 min

Large language models

  • Prompt Path What GPT stands for, next-token prediction, and how a chat is one growing text
  • Corpus Canyon The scale of training data and compute, and the three steps from pretraining to harnessing
  • Gap Gorge Next-token training pairs, and what GPT-3 does differently from earlier deep learning
Play chapter 13
Chapter 14
~25 min

Embeddings

  • One-Hot Hill Turning words into numbers: single numbers, one-hot vectors, and feature vectors that capture similarity
  • Token Thicket Learning vectors from context, and what to embed: words, characters, or subword tokens
  • Matrix Marsh Token IDs, the embedding matrix, and vector arithmetic like Paris − France + Italy ≈ Rome
Play chapter 14
Chapter 15
~25 min

Attention

  • Twin Bank Ford Why one word needs different meanings in different contexts: tokens, context, and contextualized embeddings
  • Mixing Meadow Building meaning from context: “female king”, “Lego Eiffel tower”, and which words help
  • Spotlight Ridge Attention weights, a weighted sum of context, and representations across layers
Play chapter 15
Chapter 16
~25 min

Query, key and value

  • Query Cove Which tokens are relevant: queries, keys, dot products and softmax
  • Value Ridge Values, a full attention calculator, and where Q, K and V come from
  • Root Lake The attention formula softmax(QKT / √dk) V, why we scale by √dk, and stacking layers
Play chapter 16
Chapter 17
~25 min

Self-supervised learning and foundation models

  • Nobody’s Orchard Where the training data come from: supervised vs. self-supervised learning
  • Foundation Quarry Transformers at scale, foundation models vs. one model per task, and handling a new task
  • Horizon Ridge Beyond text-only LLMs (multimodal models), and how GPT-3 learns without labels and writes one token at a time
Play chapter 17

Part 5 · Harnessing AI models

Chapter 18
~45 min

From models to applications

  • Ticket Terrace The gap between a model and an application, and five ways to make a model do what we want
  • Misfire Marsh Harness failures: unclear instructions, missing or conflicting context, unusable output, and their fixes
  • Token Budget Bluff The context window, budgeting its tokens, and why more context is not always better
Play chapter 18
Chapter 19
~45 min

Prompt engineering

  • Instruction Inlet Prompts vs. instructions, why prompting works, and where prompts come from
  • Hierarchy Heights The prompt hierarchy, prompt injection (with the 2025 hidden-text case), and when system prompts fail
  • Template Terrace Structured prompts, their typical sections, and prompt templates
Play chapter 19
Chapter 20
~45 min

Context engineering

  • Window Wharf When a prompt is not enough, choosing which context goes in, and the context budget
  • Digest Dunes Compressing context: extractive and abstractive summaries, and LLMLingua
  • Retrieval Reach In-context learning with examples, what GPT-3 found about few-shot learning, and retrieval-augmented generation (RAG)
Play chapter 20
Chapter 21
~25 min

Inference-time strategies

  • Checkpoint Hill Inference-time strategies, and generating hypotheses then verifying them
  • Thought Trail Chain of thought (few-shot and zero-shot), when it helps, and self-consistency voting
  • Schema Sound Structured outputs: JSON schemas, validation, and constrained decoding
Play chapter 21
Chapter 22
~25 min

Prompt evaluation and optimization

  • Test Set Terrace Which prompt is better? Building a test set with expected outputs, and measuring accuracy
  • Judges’ Lagoon Prompt optimization, judging outputs when there is no metric, and the optimization loop
  • Rewrite Rise Textual gradients: learning from the difference between a better and a worse prompt
Play chapter 22