How does a machine learn?
Your phone used AI dozens of times today: face unlock, maps, your feed, the spam folder. By the end of this quest you’ll know how it works, what it’s called and which AWS service builds it.
This is your warm-up. After this quest, read the full Day 1 student guide: it will feel easy.
Your progress wheel: clear a level to light up a segment
The AI family
Goal: explain AI, ML, deep learning and generative AI, and how they fit together.
Four words you hear every day. They are circles inside circles. Tap each circle.
Tap a circle to see its definition
Why ML is different from normal programming
🧑💻 Traditional programming
IF email contains “lottery” THEN spam.
🤖 Machine learning
Show 100,000 labelled emails. It learns what spam looks like.
🏁 Checkpoint
Which is always true?
What makes deep learning “deep”?
Data, and 4 ways to learn
Goal: sort data into types, and match a problem to a learning type.
Labelled or unlabelled?
Labelled data
Every example has the answer. “This email is spam.” Slow and costly to create (often humans label it).
Unlabelled data
Inputs only, no answers. A million transactions with no tags.
Structured, semi-structured or unstructured?
Structured
Rows and columns. Excel, SQL tables, bank transactions.
Semi-structured
Tagged but flexible. JSON, XML, logs.
Unstructured
No fixed format. Text, images, audio, video. Most of the world’s data.
4 ways a machine learns. Tap to flip.
🏁 Checkpoint
50,000 X-rays, each marked “pneumonia” or “normal”, are…
Grouping customers when you have no labels is…
How a model learns
Goal: describe training, the ML process, and overfitting.
🔢 Feature
An input. House size, location, rooms.
🎯 Label
The answer to predict. The price.
🧠 Model
The learned rule. price ≈ f(size…)
The ML process is a loop
Never test on what you studied
📘 Training set, ~70–80%
Textbook exercises. The model learns here.
📝 Validation set, ~10–15%
Mock tests. Tune the settings.
🎓 Test set, ~10–15%
The real board exam. Used once, at the end.
Is 99% accuracy good? Not always.
Precision
When it says “fraud”, how often is it right? (Few false alarms.)
Recall
Of all real frauds, how many did it catch? (Few misses.)
🏁 Checkpoint
99% on training data, 70% on new data. This is…
Which set gives the final, fair score?
Using the model: inference
Goal: choose between batch and real-time inference.
Batch inference
Many records at once, on a schedule. Nobody is waiting. Cheaper.
Example: nightly churn scores for all customers.
SageMaker AI: Batch Transform.
Real-time inference
One request, answer in milliseconds. Always-on endpoint.
Example: card fraud check while you pay.
SageMaker AI: real-time endpoint.
Serverless inference
No servers to manage. Scales to zero. Good for spiky or low traffic.
Asynchronous inference
Big inputs, long processing (e.g. long videos). Result delivered when ready.
🏁 Checkpoint
During inference, the model’s weights…
A chatbot reply needs…
Deep learning: the digital brain
Goal: explain how a neural network learns, and name vision and language tasks.
| 🧠 Real brain | 🤖 Neural network |
|---|---|
| Dendrites receive signals | Inputs (features) |
| Synapse strength | Weights |
| Fires only if the signal is strong enough | Activation function (e.g. ReLU) |
| Practice strengthens connections | Training adjusts the weights |
How a network learns: 4 steps, repeated
Right network for the job
🖼️ CNN
Images. Edges → shapes → objects.
📜 RNN / LSTM
Sequences, one step at a time. Older.
⚡ Transformer
Text and more. Self-attention sees all words at once. Powers LLMs.
Computer vision (CV)
Classification, object detection, segmentation, face recognition, OCR. AWS: Rekognition, Textract.
Natural language processing (NLP)
Sentiment, entities, translation, speech-to-text, chatbots. AWS: Comprehend, Translate, Transcribe, Lex.
🏁 Checkpoint
In a neural network, synapse strength becomes…
Drawing boxes around every car and person in a photo is…
Generative AI: machines that create
Goal: explain foundation models, tokens, embeddings and diffusion.
The foundation model lifecycle
How AI creates images
t = 0
🏁 Checkpoint
An LLM writes by…
Reverse diffusion…
Getting better answers
Goal: choose between prompt engineering, RAG, fine-tuning and RLHF.
The ladder: start at the bottom
| Technique | What changes | Cost | Use when |
|---|---|---|---|
| 1. 🧪 Prompt engineering | Only your instructions | 💲 | Always first |
| 2. 📚 RAG | Model gets your documents at question time | 💲💲 | Private or changing facts, with sources |
| 3. 🎓 Fine-tuning | Model weights, with labelled examples | 💲💲💲 | Consistent style, format or skill |
| 4. 📖 Continued pre-training | Weights, with lots of domain text | 💲💲💲💲 | New domain language (legal, medical) |
Instruction fine-tuning
Train on “instruction → ideal answer” pairs. Turns a base model into an assistant.
RLHF
Humans rank answers → a reward model learns their taste → the model improves.
On AWS: Amazon Bedrock Knowledge Bases does all five steps for you.
🎓 Fine-tuning = send the model back to school (changes how it behaves). 📚 RAG = hand it the right textbook page during the exam (changes what it knows today).
🏁 Checkpoint
Company policies change every month. The chatbot must use the latest ones. Use…
For consistent, factual answers, set temperature…
The AWS toolbox
Goal: pick the right AWS service, and know what drives the bill.
AI services
Order from a restaurant: call an API, no ML skills. Rekognition, Polly, Comprehend…
ML platforms
A full kitchen: build your own. SageMaker AI, Bedrock.
Infrastructure
Raw ingredients: GPUs, Trainium (training), Inferentia (inference).
Memory hooks. Tap to flip.
Generative AI on AWS
Amazon Bedrock
Many foundation models (Nova, Claude, Llama…) through one API. No servers. Plus Knowledge Bases (RAG), Agents, Guardrails.
SageMaker JumpStart
Hub of pre-trained models inside SageMaker AI. Deploy and fine-tune on your own endpoint: more control.
Amazon Q Business
Assistant for employees, connected to company data.
Amazon Q Developer
Helps developers and cloud engineers, e.g. in the AWS Console.
• Amazon Q Developer IDE plugins: no new sign-ups since 15 May 2026, support ends 30 April 2027. Replacement: Kiro. Q Developer in the AWS Console continues.
• AWS DeepRacer: the console ended in December 2025. It lives on as an open-source solution you deploy in your own account.
Tokens per month:
Estimated cost per month:
🔤 Tokens
Input + output. Keep prompts lean.
📏 Model size
Smallest model that’s good enough.
📦 Batch
Much cheaper for non-urgent jobs.
🔒 Provisioned throughput
Reserved capacity for steady, heavy use.
💤 Idle endpoints
Real-time endpoints bill while running.
🎓 Customisation
Fine-tuning adds cost. Try prompts and RAG first.
🏁 Checkpoint
Speech → text is…
Call foundation models through one API with no servers to manage:
More tutorials
Two extra pages. Your XP carries over.
Rise of the AI Agents: from talking to doing
Bonus level: agents, the think–act–observe loop, tools, safety and agents on AWS. About 12 minutes.
Open the bonus levelThe Goldilocks Problem: how does an AI know when to stop learning?
Video tutorial on the bias-variance trade-off, with a scene-by-scene walkthrough and games. About 10 minutes.
Watch the video tutorialThe big picture: mind map
Everything on one page. Click a branch to jump to its level.
Flashcards
Tap to flip. Mark each card. Keep going until the “review again” pile is empty.
Myth or fact?
Quick-fire. Decide fast.
Boss battle
10 questions. 3 hearts. Lose a heart for each wrong answer. Can you beat the Overfit Monster?
Quest progress
AI & ML Explorer
📖 Next step: read the Day 1 student guide
You now know the map. In the guide, look for:
- The 88 questions answered from basic to advanced, one module at a time
- The ML process table with the SageMaker AI tool for each step
- The comparison of prompt engineering, RAG and fine-tuning
- The full AWS service table and the cost considerations
- The Final challenge: aim for 10 out of 12