Learning Goals
3 min- Draw the nested circles: AI ⊃ ML ⊃ Deep Learning ⊃ Generative AI / LLMs.
- Define each ring with one sentence and one example.
- Explain why deep learning "took over" around 2012.
- Place tools you know (Siri, ChatGPT, a thermostat) in the right ring.
Warm-Up · The Nested Circles
5 min┌─────────────────────────────────────────────┐ │ ARTIFICIAL INTELLIGENCE │ │ any technique to make machines act "smart" │ │ ┌─────────────────────────────────────────┐ │ │ │ MACHINE LEARNING │ │ │ │ learns patterns from data │ │ │ │ ┌─────────────────────────────────────┐ │ │ │ │ │ DEEP LEARNING │ │ │ │ │ │ many-layered neural networks │ │ │ │ │ │ ┌─────────────────────────────────┐ │ │ │ │ │ │ │ GENERATIVE AI / LLMs │ │ │ │ │ │ │ │ ChatGPT, Claude, image models │ │ │ │ │ │ │ └─────────────────────────────────┘ │ │ │ │ │ └─────────────────────────────────────┘ │ │ │ └─────────────────────────────────────────┘ │ └─────────────────────────────────────────────┘
They're not competitors — they're Russian dolls. Every LLM is deep learning; every deep-learning model is ML; every ML system is AI. But not every AI is ML (some is just clever rules).
New Concept · Each Ring
14 minArtificial Intelligence (the outer ring)
Any technique that makes a machine seem intelligent — including hand-coded rules. A chess engine from the 1980s using pure search is AI but not ML. A thermostat with an if-statement is the simplest possible "AI".
Machine Learning
The subset of AI that learns from data instead of being told the rules. Everything in Lessons 9-20 (KNN, trees, regression) is "classic ML" — no neural networks needed.
Deep Learning
ML using neural networks with many layers. Each layer learns increasingly abstract features. Lessons 21-33 live here — Keras nets, image classifiers, face detection.
image → [edges] → [shapes] → [eyes, ears] → "cat"
layer1 layer2 layer3 outputGenerative AI / LLMs
Deep-learning models trained to generate new content — text, images, audio. Large Language Models (Claude, GPT) predict the next token over and over. Lessons 39-45 use them via APIs.
Why deep learning exploded ~2012
Three things arrived together:
- Data — the internet gave us billions of labelled images and texts.
- Compute — GPUs (built for games) turned out to be perfect for the matrix maths neural nets need.
- Algorithms — better training tricks (we'll meet some) made deep nets actually trainable.
In 2012 a deep net (AlexNet) crushed the ImageNet competition and the field never looked back.
Worked Example · Place the Tech
12 minTechnology Innermost ring it belongs to ───────────────────────── ────────────────────────────── A rule-based chatbot AI (not ML — just if/else) Spam filter (Naive Bayes) Machine Learning House-price regression Machine Learning A cat/dog image classifier Deep Learning Face unlock on your phone Deep Learning ChatGPT / Claude Generative AI (LLM) DALL·E / Midjourney Generative AI (image) A thermostat if-statement AI (the simplest kind)
Notice the pattern: the more fuzzy and high-dimensional the input (raw pixels, raw audio, free text), the deeper into the rings you go. Tabular data with clear columns? Classic ML usually wins — and is faster, cheaper, and easier to explain.
Reaching for deep learning or an LLM when a 5-line scikit-learn model would do better. For most tabular problems (Lessons 12-20), classic ML beats deep learning on accuracy, speed, and cost. Match the tool to the problem.
Try It Yourself
13 minReproduce the nested-circles diagram from memory, with one example in each ring.
Take 8 AI products you know and assign each to its innermost ring. Defend two of the trickier ones.
In 150 words, explain to a younger sibling why "AI suddenly got good" in the 2010s. Use the data/compute/algorithms framing.
Mini-Challenge · Tool-Picker Function
8 minWrite a Python function suggest_approach(problem) that takes a short description and returns a recommendation string based on simple keyword rules: structured/tabular → "classic ML"; images/audio → "deep learning"; free-text generation/chat → "LLM"; exact formula → "just write the rule".
Show one possible solution
# tool_picker.py def suggest_approach(problem): p = problem.lower() if any(w in p for w in ["formula", "exact", "convert", "validate"]): return "just write the rule — no ML needed" if any(w in p for w in ["image", "photo", "audio", "speech", "video"]): return "deep learning (CNN / pretrained model)" if any(w in p for w in ["chat", "summarise", "write", "translate", "explain"]): return "LLM API (Claude / GPT)" if any(w in p for w in ["table", "csv", "columns", "predict price", "classify rows"]): return "classic ML (scikit-learn)" return "unclear — gather data first, then re-decide" tests = [ "predict house price from a CSV of features", "detect a cat in a photo", "summarise a news article", "convert celsius to fahrenheit", ] for t in tests: print(f" {t:<45} → {suggest_approach(t)}")
Non-negotiables: at least four branches, a sensible fallback, tested on several inputs. The lesson is matching the tool to the problem, not always reaching for the fanciest one.
Recap
3 minAI ⊃ ML ⊃ Deep Learning ⊃ Generative AI. AI is any "smart" technique (including rules). ML learns from data. Deep learning uses many-layered neural nets for fuzzy, high-dimensional inputs. LLMs generate content. Deep learning exploded when data, compute, and algorithms aligned around 2012. Pick the smallest ring that solves your problem.
Vocabulary Card
- deep learning
- ML using neural networks with many layers, each learning more abstract features.
- neural network
- A model loosely inspired by brain neurons — layers of weighted connections.
- LLM
- Large Language Model — a deep net trained to predict the next token of text.
- generative AI
- Models that produce new content (text, images, audio) rather than just classifying.
Homework
4 minFind one news headline about AI from this month. Write a paragraph: which ring is the technology in, what data it learned from, and one risk. Bring it to the next session.
This is a reflection task — there's no single right answer. A strong response names the correct ring, identifies the training data honestly, and raises a specific (not generic) risk.