learn-ai-literacy-with-phoebe / Session 1 of 8
Learn AI Literacy with Phoebe · Session 1 of 8

What AI actually is: the nested dolls

Everyone says "AI" and means five different things. This session gives you the map: how AI, machine learning, deep learning and generative AI nest inside each other, the three eras that got us here, and the single most useful question a leader can ask - is this even an AI problem? The sequel to the data literacy course; here the machines start to learn.

🟢 Foundation Leaders & practitioners No code 45 min live + self-study
0-3 · Welcome 3-18 · Concepts 18-40 · Apply-along 40-45 · Q&A
Part 0

Why this course exists

The data literacy course ended at a wall: deep learning, recommendations, search, and generative AI were "out of scope, coming next". This is next. AI literacy is not learning to build models - it is knowing what these systems actually do, why they work, where they fail, and when the answer is "do not use AI for this". Over 8 sessions you will look inside neural nets, embeddings, recommenders, search, and LLMs - each with a hands-on toy, no code anywhere. Leaders get the judgment; practitioners get a light second lane of depth marked Self-study.

Live - presented in session Self-study - practitioner depth ★ Apply-along - interactive, everyone plays Sequel to learn-data-literacy
★ What you walk out with today The nested-dolls model (AI > ML > deep learning > GenAI) so no vendor pitch can confuse you again, the three-era story of how we got here, and the right-tool test - the question that saves more money than any model: "is this actually an AI problem, or would a rule or a spreadsheet do?"
Part 1 · the words untangled

AI, ML, deep learning, GenAI 5 min live

These four words are not synonyms and they are not rivals - they nest. Every generative AI is a deep learning model; every deep learning model is machine learning; all machine learning is AI. But not the reverse. Getting the nesting right is the first defense against hype.

Artificial intelligence any machine mimicking smart behavior Machine learning learns patterns from data Deep learning many-layered neural nets Generative AI makes new content The which-doll test: Learning from data? → ML. Deep neural nets? → deep learning. Creating new content? → GenAI. None of these? → just a rule.
🔍 Click to zoom - four words that nest, not compete
LiveThe nesting, and why it defuses hype3 min▶
  • AI is the broad goal: machines doing things that seem intelligent. A thermostat and a chess rulebook technically qualify - "AI" alone tells you almost nothing.
  • Machine learning narrows it: systems that learn patterns from data instead of following hand-written rules. This is where the data literacy course lived (regression, clustering, churn models).
  • Deep learning is ML using many-layered neural networks - the engine behind modern vision, speech, and language. Session 2 opens it up.
  • Generative AI is deep learning that creates new content (text, images, code). ChatGPT and image generators live in this innermost doll. It is the smallest circle, not the whole field.
Real world

When a vendor says "our product is powered by AI", the literate question is "which doll?". A fraud filter might be simple ML (great, explainable, cheap). A "smart assistant" might be a wrapped LLM (powerful, but can hallucinate). The word "AI" is marketing; the doll is the fact.

Self-studyNarrow vs general: the AGI myth-check2 min read▶

Every AI in production today is narrow: brilliant at one task, useless outside it. The chess engine cannot drive; the language model cannot truly reason about physics it has not seen. Artificial general intelligence (AGI) - one system matching humans across any task - does not exist and its timeline is genuinely contested among serious researchers.

  • Modern LLMs feel general because language touches everything, but they are still pattern machines with sharp edges (Session 6).
  • The leadership takeaway: buy and deploy for narrow, well-defined tasks with clear success metrics. "It will figure it out" is not a spec.
Part 2 · how we got here

Three eras of AI 5 min live

AI is not new - it is on its third act. Knowing the arc tells you why this moment feels different, and why "AI winters" happened before (over-promise, under-deliver, funding collapse). Today's era is real, but the pattern is worth remembering.

1 · Rules (1950s-80s) experts hand-code every if-then brittle; breaks on the unexpected "expert systems", then winter 2 · Learning (1990s-2010s) machines learn from labeled data deep learning breaks out ~2012 vision, speech, translation 3 · Foundation (2018-now) one huge model, many tasks transformers + scale + GenAI the era you are living in Each era did not replace the last - rules and classic ML still run most of the world. The newest doll is the loudest, not the biggest by usage. The lesson of the two winters: value follows working deployments, not demos. Session 8 returns to separating ROI from hype.
🔍 Click to zoom - three acts, two winters, one throughline
LiveWhy 2012 and 2018 were hinge years3 min▶
  • 2012 - deep learning breaks out. A neural net (AlexNet) crushed the ImageNet vision contest. The ingredients were not new ideas but new scale: massive labeled data + GPU compute. Session 2 tells this story.
  • 2017-2018 - the transformer + foundation models. A new architecture (the transformer) let models train on the entire internet and handle any language task. GPT and its cousins followed. Session 6 opens this up.
  • The throughline: progress came from scale (data + compute) as much as cleverness. That is why "just add more data" is a real strategy - and why compute cost is now a boardroom line item (Session 8).
The exec move When someone pitches an AI breakthrough, ask "what is genuinely new - the idea, or the scale?" Often it is scale, which means the real questions are cost, data access, and defensibility, not magic.
Part 3 · the money-saving question

Is this even an AI problem? 5 min live

The most valuable AI skill for a leader is knowing when NOT to use it. AI is probabilistic, expensive, and hard to explain. If a rule, a lookup, or a simple calculation solves the problem, that is not a consolation prize - it is the right answer. Monica Rogati's AI Hierarchy of Needs (from the data literacy course) says it plainly: you need the boring foundation before the fancy top.

LiveThe right-tool test, and the hierarchy under it3 min▶
  • Rules beat AI when the logic is known and stable: tax brackets, shipping thresholds, "flag orders over $10k". Deterministic, explainable, free. Do not train a model to do arithmetic.
  • Classic ML beats GenAI when you have structured data and a clear target: churn, forecasting, fraud scoring. Cheaper, faster, auditable (the whole data literacy course).
  • Reach for deep learning / GenAI when the input is messy and human: free text, images, speech, open-ended generation. That is where the newest dolls earn their cost.
  • The hierarchy: reliable data collection and cleaning (data literacy) sit UNDER any AI. Most failed AI projects are actually failed data projects wearing a fancy hat.
Real world

A retailer wanted "AI" to decide reorder quantities. A two-line rule (reorder when stock < 2 weeks of average sales) captured most of the value in an afternoon, explainable to every store manager. AI came later, for the genuinely hard 10% (promotions, weather, new products). Right tool first, model second.

Self-studyThe cost of using AI where you should not2 min read▶

Reaching for AI reflexively has real costs beyond the bill: you inherit unpredictability (it can be wrong in surprising ways), opacity (hard to explain to auditors or customers), and a maintenance burden (models drift, prompts break, vendors change). A rule you can read in one line has none of these. The literate default is: simplest tool that meets the bar, escalate only when it does not. Session 8 turns this into a use-case scorecard.

Apply-along 1 of 3

Which doll is it? ★ 8 min · everyone sorts

Eight systems from real life. Place each in the smallest doll that describes it - and notice which ones are not AI at all.

The debrief Two of eight were not AI at all - and that is the point. A data-literate organization reaches for the smallest doll that works. "Can we use AI for this?" should always be preceded by "should we?".
Apply-along 2 of 3

Audit an AI claim ★ 7 min · everyone digs

Find a real "AI-powered" claim - a product page, a vendor email, a headline you saw this week. Read it aloud.

Ask the three literate questions: which doll is it really (rule / ML / deep / GenAI)? what is genuinely new (idea or scale)? and what task exactly - narrow and measurable, or vague?

Rate the claim: substantive, or "AI" as a sticker on something simpler? Most land in between - and now you can say where and why.

Share one. The room will notice how often "AI" is a rule with better marketing - and occasionally, something genuinely deep.

Apply-along 3 of 3

Run the right-tool test on your own backlog ★ 7 min · everyone decides

List 3 things someone in your org has proposed to "use AI" for (or that you are curious about).

For each, walk the ladder: could a rule do it? classic ML on data you have? does it genuinely need deep learning or GenAI (messy, human input)?

Mark the honest verdict. Flag any that are really data problems in disguise (no clean data = no AI, whatever the doll).

Keep this list. Session 8 turns it into a scored use-case matrix; the sessions between explain the dolls you flagged.

Real world

Teams that run this test before the vendor call save quarters of wasted pilots. "We tried AI and it did not work" is usually "we pointed a model at a problem a rule owned, on data we had not cleaned". The test is cheap insurance.

After the session

This week ◐ 25 min total

Check yourself

Three questions before you go 🎯 ◐ 90 seconds

1 · A vendor says their tool is "powered by generative AI". What does that tell you about the nesting?

GenAI is the smallest circle inside deep learning inside ML inside AI. Powerful for messy, generative tasks - not a synonym for "the best AI".

2 · A team wants to "use AI" to calculate shipping cost from weight and destination zone. Best response?

The right-tool test: known, stable logic belongs in a rule. Never train (or prompt) a model to do arithmetic a formula nails for free.

3 · What mostly caused the 2012 deep-learning breakout?

The ideas were decades old; data and GPU scale made them work. That is why "is it the idea or the scale?" is such a sharp question.

Source material

Concepts covered & where they come from

This course teaches the working intuition of each idea and credits the canon. Leaders get judgment; practitioners get the Self-study lane and the names to go deeper.

✓AI / ML / deep learning / GenAI nestingthe four dolls + the which-doll test - Part 1 + Apply-along 1
✓The three eras (rules / learning / foundation)the arc, the two winters, why 2012 & 2018 mattered - Part 2
✓The right-tool testwhen NOT to use AI - Part 3 + Apply-alongs 2 and 3
◐AI Hierarchy of Needs (Monica Rogati)data foundation under any AI - Part 3, callback to data literacy
◐Narrow vs general / AGIself-study card - the honest state of the field

Session 1 cheat sheet · pin this

The nested dollsAI > machine learning > deep learning > generative AI. Each is inside the last. "AI" alone tells you nothing - ask which doll.
Which doll testLearning from data? ML. Deep neural nets? deep learning. Creating content? GenAI. None of these? just a rule.
Three erasRules (hand-coded) → learning (from data, deep learning ~2012) → foundation models (transformers + scale, ~2018).
Idea or scale?Most recent leaps came from more data + compute, not new math. So ask about cost, data, defensibility.
The right-tool testRule > classic ML > deep/GenAI. Use the smallest that works. Two of eight "AI" tasks were not AI.
Hierarchy of needsClean, reliable data sits under any AI. Most failed AI projects are failed data projects in disguise.