DAY 04 · AI & ML ON GOOGLE CLOUD

From "what is AI" to building one. 🛠️

Today we lay the foundation: the difference between AI types, how Google Cloud is built for AI, how models learn — and you finish by training your first ML model in SQL.

Today — We Start With Module 1 🗺️

Module 1 is the whole AI Foundations course, distilled. It maps 1-to-1 to the official Google Skills Boost chapters.

☕

1 · Use Case

AI solving real business problems

🔮

2 · AI on GCP

Predictive vs Generative + the architecture

🏗️

3 · Infrastructure

Compute, storage, TPUs

🧠

4 · AI Models

Supervised vs unsupervised

🪄

5 · BigQuery ML

Train a model with SQL

🔜

Modules 2–4

Generative AI · added next

Five chapters, one module. Each ends with a quick check. Use ← → to move. Aim for 100+ ⭐.

Chapter 1 · A Use Case

Meet "Coffee on Wheels" ☕🚚

An international company selling coffee from trucks in London, New York, San Francisco & Tokyo. They hired a data company (Data Beans) to fix three problems with AI:

📍

Location & Routes

Where to park trucks; best routes given weather & traffic

📈

Sales Forecasting

Predict sales & monitor performance in real time

📣

Marketing

Auto-generate campaigns — text & images

Why start here? Every abstract concept today — prediction, generation, BigQuery, Vertex AI — shows up in this one app. Keep the coffee trucks in mind.

What the App Actually Does 🖥️

Data Beans built a single dashboard. Pick a city — say London — and it shows revenue, operating margin and the number of trucks. From there, AI is doing five jobs at once:

🗺️

Smart routing

Cold day forecast? It reroutes trucks to covered areas. Football match nearby? It avoids the congestion.

🚚

Per-truck insight

Click a truck for its street view and a revenue forecast for that spot.

🍩

Live menu watch

An item underperforming? Hit Generate and AI suggests a replacement.

📣

One-click marketing

Auto-creates a campaign — text + images — and emails it to the right customers.

📑

Instant reports

Generates an operational report and exports it straight to Google Slides.

Notice the mix: some features predict (sales, routes), others generate (menus, campaigns). That's the whole of AI in one app.

Key Findings From the Use Case 🔑

Strip away the coffee and three lessons remain — they hold for almost any AI project:

🔀

Predictive + Generative together

The best apps combine both — forecast and create. Rarely just one.

🧩

Products collaborate

Gemini acquires & generates · BigQuery analyses · Vertex AI models · agents act (e.g. push to Looker).

⚡

Outcome: efficiency

Manual decisions became automated — freeing the team to innovate on the customer experience.

The takeawayYou don't need to invent new AI — you orchestrate existing Google products around a real business problem. You can do the same for your work.

Hold this question: which parts were predictive and which were generative? Chapter 2 makes that split exact.

How One App Pulls It Off 🔧

The pattern behind the whole demo — and behind most real AI apps:

IN

Multimodal Input

Text, reviews, images, live street view

→

BRAIN

Predict & Generate

Forecast sales · write campaigns

→

OUT

Visual Output

Dashboards & reports to act on

The Google products doing the work

💎

Gemini

Multimodal input & generation

🗄️

BigQuery

Data analytics

🏭

Vertex AI

ML development

📊

Looker

Visualise insights

Try It · Chapter 1

Match Each Google Product to Its Job 🧩

🕹️Drag each product onto the job it does in the app. Goal: see how Gemini, BigQuery, Vertex AI & Looker split the work. Correct drop → +5⭐.

Gemini

BigQuery

Vertex AI

Looker

💎 Understands images + writes content

🗄️ Stores & analyses the data

🏭 Builds & serves the ML models

📊 Turns insights into dashboards

Memory hook: Gemini talks, BigQuery stores, Vertex AI models, Looker shows.

Chapter 2 · AI on Google Cloud

Why Google for AI? 🟦🟥🟨🟩

📜

Heritage

AI has powered Search, Maps & Workspace for years

🚀

Innovation

Gemini, Vertex AI, NotebookLM — frontier tools

⚖️

Responsible AI

Bold but safe: fair, accountable, transparent

Responsible AI isn't a footnote. Google frames every product around it — bold innovation, responsible deployment, collaborative progress. Expect it on the exam.

The Big One: Predictive vs Generative AI ⭐

Almost every AI problem is one of these two. Knowing which is which is half of today.

🔮 Predictive AI

analyse & predict

Also called traditional or discriminative AI
Learns from existing data to classify or forecast
"What will happen?" / "Which group is this?"
e.g. sales forecasting, churn prediction, route/traffic prediction

✨ Generative AI

create & take action

Creates new content: text, images, video, summaries
Built on foundation models (e.g. Gemini)
"Make me…" / "Write…" / "Summarise…"
e.g. marketing copy, chatbots, campaign images

One line to rememberPredictive AI analyses and predicts. Generative AI creates and acts.

You Often Use Both Together 🤝

There's no hard line. The strongest apps chain them — predictive output becomes part of the generative prompt.

STEP 1

Predictive

Forecast which customers will churn

→

STEP 2

Generative

Chatbot helps sales explore those predictions

STEP 1

Predictive

Identify customer segments

→

STEP 2

Generative

Write personalised marketing per segment

Decision shortcut: need a number or a category → predictive. Need new content or automation → generative.

Try It · Chapter 2

Predictive or Generative? 🔮✨

🕹️Drag each task into Predictive or Generative. Goal: train your ear for the trigger words. Correct → +5⭐.

Forecast next month's coffee sales

Write a promo email with an image

Flag which customers may churn

Summarise 500 customer reviews

Predict tomorrow's truck traffic

Generate a chatbot reply

🔮 Predictive AI

✨ Generative AI

Trigger words: "forecast / predict / flag / classify" → predictive. "write / summarise / generate / reply" → generative.

The Google Cloud AI Architecture 🏛️

Three layers, bottom to top. Each layer serves a different kind of user.

· Applications & Solutions out-of-the-box · for business users & analysts
· AI Development Vertex AI · Gemini · BigQuery · for developers & data scientists
· AI Infrastructure compute · network · storage · the solid ground

Read it like a building. Infrastructure is the foundation, Development is where you build, Applications is the storefront. We spend the rest of Module 1 on layers 1 & 2.

Chapter 3 · AI Infrastructure

Inside the Foundation Layer 🏗️

Infrastructure itself has three sub-layers:

Data & AI products BigQuery · Vertex AI · Looker
Compute & Storage decoupled — scale independently
Networking & Security the base under everything

Key idea: decoupled compute & storageUnlike your laptop, the cloud scales compute and storage separately. Need more processing but the same data? Scale only compute. This is the superpower of cloud computing.

Compute: Control ↔ Convenience ⚙️

Pick how much you manage vs how much Google manages.

MOST CONTROL

Compute Engine

Like running your own server (VMs)

→

BALANCE

GKE

Managed containers + orchestration

→

MOST CONVENIENCE

Cloud Run

Serverless — Google runs it all

The chips that power it

🖥️

CPU

General-purpose, everyday computing

🎮

GPU

Parallel work — graphics & some ML

⚡

TPU

Google's custom AI chip (2016). Built for the matrix math in ML — faster & more energy-efficient

Storage & the Data-to-AI Workflow 🗄️

📦 Unstructured data

no fixed shape

Documents, images, audio, video
→ store in Cloud Storage

🧮 Structured data

tables, rows, columns

Transactions, records, metrics
→ BigQuery, AlloyDB, Cloud SQL

The data-to-AI workflow

1

Ingest & Process

Pub/Sub · Dataflow

→

2

Store & Analyse

BigQuery · Looker

→

3

Activate with AI

Predict · Generate

BigQuery is the star: built for structured data, great with JSON, and can even query images/logs in Cloud Storage via external tables.

Chapter 4 · AI Models

AI vs ML vs Deep Learning vs GenAI 🧠

People mix these up. They're actually nested — each sits inside the one above. Tap any layer to expand.

🤖 Artificial Intelligence any computer mimicking human intelligence

📚 Machine Learning learns from data, not explicit rules

🕸️ Deep Learning neural networks with many layers

✨ Generative AI foundation models / LLMs like Gemini

AI is the umbrella. ML is AI that learns. Deep learning is ML with layered neural nets. GenAI uses huge deep-learning models (LLMs) to create.

Supervised vs Unsupervised Learning 🏷️

The whole difference: do your training examples come with answers (labels) or not?

	🏷️ Supervised	🔍 Unsupervised
Data	Labelled (has answers)	Unlabelled (no answers)
Driven by	A task / goal	The data itself
Finds	A known target	Hidden patterns
Example	"Is this a cat or dog?" (you labelled the photos)	"Group these dog breeds" (you don't know them)

The one testLabelled data → supervised. No labels → unsupervised. That's it.

The 5 Model Types You Must Know 🎯

🕹️Tap each card to open it — the problem it solves + its classic algorithm. Goal: link each model type to its use. (Explore freely — no points.)

🏷️ Classification

SUPERVISED · tap to expand

Predicts a category — cat vs dog, churn yes/no. Classic model: logistic regression.

📈 Regression

SUPERVISED · tap to expand

Predicts a number — future sales, price. Classic model: linear regression.

🧩 Clustering

UNSUPERVISED · tap to expand

Groups similar data points — customer segmentation. Classic model: k-means.

🔗 Association

UNSUPERVISED · tap to expand

Finds things that go together — "people who buy X also buy Y". Algorithm: Apriori.

📉 Dimensionality Reduction

UNSUPERVISED · tap to expand

Shrinks many features into fewer — simplify an insurance quote. Technique: PCA.

💡 Quick recall

CHEAT · tap to expand

Category → logistic · Number → linear · Groups → k-means · Goes-together → Apriori · Fewer features → PCA.

Try It · Chapter 4

Pick the Right Model 🎯

🕹️Drag each scenario onto the model that fits. Goal: pick by the answer shape — a category, a number, or a group. Correct → +5⭐.

Will this customer buy? (yes/no)

Predict next quarter revenue ($)

Group customers into segments

Forecast a product's future price

Spam or not spam?

Find natural shopper groups

🏷️ Classification
logistic

📈 Regression
linear

🧩 Clustering
k-means

Category answer → classification. A number → regression. "Group / segment" → clustering.

Chapter 5 · BigQuery ML

Train a Model… in SQL 🪄

BigQuery is two things in one: a storage warehouse and a fast SQL engine. BigQuery ML lets you build models where the data already lives — no Python, no data movement.

The 5 phases of an ML project

1

ETL

Get data into BigQuery

→

2

Features

Preprocess · one-hot encoding

→

3

Create

CREATE MODEL

4

Evaluate

ML.EVALUATE — accuracy, precision, recall

→

5

Predict

ML.PREDICT — adds predicted_

One-hot encoding = turning categories ("London/Tokyo") into numbers a model can use. BigQuery ML does it for you automatically.

Try It · Chapter 5

Tap Each Keyword to Learn It 🧪

🕹️Tap any yellow keyword → its job shows below. Goal: learn the anatomy of a BigQuery ML statement, piece by piece. Each new keyword → +2⭐.

CREATE MODEL ecommerce.purchase_model OPTIONS(model_type='LOGISTIC_REG', input_label_cols=['will_buy']) AS SELECT * FROM ecommerce.visitor_data; ML.EVALUATE(MODEL ecommerce.purchase_model); ML.PREDICT(MODEL ecommerce.purchase_model, ...);

Click a yellow keyword above to see what it does.

This exact pattern — CREATE → EVALUATE → PREDICT — is your homework lab: predicting visitor purchases on the Google Merchandise Store.

Knowledge Check · Module 1

Lock It In 🧪

Q1"Forecast next month's sales" is which kind of AI?

Predictive AI

Generative AI

Responsible AI

Neither

Q2What is a TPU built for?

Rendering graphics

General everyday computing

The matrix math in machine learning

Storing files

Q3"Compute & storage scale separately" describes:

Decoupling

One-hot encoding

Overfitting

Clustering

Q4Predicting a continuous number (future spending) uses which model?

Logistic regression

Linear regression

K-means

Apriori

Q5Deep learning is best described as:

The same thing as AI

A type of unsupervised data

A storage format

ML with many neural-network layers

Q6Which command trains the model?

ML.PREDICT

ML.EVALUATE

CREATE MODEL

SELECT

Q7ML.EVALUATE reports which metrics?

Latency & cost

Rows & columns

Storage & compute

Accuracy, precision, recall

Q8After ML.PREDICT, your label column comes back as:

deleted

unchanged

prefixed with predicted_

turned into an image

Module 1 — You Can Now Explain… ✅

Predictive vs Generative 3-Layer Architecture Decoupled Compute/Storage TPU

AI ⊃ ML ⊃ DL Supervised vs Unsupervised Classification / Regression Clustering

CREATE MODEL ML.EVALUATE ML.PREDICT One-hot encoding

Finish Module 1 in 2 stepsDo the hands-on lab first — then take the graded quiz.

🧪 Step 1 · Lab — Predict Visitor Purchases with BigQuery ML (~60 min)

Hands-on the exact CREATE → EVALUATE → PREDICT flow we just covered, on the Google Merchandise Store data. Open the lab →

📝 Step 2 · Official Module 1 Quiz →

After the lab, take the graded quiz on Skills Boost to confirm you're exam-ready. skills.google › course 593 › quiz 617884

Foundation locked in. 🧱

MODULE 1 COMPLETE

Next: Generative AI. ✨

You've got the foundation. Module 2 is ready now — generative AI, prompts, tuning, and agents. Modules 3 & 4 drop in next.

💎

M2 · Generative AI

Start now →

🛤️

M3 · AI Dev Options

Coming soon

🔄

M4 · AI Dev Workflow

Coming soon