DAY 03 · GENERATIVE AI LEADER

You've got this. ✨

Five short modules. Just the essentials you need to pass the GAIL exam. No fluff, no overwhelm — clear ideas you can remember tomorrow.

Here's What We'll Cover 🗺️

🧠

M1 · The Basics

What Gen AI is, in plain English

🛠️

M2 · Google's Tools

Gemini, Vertex AI, Agent Studio

⚙️

M3 · Make It Work

Prompting, grounding, RAG

🛡️

M4 · Make It Safe

Responsible AI, SAIF basics

📈

M5 · Make It Real

Business value & strategy

🏅

Then: Homework

Skills Boost path — earn the badge

No tricks. The exam tests what we'll cover. If you follow along + do the Skills Boost exercises over the next few days, you'll pass.

Before We Start

Setting Up Your Skills Boost Account 🏅

All homework + the official badge live on Google Skills Boost. 60 seconds of setup, then you're sorted for both days.

1️⃣

Sign in

Go to skills.google · use your Google account · accept the terms once.

2️⃣

Find the path

Search "Generative AI Leader" · enrol (free) · star it so it stays on your dashboard.

3️⃣

Earn the badge

Finish all videos + quizzes · badge appears on your profile · share on LinkedIn.

No live labs todayDay 3 is concepts + quizzes. We'll point you at specific Skills Boost videos at the end of each module — take your time with them over the next few days. Tomorrow's Day 4 has the hands-on labs.

Open Skills Boost in another tab now — keeps it ready when we reach the homework cards.

MODULE 01

The Basics 🧠

What Gen AI is, what it's not, and the three words that unlock everything else.

Module 1 · In This Module

Four Ideas. That's the Module. 🧠

AI → ML → Deep Learning → Gen AIThe nesting relationship the exam loves.

Discriminative vs GenerativePredicts a label vs creates something new.

Foundation Model · LLMTwo terms that unlock every modern AI product.

Supervised · Unsupervised · Reinforcement · Self-supervisedFour ways a model can learn — including how LLMs train themselves.

Small list. Every single item here appears on the exam.

Three Words. That's It. 🎯

If you remember three things from today, make it these.

Generative AI = creates new contentText, images, audio, video, code. Different from older AI that just classifies things.

Foundation Model = a big, pre-trained modelOne model, many uses. Gemini is Google's foundation model family.

LLM = Large Language ModelA foundation model that predicts the next token. That's how all chatbots work.

If you can say these three in your own words, you've already got Module 1 covered.

Discriminative vs Generative ⚖️

The exam loves this distinction. Memorise it once. Done.

🎯 Discriminative

Predicts a label.

"Spam or not?"
"Cat or dog?"
"Will this customer churn?"

VS

🎨 Generative

Creates something new.

"Write me a poem"
"Draw a cat in a hat"
"Summarise this contract"

Easy memory trick — Discriminative discriminates (picks). Generative generates (makes).

Four Types of Learning 🎓

🏷️

Supervised

Labelled data. Model learns input→output. E.g. spam filter with "spam / not spam" labels.

🔍

Unsupervised

No labels. Finds hidden patterns. E.g. grouping customers by behaviour.

🎮

Reinforcement

Trial and error with rewards. Agent learns to maximise score. E.g. AlphaGo.

🔮

Self-supervised

Creates its own labels from raw data. How LLMs train — predict the next word, no human labels needed.

Exam shortcut: "How do LLMs learn?" → Self-supervised. "Labelled cat photos" → Supervised.

Try It · Module 1

Sort These Tasks 🧩

Drag each task into the right bucket. Correct drop = +5⭐.

Detect spam emails

Write a poem about KL

Predict customer churn

Summarise this contract

Classify cat vs dog

Generate a logo image

🎯 Discriminative →

🎨 Generative →

All 6 in the right bucket? You've nailed the most-tested distinction on the exam.

Knowledge Check · Module 1

5 Quick Wins 🧪

Q1"Predict whether this customer will churn" is what kind of model?

Discriminative

Generative

Reinforcement

Unsupervised

Q2A large, broadly pre-trained model adaptable to many tasks is called a:

Custom model

Foundation model

Expert system

Rule-based system

Q3"LLM" stands for:

Large Language Model

Latent Logic Machine

Linear Learning Module

Layered Linguistic Model

Q4Which is the SMALLEST set?

Artificial Intelligence

Machine Learning

Deep Learning

Generative AI

Q5"Cat photos labelled 'cat'" → train a classifier. That's what type of learning?

Supervised

Unsupervised

Reinforcement

Self-supervised

Module 1 — Recap ✅

Generative AI Foundation Model LLM Discriminative

If you can pick these terms out of a multiple-choice question, you've nailed Module 1. That's already ~20% of the exam.

📺 Reinforce on Skills Boost

C2Understanding AI, ML, and gen AI~5 min C2Types of learning~4 min C2Deep learning, gen AI & foundation models~6 min 📝 Quiz 1 · Course 2 (~5 min)

MODULE 02

Google's Toolkit 🛠️

Five products to remember. That's the whole module.

The 5 Products You MUST Know 🌟

♊

Gemini

Google's foundation model family. Pro, Flash, Nano.

🏭

Vertex AI / Agent Platform

Google's ML platform. Where everything lives.

🎨

Agent Studio

Prompt design workbench. Was "Vertex AI Studio".

🏢

Gemini Enterprise

Turnkey enterprise AI. Search across all internal data, minimal setup.

📓

NotebookLM

Research assistant grounded in your docs.

💡

That's all!

Five products. The exam asks about these.

Don't try to memorise every Google AI product. These five = 80% of exam questions on tools.

Gemini Picking Guide ♊

Three names. Three jobs. Easy.

🦾

Gemini Pro

Hard reasoning, big context. Use when accuracy matters.

⚡

Gemini Flash

Cheap, fast, high volume. Default choice.

📱

Gemini Nano

Runs on a phone. Offline, private.

Default ruleStart with Flash. Move to Pro only if Flash isn't accurate enough.

More Tools You'll See on the Exam 🛠️

Six more cards that complete the D2 product picture.

💼

Gemini for Workspace

AI inside Gmail, Docs, Sheets, Meet & Slides. Productivity-first — no code needed.

🔬

Google AI Studio

Free playground at ai.google.dev — test prompts, tweak temperature & top-K before writing code.

💎

Gems

Saved custom AI personas in Gemini Advanced. Write instructions once, reuse as a named assistant.

☁️

Gemini for Google Cloud

Gemini built into BigQuery, Vertex AI & Cloud Console for dev & data teams.

🖼️

Multimodal AI

One model handles text + images + audio + video. Gemini is multimodal. Key D1 concept.

🆚

Two Studios — Different!

Agent Studio (Cloud Console) = build & deploy agents. Google AI Studio (ai.google.dev) = experiment with prompts.

Quick rule: Workspace = anyone · AI Studio = devs prototyping · Gems = power users · Gemini for GCloud = engineers.

Try It · Module 2

Tap to Reveal — When to Use Each 🃏

Click each card. The "use when…" answer flips open underneath.

🦾

Gemini Pro

Heavy reasoning, long context

📋 Use when: complex multi-step analysis (legal contracts, research synthesis, ambiguous customer queries). Slower & pricier but most accurate.

⚡

Gemini Flash

Cheap, fast, high volume

📋 Use when: 1M+ emails a day, real-time classification, anything where latency + cost matter more than peak accuracy. Default for most workloads.

📱

Gemini Nano

On-device, offline

📋 Use when: privacy-critical data, no internet, ultra-low latency. Runs on Pixel and similar phones.

🏢

Gemini Enterprise

Turnkey for enterprise

📋 Use when: company-wide rollout, search across all internal data, want minimal engineering effort with enterprise-grade governance.

📓

NotebookLM

Research with citations

📋 Use when: research assistant for a defined corpus (your PDFs, contracts, sources). Every answer cites the exact source page.

🎨

Agent Studio

Prompt & agent workbench

📋 Use when: designing, testing, tuning prompts and agents in Google Cloud with full enterprise governance.

After tapping all 6, you've seen the "use when" for every product. That's all the exam tests on this topic.

Knowledge Check · Module 2

5 Quick Wins 🧪

Q1You need to process 1 million support emails per day, cheaply. Pick:

Gemini Pro

Gemini Flash

Gemini Nano

Custom training

Q2Which Google product is a turnkey enterprise AI experience?

BigQuery

Cloud Run

Gemini Enterprise

Looker Studio

Q3Which is the Gen AI prompt-design workbench in Google Cloud Console?

Agent Studio

Cloud Shell

Vertex AI Workbench

BigQuery Studio

Q4A research assistant grounded ONLY in your uploaded PDFs is:

Google AI Studio

Gemini app

NotebookLM

Vertex AI Search

Q5Need an AI model that runs offline on a phone. Pick:

Gemini Pro

Gemini Flash

Gemma open

Gemini Nano

Module 2 — Recap ✅

Gemini Pro Gemini Flash Gemini Nano Vertex AI Agent Studio Gemini Enterprise NotebookLM Gemini for Workspace Google AI Studio Gems Multimodal

Eleven product/concept names. That's the whole D2 picture. You know them all now.

📺 Reinforce on Skills Boost

C1Vertex AI & multimodal gen AI~6 min C3The platform layer + The model layer~8 min C4NotebookLM + NotebookLM in action~7 min C5Gemini Enterprise~5 min 📝 Quiz 2 · Course 3 (~5 min)

MODULE 02b

AI Agents 🤖

What changes when AI can act — not just answer.

Agents · In This Section

From Chatbot to Agent 🤖

ChatbotYou ask → it answers. One shot. Passive.

AgentYou give a goal → it plans, acts, observes, repeats until done. Active.

An agent is a gen AI model connected to tools — search, APIs, databases, code execution. It doesn't just talk. It does things.

The Agent Loop 🔄

Every agent runs this cycle — over and over until the goal is complete.

1 · Perceive → reads the goal and current context

2 · Plan → breaks the goal into steps

3 · Act → calls a tool (search, API, code, email)

4 · Observe → checks the result, decides next step

Real example"Book cheapest flight to Singapore Friday" → searches flights → compares prices → confirms booking. All automated.

Google's Agent Stack 🏗️

🎨

Agent Studio

Design and test agent flows. No-code / low-code. Part of Vertex AI.

🏭

Vertex AI / Agent Platform

Deploy agents at enterprise scale with monitoring and governance.

🏢

Gemini Enterprise

Pre-built agents for HR, finance, and sales workflows. Zero engineering needed.

Think of it as a spectrum — Agent Studio for builders, Gemini Enterprise for business users who just want it to work.

Knowledge Check · Agents

3 Quick Wins 🧪

A1What makes a Gen AI agent different from a basic chatbot?

It uses a larger model

It always gives longer responses

It takes multi-step actions using tools to reach a goal

It runs faster than standard models

A2Which Google tool is the no-code/low-code workbench for designing agent flows?

Agent Studio

Cloud Shell

BigQuery ML

Vertex AI Workbench

A3The correct sequence of the agent loop is:

Plan → Act → Perceive → Observe

Act → Observe → Perceive → Plan

Observe → Perceive → Act → Plan

Perceive → Plan → Act → Observe

Agents — Recap ✅

Agent Agent Loop Agent Studio Gemini Enterprise

Agents = gen AI that acts. The exam tests whether you know the difference — chatbot answers once, agent loops until the goal is done.

📺 Reinforce on Skills Boost

C3Agents and gen AI powered applications + How agents work~8 min C5Building agents + Customer Experience Agent Studio~12 min 📝 Quiz 1 · Course 3 (~5 min)

MODULE 03

Make It Work ⚙️

How to actually get Gen AI to do what you want.

Three Magic Techniques 🎩

✍️

Prompting

Write better instructions. Cheap, fast, try first.

📚

RAG

Give the model your private docs to read. No hallucination.

🎯

Fine-tuning

Re-train on your data. Last resort, expensive.

Golden RuleTry in this order: Prompting → RAG → Fine-tuning. Stop when it works.

If a question asks "best technique for…", look at cost first. Cheapest that works = right answer.

Prompting Techniques Toolkit 🎯

Prompting has sub-types. The exam names them — you need to recognise each.

📸

Few-shot

Include 2–5 example pairs in the prompt. "Here's the pattern — now do it for this input."

🎭

Role prompting

Give the model a persona. "You are a senior lawyer…" — outputs match that role's style & expertise.

⛓️

Prompt chaining

Output of step 1 → input of step 2. Break complex tasks into a sequential pipeline.

Temperature — One Number to KnowRange 0 → 2. Near 0 = deterministic (same answer every time, good for factual tasks). Near 2 = creative & varied. Default ≈ 1. Exam tip: "need consistent / factual" → temperature near 0.

Exam shortcut: "consistent format" → few-shot · "act as expert" → role prompting · "multi-step workflow" → prompt chaining · "deterministic output" → low temperature.

Grounding & Hallucination 🧯

HallucinationWhen AI confidently makes up facts. The #1 risk.

GroundingConnect the model to a real source (Google Search, your docs). It cites where the answer came from.

Whenever you see "stop hallucination" on the exam → answer is almost always grounding or RAG.

Try It · Module 3

Which Technique Fits? 🎯

Drop each scenario in the cheapest technique that solves it. Wrong drop bounces back — try again.

"Translate this paragraph to Malay"

"Answer questions about our private HR policy"

"Generate replies in our exact brand voice, at scale"

"Summarise an open news article"

"Live chatbot for current product inventory"

"Output strict legal-clause format every time"

✍️ Prompt

📚 RAG

🎯 Fine-tune

Pattern lock-in: generic tasks → Prompt. Fresh/private facts → RAG. Domain tone or structured output at scale → Fine-tune.

Knowledge Check · Module 3

5 Quick Wins 🧪

Q1The model needs up-to-date answers from your private docs. Best technique?

Fine-tune a custom model

RAG (Retrieval-Augmented Generation)

Increase model size

Lower temperature to 0

Q2When the model confidently makes up facts, that's called:

Overfitting

Drift

Hallucination

Skew

Q3Correct order to try Gen AI techniques (cheapest first)?

Prompting → RAG → Fine-tuning

Fine-tuning → RAG → Prompting

RAG → Fine-tuning → Prompting

Fine-tuning → Prompting → RAG

Q4Connecting model output to verifiable real sources is called:

Tuning

Caching

Grounding

Pre-training

Q5You want the most deterministic, factual output. Set temperature to:

2.0 (maximum)

1.0

Near 0

It doesn't matter

Module 3 — Recap ✅

Prompting Few-shot Role prompting Prompt chaining RAG Fine-tuning Grounding Hallucination Temperature

Nine techniques. That's every D3 concept the exam tests. You've seen them all.

📺 Reinforce on Skills Boost

C1Prompting models + Prompting the Gemini app~10 min C2Techniques to overcome limitations~6 min C4Prompting techniques + Grounding and RAG~12 min 📝 Quiz 2 · Course 2 (~5 min)

MODULE 04

Make It Safe 🛡️

Responsible AI in plain words. The exam wants you to know two frameworks.

Two Frameworks. Just Two. 📜

Google's AI Principles (2018)7 ideas. Don't memorise all 7. Remember the spirit: do no harm, be accountable, be private, be fair.

SAIF — Secure AI Framework6 elements. Just remember it exists and means "security best practices for AI systems." The exam loves the acronym.

If a question mentions "secure AI framework" — the answer is SAIF. If it mentions "Google's principles" — pick the answer that sounds most ethical. Done.

The Real Safety Lever — HITL 👤🔁🤖

Human-in-the-Loop (HITL)A human reviews or approves AI output before it has consequence.

Why it matters: it's the answer to almost every "should we deploy this risky AI?" question on the exam. The right answer is rarely "yes" or "no" — it's "yes, with HITL".

Memory hook: Whenever the scenario says "high-stakes decision" or "regulated industry" → look for the option that includes HITL.

Risk Register 📋

A living document that lists, scores, and assigns ownership to every AI risk in your project.

⚠️

Identify

List all risks: hallucination, bias, data leakage, prompt injection, knowledge cutoff.

📊

Score

Rate each risk: Likelihood × Impact = Priority. High priority = act first.

👤

Own

Assign one person responsible for monitoring and mitigating each risk.

Exam cue: "How does a team track and manage AI risks?" → Risk Register.

Try It · Module 4

Rate the Risk ⚖️

Tap your guess. The card reveals the real answer and why.

🏦 An AI auto-approves home loans up to RM 50,000 with no human review.

HIGH — regulated financial decisions require bias audits + explainability + HITL. Solo deployment violates equal-credit rules in most jurisdictions.

📰 An AI summarises today's public news headlines into a 3-bullet brief.

LOW — public data, no consequence to individuals, output is informational. Hallucination is a nuisance, not a harm.

🏥 An AI suggests possible diagnoses to a doctor, who confirms before treatment.

MEDIUM — high-stakes domain, but the doctor (HITL) is the decision-maker. AI is decision support, not decider.

👥 An AI screens job applicants and auto-rejects below a threshold score.

HIGH — hiring is regulated. Bias risk is enormous. Autonomous rejection is illegal in many jurisdictions. Always HITL.

Pattern: ask "would a wrong decision change someone's life?" If yes → HIGH → demands HITL.

Knowledge Check · Module 4

5 Quick Wins 🧪

Q1Google's framework for securing AI systems is called:

RAG

SAIF

MLOps

HITL

Q2A bank wants to auto-approve loans with AI. Best move?

Ship it — speed wins

Block it forever — too risky

Pilot with Human-in-the-Loop review

Use a smaller model

Q3"HITL" stands for:

Human-in-the-Loop

High-Intensity Training Loop

Hosted Inference Test Layer

Hyper-Iterative Tuning Loop

Q4The single strongest safety lever for high-stakes AI deployments is:

A bigger model

More training data

Stricter API rate limits

Human-in-the-Loop review

Q5Google's AI Principles were first published in which year?

2010

2018

2022

2024

Module 4 — Recap ✅

AI Principles SAIF HITL Hallucination Bias

You've now got the safety vocab. These show up in 4-6 exam questions.

📺 Reinforce on Skills Boost

C1Keeping humans at the forefront of gen AI~6 min C2Google strategies for foundation model limitations~5 min C2Secure AI + Responsible AI~10 min 📝 Quiz 3 · Course 2 (~5 min)

MODULE 05

Make It Real 📈

How a leader takes Gen AI from idea to business value.

The 5 Stages — Memorise These 🛤️

Discover → Prototype → Pilot → Scale → Transform Every Gen AI project goes through these stages. Discover a problem. Prototype cheap. Pilot with real users. Scale what works. Transform the whole workflow.

Exam tip: any "what stage should they be in?" question — look at the scenario verbs. "First idea" = Discover. "Live users" = Pilot. "Company-wide" = Scale or Transform.

The Value Equation 🧮

Value = (Time saved + Revenue lift) − (Build + Run + Risk) Track cost per resolved outcome, not cost per token. Run cost dominates at scale.

If a question asks "best KPI to track" → look for an option that measures business outcomes (cost per ticket, time-to-decision), not technical things (accuracy, parameters).

Try It · Module 5

Match the Stage 🛤️

Drop each scenario into the right stage of the adoption path.

"Walking the floor to spot workflow bottlenecks"

"Built a working demo in Agent Studio last week"

"50 staff using it daily with weekly feedback"

"Rolled out to 5,000 employees across 3 countries"

"Whole support department restructured around AI agents"

1 · Discover

2 · Prototype

3 · Pilot

4 · Scale

5 · Transform

Stage = scope of users. Few people exploring → Discover. One real workflow with feedback → Pilot. Entire department redesigned → Transform.

Knowledge Check · Module 5

5 Final Wins 🧪

Q1A team just built their first working Gen AI demo. What stage are they at?

Discover

Prototype

Scale

Transform

Q2Best KPI for an AI customer-service deployment?

Cost per token

Cost per resolved interaction

Number of model parameters

Training accuracy

Q3The 5 stages of Gen AI adoption, in order:

Discover → Prototype → Pilot → Scale → Transform

Prototype → Discover → Scale → Pilot → Transform

Pilot → Discover → Prototype → Transform → Scale

Transform → Scale → Pilot → Prototype → Discover

Q4At scale, which cost dominates most often?

Build cost (one-time)

Engineering salaries

Per-call inference cost × volume

Hardware purchase

Q5After a Pilot succeeds with real users, the next stage is:

Discover

Prototype

Scale

Re-pilot

Module 5 — Recap ✅

Discover Prototype Pilot Scale Transform Value Equation

📺 Reinforce on Skills Boost

C1The AI-first company + Building a successful gen AI strategy~12 min C3Gen AI project resources + Choosing & maintaining your solution~10 min C5Strategy: Planning ahead~6 min 📝 Quiz 3 · Course 1 (~5 min)

You've got the whole exam framework. 🎯

Screenshot This 📸

Your Exam Cheat Sheet 📋

All the key concepts on one screen. Take a photo. Review over the next few days.

🧠 The Basics

Generative AIFoundation ModelLLMMultimodalDiscriminativeSupervised

🛠️ Google's Tools

Gemini ProGemini FlashGemini NanoGemini for WorkspaceGemini for Google CloudGoogle AI StudioGemsNotebookLMVertex AIGemini Enterprise

⚙️ Make It Work

PromptingFew-shotRole PromptingPrompt ChainingTemperatureRAGGroundingFine-tuningHallucination

🛡️ Make It Safe

AI PrinciplesSAIFHITL

📈 Make It Real

DiscoverPrototypePilotScaleTransformCost/Outcome

Match every keyword to its meaning and you'll pass the exam. Truth.

Exam Day — Know What to Expect ⏰

📋

Format

50–60 multiple choice questions

⏱️

Time

90 minutes · ~90 sec per question

🎯

Pass Mark

~75% · aim for 80% to be safe

🧠 Test-taking tips

1. First read — answer everything you know fast. Flag hard ones.
2. Second pass — return to flagged questions with fresh eyes.
3. Eliminate — wrong answers are usually obvious. Cross them off.
4. Trust patterns — "HITL" answers safety questions. "RAG" answers freshness questions. "Cheapest that works" wins technique questions.
5. Never leave blank — no negative marking. Guess if you must.

90 minutes for up to 60 questions = generous. You have time. Breathe. You've prepared. You'll do well.

Mock Exam 🏁

Full Exam Simulation — 20 Questions

⏱️

Time

Set a 25-minute timer now

📵

Rules

No slides, no notes — exam conditions

🎯

Target

15 / 20 = exam-ready

Domain mix (mirrors real exam)

D1 Fundamentals 30% · D2 Google Products 35% · D3 Techniques 20% · D4 Business/Safety 15%

Scroll down — all 20 questions on this slide. Answers revealed on click.

Mock Exam · Questions 1 – 20

Go. ⏱️

M1Which best defines a "foundation model"?

A large model pre-trained on broad data, adaptable to many downstream tasks

A task-specific model trained from scratch on labelled examples

A rule-based expert system encoding domain knowledge

A model that only handles structured tabular data

M2What is the relationship between AI and machine learning?

They are identical — the terms are interchangeable

ML is older; AI is the modern replacement

AI is the broad field; ML is a specific subset where systems learn from data

AI requires custom hardware; ML is purely software

M3NotebookLM is best described as:

A code notebook for training and deploying ML models

An AI research assistant grounded in your own uploaded documents

A no-code agent builder for business workflows

A model fine-tuning interface on Vertex AI

M4Which statement best describes "generative AI"?

AI that classifies inputs into predefined categories

AI that detects anomalies in time-series data

AI that optimises logistics routes using reinforcement learning

AI that creates new content — text, images, audio, code — from patterns learned in training

M5Google's no-code platform for building and deploying AI agents is:

Vertex AI Workbench

Google AI Studio

Agent Studio

BigQuery ML

M6"Few-shot prompting" means:

Including 2–5 example input/output pairs so the model learns the desired format

Using the shortest possible prompt to save tokens

Sending the same prompt a few times and picking the best result

Setting temperature low so the model is less random

M7In the agent loop, what happens during the "Observe" step?

The agent searches for available tools to use

The agent decomposes the goal into sub-tasks

The agent sends a reply to the user

The agent evaluates the result of its last action

M8Gemini for Workspace is primarily built to:

Train and deploy custom ML models on Google Cloud

Boost productivity in Gmail, Docs, Sheets, Meet, and Slides

Replace Vertex AI for large enterprise AI workloads

Provide on-premise AI infrastructure for regulated industries

M9Which Google product lets an app search private documents and cite sources in its responses?

Gemini for Workspace

Google AI Studio

Vertex AI Search with Grounding

Cloud Natural Language API

M10Which scenario best illustrates unsupervised learning?

A spam filter trained on human-labelled email examples

A model earning reward points for winning chess games

A language model predicting the next token in a sentence

A model grouping customers by purchase behaviour — no labels given

M11"Prompt chaining" means:

Using the output of one prompt as the input to the next, step by step

Saving multiple prompts in a reusable library (Gems)

Running the same prompt in parallel across multiple models

Prepending chain-of-thought examples to every prompt

M12A responsible AI practice for a high-stakes content moderation model is:

Disable safety filters to minimise false positives

Add a human review queue for borderline decisions

Switch to the largest model to reduce errors automatically

Publish model weights publicly to build external trust

M13What is the primary purpose of "role prompting"?

To enforce a specific JSON output schema

To break a complex task into smaller sequential steps

To assign the model a persona so responses match a specific expertise or tone

To improve factual accuracy by injecting citations

M14Fine-tuning a foundation model is most appropriate when:

You need the model to access live web data in real time

The model must adopt a specific style, domain vocabulary, or task format from examples

You want to add external database connectivity to the model

Prompting alone produces responses that are too long

M15Which is a core element of a gen AI Risk Register?

The model's parameter count and training compute cost

A library of approved prompt templates and guardrails

Vendor SLA commitments and uptime guarantees

Each risk scored by likelihood × impact, with a named owner assigned

M16Which Google tool lets developers tweak temperature and top-K and preview outputs before writing any code?

Google AI Studio (Parameter Playground)

NotebookLM

Vertex AI Pipelines

Cloud Build

M17A "multimodal" AI model can:

Run efficiently across multiple CPUs in parallel

Understand and generate content in multiple spoken languages

Apply multiple safety policies at once

Process and generate across multiple data types — text, image, audio, and video

M18An "AI-first company" is best described as one that:

Embeds AI into strategy and operations as a core competitive advantage

Builds all its software using AI-generated code

Employs only AI researchers in leadership roles

Sells AI infrastructure exclusively to other businesses

M19The primary limitation that RAG addresses in large language models is:

Slow inference speed on long documents

Inability to follow multi-step instructions

A fixed knowledge cutoff with no access to live or private data

Generating responses that are too verbose

M20"Gemini for Google Cloud" is primarily aimed at:

Consumer productivity tasks in personal Gmail and Drive

Developers and data teams building on Google Cloud services

Small businesses without in-house IT departments

Marketing teams creating social media content

15+ correct = exam-ready, go book it. 12–14 = strong, 1 more Skills Boost pass — take your time. Under 12 = revisit your weak domains — the recap cards tell you exactly where to go.

Your Next Step 📝

One Job — Earn the Badge 🏅

Now that the framework is in your head, at your own pace over the next few days, go deep with Google's official path.

🔗 Generative AI Leader Certification Path

skills.google/paths/1951

5 short courses · ~7 hours · Google Cloud badge on completion

Today gave you the map. Skills Boost gives you the details. Do it over the next few days — easy when you have both.

DAY 3 COMPLETE

You've got this. 💪

30 exam concepts. 5 modules. 1 framework. The GAIL exam is now something you can pass. Do the homework, sleep well, smile in the morning.