GenAI Leader cheat sheet — every exam topic on one page

Fundamentals of gen AI

Generative AI Concepts: Foundation Models, LLMs, and Business Use Cases

• Generative AI is a branch of machine learning that creates new content; traditional ML classifies or predicts, and one organization often needs both.
• NLP covers non-generative language tasks too: a model that only tags or scores text is NLP and ML, not gen AI.
• A foundation model is pretrained on broad data and adapted to many tasks; an LLM is pretrained self-supervised on unlabeled text, then refined with labeled instructions and RLHF.
• Multimodal means several data types in one model (Gemini); multilingual means several languages. When inputs of different types must be checked against each other, one multimodal model sees them together, while a chain of models passes along only a lossy summary.
• Diffusion models generate images by iteratively denoising random noise — the concept behind Imagen; Veo generates video. GANs and VAEs are other generative approaches.
• Prompt engineering edits the words, grounding/RAG adds retrieved documents, prompt tuning learns a soft prompt with weights frozen, and fine-tuning changes weights.
• Supervised learning uses labeled data, unsupervised finds patterns (and anomalies) in unlabeled data, reinforcement learns from rewards and penalties.
• The ML lifecycle runs ingestion (Pub/Sub, Cloud Storage), preparation (Dataflow, BigQuery), training, deployment, and management (Gemini Enterprise Agent Platform model monitoring for drift) — as a loop.
• Choose a foundation model by modality, context window, security, availability, cost, performance, fine-tuning, and customization; set the quality bar first, then let cost decide among models that meet it.
• Gen AI business value maps to four verbs: create, summarize, discover, and automate.

Data Types in Gen AI: Quality, Structured vs Unstructured, Labeled Data

• Gen AI output quality is bounded by data quality, and a generative model fails silently: bad data produces confident, plausible, wrong answers rather than errors.
• Poor data quality costs trust, introduces bias, inflates preparation cost and delay, misleads decisions, and leaves use cases out of reach.
• A deployed foundation model does not learn from the requests it serves; behavior changes only through deliberate tuning or updated grounding data.
• Six characteristics: completeness (missing data or groups), consistency (sources disagree), relevance (wrong subject or outdated), availability (cannot be reached), format (cannot be processed), cost (preparation outweighs value).
• Under-represented groups in training data are a completeness gap that shows up as bias against those groups.
• Structured data fits rows and columns (transactions, CRM records, inventory); unstructured data is free-form (emails, contracts, images, audio, video), makes up the large majority of enterprise data, and is what foundation models unlock.
• Structure belongs to the content, not the container: free-text reviews in a database column are still unstructured.
• Labels are the correct outputs for a task; metadata and record IDs are not labels. Supervised tuning needs input-output pairs.
• Structured/unstructured and labeled/unlabeled are independent: free-text call notes paired with deal outcomes are unstructured and labeled.
• Foundation models pretrain self-supervised on unlabeled data; expert-judgment tasks still need expert labels, which is a cost and time decision.
• Solve access before format, keep grounding sources synchronized and superseded content retired, and monitor answers after launch.

The Five Layers of the Gen AI Landscape and What They Mean for Business

• The gen AI landscape has five layers: infrastructure, models, platforms, agents, and applications, and each builds on the layer below.
• Infrastructure (TPUs, GPUs, AI Hypercomputer with its optimized software and consumption models, Google Cloud data centers) is rented from a cloud provider, not built or bought.
• The models layer is a choice among foundation models like Gemini, Gemma, Imagen, and Veo, not a mandate to train your own.
• Openness is not location: an on-site data requirement is met only when the model actually runs on the organization's own servers (or on Google Distributed Cloud).
• Classify offerings by role: Model Garden and Agent Search are platform tooling for builders, even though one lists models and the other has "agent" in its name.
• Platforms such as Agent Platform with Model Garden are where enterprises build custom, data-grounded solutions with control over security and governance.
• Agents combine a model with reasoning and tools to complete multi-step tasks; start with high-volume, goal-clear work with live system access and low-risk, reversible actions, keeping people in the loop for money-moving steps.
• Applications like the Gemini app, Gemini for Google Workspace, and Gemini Enterprise deliver the fastest time to value with no development team.
• An employee Gemini Enterprise deployment serves staff; customer-facing experiences need a customer-facing product or a custom build.
• Lower layers trade speed for control and differentiation; higher layers trade customization for speed and low cost.
• Most organizations operate at two layers: applications for broad productivity, plus platforms or agents for a few differentiating use cases.

Google's Foundation Models: Gemini, Gemma, Imagen, and Veo Explained

• Google's four named foundation models are Gemini (multimodal reasoning), Gemma (lightweight open models), Imagen (image generation and editing), and Veo (video generation).
• Gemini is natively multimodal with a long context window; one long-context request keeps links between distant parts of a document that chunking would lose.
• Function calling lets Gemini use external tools and APIs, grounding with Google Search adds retrieved web facts, and as an agent's reasoning engine Gemini chooses tools, checks results, and re-plans, with no need for open weights.
• Gemma's defining trait is open weights the organization possesses and modifies; Gemini Nano, Flash-Lite, and Gemini on Google Distributed Cloud (Google-supplied hardware in your facility) are small, cheap, or on-premises but never release weights.
• Tuning Gemma yields weights you keep and deploy anywhere; tuning Gemini yields a platform-served model; grounding and few-shot prompting change no weights.
• Imagen generates new images and edits existing ones (instruction-based edits, inpainting, outpainting); Google is retiring Imagen in favour of Gemini image models, though the exam guide still names Imagen.
• Veo generates new video, with synchronized audio on current models; analysing existing footage is Gemini's job.
• Output modality usually decides among Gemini, Imagen, and Veo; openness matters between Gemini and Gemma; a small Gemma tuned for one narrow, high-volume task is cost-efficient, but cost alone does not justify self-hosting general-purpose work.
• Beyond modality and openness, weigh context window, cost and performance, security and data handling, availability, and customization needs.
• The models are complements: real projects combine them, and all four are accessible to builders through Model Garden on Agent Platform.

Google Cloud's gen AI offerings

Google Cloud's Gen AI Strengths: Platform, Infrastructure, and Openness

• Google's AI-first approach and research heritage — including the Transformer architecture — create a short path from breakthrough to product, so customers get state-of-the-art capability and a strong roadmap.
• Enterprise-ready means five qualities: responsible, secure, private, reliable and scalable — match the quality to the concern a scenario stresses.
• Google's AI Principles cover responsible use (fairness, accountability, harm); SAIF covers AI security (prompt injection, data poisoning, model theft).
• The comprehensive ecosystem integrates the same Gemini models across Google products and offers every layer from one provider; the open approach is about model choice and reduced lock-in — standardizing on one model is its opposite.
• Model Garden on Gemini Enterprise Agent Platform lists Google, open (Gemma, Llama) and partner (Claude) models; Google AI Studio, Gems and Gemma on Kaggle offer only Google's models.
• Gemma's weights are open to download and adapt; Gemini's are not, even when Gemini runs on Google Distributed Cloud in a customer's facility.
• AI-optimized infrastructure combines the AI Hypercomputer, custom TPUs, GPUs (so GPU code need not be rewritten) and efficient data centers, consumed on demand — committed use discounts are optional.
• Customer data is not used to train Google's foundation models without permission — a default commitment, not an opt-out; residency decides where data lives, CMEK who can decrypt, audit logs who did what, permissions-aware grounding what an agent may reveal.
• Pre-built as is for speed, customize for own data and brand voice, custom-build for proprietary processes with developers and full control; fine-tuning a model does not make it an agent.
• No-code tools remove coding, pre-trained models remove training, APIs remove ML expertise but still need application code.

Prebuilt Gen AI at Work: Gemini App, Gemini Enterprise, and Workspace

• Prebuilt offerings are finished, Google-maintained products — fastest time to value, no ML skills required — and the entry point of Google Cloud's gen AI spectrum.
• The Gemini app is a standalone multimodal assistant for individuals; its premium tier (Gemini Advanced, now Google AI Pro) adds the most capable models and longer context for very long material.
• Gems are custom, reusable versions of Gemini configured once with instructions and optional files; they are free for all Gemini app users, shareable, usable in Workspace side panels, and do not retrain the model.
• Gemini Enterprise grounds answers by retrieval at question time across Google and third-party repositories (SharePoint, Confluence, Jira, ServiceNow and more) — never by retraining on company files — and, for connectors that carry access controls, with the asking user's permissions.
• Multimodal search covers everything connected; a Gemini Notebook Enterprise notebook answers only from a curated source set; the Gemini Notebook API lets internal applications create, share, and populate notebooks and generate audio overviews.
• Agent Designer builds shared, governed Gemini Enterprise agents without code; sharing an agent never shares data a user cannot already open.
• Gemini for Google Workspace embeds AI (such as Help me write) in Gmail, Docs, Sheets, Slides, Drive, Meet, and Chat, works only on content the user can access, and does not use customer content to train models outside the organization without permission.
• Use prebuilt for common employee productivity; reserve custom builds on the Agent Platform for customer-facing, differentiating features tied to proprietary systems.
• Business value scales with scope: personal productivity from the Gemini app, workforce time savings from Workspace, and unlocked institutional knowledge from Gemini Enterprise.

Improving Customer Experience with Agent Search and Customer Engagement Suite

• Google Cloud improves customer experience through two families: external search offerings (Agent Search, grounding with Google Search) and the Customer Engagement Suite.
• Agent Search, on the Gemini Enterprise Agent Platform, embeds Google-quality generative search — with prebuilt RAG and grounded, cited answers — into a company's own sites and apps for customers; Gemini Enterprise serves employees.
• Agent Search grounds answers in your enterprise content; grounding with Google Search retrieves public-web results at response time without retraining the model. Grounding reduces errors but does not eliminate them, and answers are only as current as their sources.
• Fine-tuning shifts model tendencies but gives no source references and goes stale between retrains; mandatory steps belong in deterministic flows.
• Conversational Agents resolve customer requests directly via chat and voice, blending deterministic flows with grounded generative answers, and hand off to humans with context carried over.
• Agent Assist is a real-time copilot inside one live conversation: proactive knowledge, suggested responses, coaching guidance, and automatic summaries. It speeds contacts without reducing their number.
• Conversational Insights analyzes completed conversations across the whole volume for contact drivers, sentiment, and quality, replacing hand-scored samples with evidence-based coaching.
• Contact Center as a Service is Google's cloud contact-center platform: elastic capacity, no on-premises telephony, natively integrated AI. The AI components also work with partner platforms, and they can only help with contacts the platform connects.
• Business benefits by component: higher search conversion and deflection (Agent Search), lower cost per contact (Conversational Agents), shorter handle time and faster new-agent ramp-up (Agent Assist), fewer contacts through root-cause fixes and better coaching (Conversational Insights), and higher customer satisfaction overall.

How Google Cloud Empowers Developers: Agent Platform and RAG

• Agent Platform is Google Cloud's unified environment for discovering models, grounding them in company data, and building, deploying, and managing gen AI agents.
• Model Garden offers managed Gemini models as a service, open-weight models such as Gemma to download and self-host, and partner models such as Claude billed and governed through your own Google Cloud project.
• Pick the smallest model that does the job: tuning, temperature, and Provisioned Throughput leave a model's size, inherent speed, and per-token price unchanged — tuning can only trim cost through shorter prompts.
• Agent Search delivers out-of-the-box, Google-quality search and grounded, source-referenced answers over an organization's own data; judge it by business outcomes such as support contacts avoided, not by usage volume.
• Prompt first, tune when prompting falls short and examples exist, use AutoML for company-specific image or tabular labels without ML engineers, and write custom training last.
• Prebuilt RAG with Agent Search is the managed retrieval path; RAG APIs are components — parsing, embeddings, retrieval, ranking — that can be adopted one at a time.
• Stale grounded answers are fixed in the data store — add the new source, remove the superseded one — not by tuning or instructing the model.
• A custom agent acts only through tools connected to the target system; instructions alone can only describe the steps.
• Take the most packaged offering that meets the need; every step toward custom adds engineering cost a requirement must justify.

Gen AI Agent Tooling: Extensions, Functions, Data Stores, and Plugins

• Tools are the agent's connection to the external environment — without them a model can only generate text from frozen training data; with them an agent can retrieve live information and complete tasks.
• Extensions are standardized bridges to external APIs, executed by the agent platform on the agent's behalf.
• Functions keep execution on your side: the model outputs which function to call and with what arguments, but your application validates and runs the action — the governed choice for consequential operations.
• Data stores ground an agent in company documents and data, delivering current, cited answers that stay accurate through content updates — tuning always lags the latest change, and prompt-pasted content must be re-edited by hand and cannot hold a large corpus.
• Plugins extend an agent into outside applications and services, reusing capabilities the business already has.
• Cloud Storage backs files and documents, databases back live records, Cloud Functions (now also called Cloud Run functions) runs a single function from source per request or event, Cloud Run hosts containerized services, and Agent Platform is where agents are built and governed.
• The Google Cloud API Library is where you discover and enable Google APIs for a project; Model Garden is for models and Agent Garden for sample agents.
• Pre-built AI APIs — Speech-to-Text, Text-to-Speech, Translation, Document Translation, Document AI, Cloud Vision, Video Intelligence, Natural Language — give agents proven abilities with no training data or ML expertise; the trade-off is limited customization.
• Vision OCR returns raw text while Document AI returns structured form fields; Vision sees each image alone while Video Intelligence follows shots and objects over time.
• Google AI Studio is for quick prototyping with Gemini models; Agent Studio is for building grounded, governed, production gen AI agents.

Techniques to improve gen AI model output

Overcoming Foundation Model Limitations: Grounding, RAG, and HITL

• Data dependency is about coverage (what was never in training data, such as private documents); the knowledge cutoff is about time (what happened after collection). Both are closed by grounding or RAG.
• Models often invent rather than decline when knowledge is missing, so cutoff gaps usually surface as hallucinations.
• Grounding fixes fabricated facts only when its source holds the facts; lower temperature reduces variability, not fabrication.
• RAG keeps answers current and traceable; stuffing a whole library into every prompt hits context and cost limits, and fine-tuned "citations" are not real attribution.
• Match limitation to remedy: wrong or stale facts → grounding/RAG (retrieve, augment, generate); tone or format → prompt engineering; a persistent domain gap → fine-tuning; high-stakes or fairness-sensitive decisions → human in the loop as a standing safeguard.
• The cost-effective sequence is prompt engineering first, grounding and RAG next, fine-tuning last — and only when a real gap remains.
• Grounding, prompting and fine-tuning reduce errors; human review catches an individual error before it takes effect; monitoring detects problems after the fact.
• Value KPIs measure business outcomes, not usage or public benchmark scores; drift is gradual change with an unchanged deployment, not an outage or a single wrong answer.
• Pin a stable model version, evaluate each newer version against your KPIs, then promote; stable versions retire, and Google patches managed models while self-hosting moves patching to your team.
• The Agent Platform Feature Store serves the same governed feature values to training, testing and production, preventing training–serving inconsistency.

Prompt Engineering Techniques: Few-Shot, Role, and Chain-of-Thought

• Prompt engineering is deliberately designing the model's input — instruction, context, format, examples — and is the fastest, cheapest way to improve gen AI output.
• Prompting is iterative: check each refined output against every requirement, and when a fix drops something the prompt never stated, add it explicitly rather than reverting.
• Fine-tuning (changes weights), model selection, generation settings and human review can all improve results, but none of them is prompt engineering — and prompting should be tried first because it is cheapest.
• Prompting cannot supply missing or outdated facts: supply current data with the request (grounding), which is cheaper than fine-tuning and does not need repeating when facts change.
• Zero-shot means no examples, one-shot means one, few-shot means several; add examples when output misses your format or standard, and use the fewest that work.
• Examples teach whatever they share: one or unrepresentative examples can make the model copy surface features, so use specific, varied examples that cover every case — but not so many that it overfits.
• Role prompting assigns a persona that steers tone, depth, and vocabulary — but it changes how the model communicates, not what it factually knows.
• Prompt chaining breaks complex work into linked steps with inspectable intermediate outputs, moving quality control into the middle of the process.
• Chain-of-thought prompting asks for step-by-step reasoning before the answer, improving multi-step logic and making conclusions auditable.
• ReAct prompting interleaves reasoning with actions such as search or tool lookups, and is the conceptual foundation of AI agents.
• Techniques combine in practice: a role plus few-shot examples inside a chain, with ReAct gathering live information, is a typical production pattern.

Grounding Techniques: RAG, Google Search, and Sampling Settings

• Grounding connects LLM responses to verifiable data sources, reducing hallucinations and keeping answers current without retraining the model.
• First-party data is what your organization creates and owns, third-party data is acquired from external providers, and world data is broad public knowledge such as Google Search results — the category follows the data's origin, not the app's audience.
• RAG retrieves relevant content at query time and sends it with the prompt, making output more accurate, current, and attributable — but only as current as the indexed corpus (stale documents and un-refreshed indexes both leak through), and the retrieved passages add input tokens.
• Pre-built RAG with Agent Search is the fastest path to grounding on content you supply (first-party or licensed third-party); RAG APIs give developers component-level control over a custom retrieval pipeline.
• Grounding with Google Search grounds responses in fresh public web data and returns source links, but it cannot reach internal, unpublished information — that needs RAG.
• Temperature reshapes how sharply the model favours likely tokens (lower = more consistent, but not guaranteed identical); top-p keeps the smallest set of tokens whose probabilities reach p, so the pool adapts to the model's certainty, unlike top-k's fixed count.
• Safety settings set a blocking threshold per harm category (harassment, hate speech, sexually explicit, dangerous content), tuned per application; a response that hits the max output token limit is cut off mid-sentence.
• Gemini pay-as-you-go text usage is billed per token for both input and output, so token count and output length are how leaders control spend and response size.
• Fine-tuning changes how a model behaves (tone, structure, task style); RAG supplies current, citable facts at request time — and the two can be combined.

Business strategies for a successful gen AI solution

Implementing a Transformational Gen AI Solution: The Google Cloud Steps

• A transformational gen AI solution starts from a business need and works backward to the technology, never the reverse.
• The main solution types are text generation, image generation, code generation, and personalized user needs; personalization means content generated from each person's own data, not ranking existing items or sending everyone their own copy of input-only output.
• Business requirements (outcomes, users, budget, deadlines, compliance obligations) and technical constraints (data quality, integration, skills, infrastructure) jointly determine the right solution; compliance obligations are requirements that narrow the technical options.
• Choose the simplest build depth that meets the requirement: ready-to-use, then grounded, then tuned; training a foundation model from scratch is rarely justified for a business.
• Grounding retrieves current documents at answer time, so revisions flow into answers; tuning shapes behavior at training time and is not a way to keep facts current.
• The recommended integration sequence is: define vision and use cases, assess readiness, select the solution, pilot, establish governance and security, integrate and scale, then measure and iterate.
• A pilot converts uncertainty into evidence cheaply; a pilot that prevents a costly mistake is a success.
• A pilot with fully privileged users never tests access boundaries; design controls for the wider audience before rollout, then expand in stages under them.
• Impact needs KPIs defined and a baseline measured before launch; ROI weighs total value against total cost, including costs paid from other budgets.
• Most gen AI failures are strategic: map a failing initiative back to the recommended step it skipped.

Secure AI and SAIF: Protecting AI Systems on Google Cloud

• Secure AI means protecting an AI system's data, models, infrastructure, and interactions from attacks and misuse across the entire ML lifecycle; it is not AI for security, responsible AI, or perimeter hardening alone.
• Data poisoning lives in training data and is fixed by clean, verified data; evasion crafts inputs against a trained model at inference and survives retraining.
• Prompt injection crafts input to override a system's instructions or extract hidden context; prompt engineering is the legitimate craft of writing better prompts.
• Sensitive data can leak from grounding sources retrieved at answer time, not only from training data; scope grounding to what the requesting user may read.
• Read restrictions and encryption protect confidentiality, not integrity: poisoned data arriving through a legitimate path needs provenance checks and outlier screening.
• Security must cover every lifecycle stage, data, training, deployment, and monitoring; an attacker targets the weakest stage, so that stage defines the system's overall risk.
• SAIF is Google's voluntary framework of six core elements: expand strong security foundations, extend detection and response, automate defenses, harmonize platform-level controls, adapt controls (red teaming, fast feedback loops), and contextualize AI risks in business processes. It issues no certification.
• SAIF's business benefits: reduced breach risk, faster and more confident AI adoption, stronger compliance posture, and preserved customer trust.
• Tools: secure-by-design infrastructure (protected baseline, default encryption), IAM (least privilege, per identity), Security Command Center (organization-wide posture findings such as a public bucket), workload monitoring (fast alerts on what authorized callers actually do).
• Security on Google Cloud is shared: Google secures the platform; the customer governs access, data, and acceptable use. Default encryption does not protect data from anyone the customer has granted access to.

Responsible AI in Business: Transparency, Privacy, Bias, and Fairness

• Responsible AI is a business requirement, not just an ethical ideal: it protects trust, brand reputation, and regulatory standing.
• Google's AI Principles (revised February 2025) rest on three principles: bold innovation, responsible development and deployment, and collaborative progress, together. They guide Google's own work and tooling; the deploying business keeps accountability and legal liability for its use.
• Transparency is system-level (AI is used, how it works); explainability is about why a specific output was produced.
• Anonymization is irreversible for anyone and must deal with quasi-identifiers such as postcode and birth date; pseudonymization keeps a separately held mapping, so the data is not anonymous while re-identification is reasonably possible.
• Prompts are a disclosure: the remedy is an approved enterprise service with data-use commitments plus usage guidance, because output-side controls cannot undo data already sent.
• Fairness is evidenced only by measuring outcomes and errors per group; overall accuracy, dropping demographic fields (proxies remain), representative inputs or an intent review are not evidence.
• Human oversight for high-stakes outputs means review before the output takes effect, tiered by risk; model confidence and safety filters do not substitute for it.
• Per-prediction feature attributions (for example Vertex Explainable AI) explain individual decisions from the model's own computation; a model card is transparency, and a gen-AI-written rationale may not reflect what drove the decision.