Generative AI Leader Cheat Sheet

A large-scale, pre-trained, general-purpose model adaptable to multiple downstream tasks

Every LLM is a foundation model specialized in text; not every foundation model is an LLM (e.g. Imagen)

Processes and/or generates more than one data type at once (text+image, text+video). Gemini is Google's example

Architecture that gradually reverses noise to generate image/video (Imagen, Veo)

Tuning = trains prompt embeddings (requires training). Engineering = hand-writing the text (no training)

Data collection → preparation → training/selection → deployment → continuous monitoring

Structured = rows/columns (tables, SQL). Unstructured = free text, images, audio

Labeled = includes the correct answer (supervised). Unlabeled = the model discovers patterns on its own (unsupervised)

Infrastructure → Models → Platforms → Agents → Applications

Google designs products with AI as a core criterion, not an afterthought; pioneer of the Transformer architecture

A responsible, secure (SAIF), private, and scalable platform, built for the requirements of large organizations

Google's own chip designed to accelerate machine learning operations

A supercomputing system that integrates hardware, software, and ML frameworks for large-scale AI workloads

Gemma (open weights) + Model Garden (third-party models) + open source standards (Kubernetes, TensorFlow)

Low-code/no-code (Agent Studio) + pre-trained models + simple APIs, without requiring ML expertise

Enterprise platform: connects the model to internal data, multimodal search, and custom agents

Packaged customer experience solutions (applications layer of the gen AI landscape)

Developer tools: Model Garden, Agent Search (component), and Agent Platform AutoML

Extensions, functions, data stores, and plugins — the pieces an agent uses to act on its environment

Agent Studio = build complete enterprise agents. Google AI Studio = quickly prototype prompts/models

Environment for prototyping prompts and testing models within Agent Platform, formerly Vertex AI (platforms layer)

Combines OCR with natural language understanding to process documents (invoices, forms)

Data dependency, knowledge cutoff, bias, fairness, hallucinations, edge cases

They differ only in the number of examples included: 0, 1, or several

Orthogonal techniques: they can be combined with zero/one/few-shot in the same prompt

Asks for step-by-step reasoning before the answer; improves math and logic tasks

Combines reasoning with external actions (function calling) and observation of results

First-party (own), third-party (licensed/third-party), world data (real-time web search)

Retrieves relevant fragments from an external source and uses them as context before generating the response

Token count, temperature, top-p, safety settings, output length

Google's principles for deploying AI securely by design: expanded controls, automated defenses, harmonization, continuous adaptation

Anonymization = irreversible. Pseudonymization = reversible with a key (still personal data)

The organization that deploys the system — not the model — is responsible for its outcomes

Being able to describe in understandable terms which factors influenced a system's decision

Human review and approval before using a high-risk output; reinforces accountability

Manipulating the model's behavior with hidden instructions in the prompt or in a processed document (RAG)

The official exam guide is explicit: strategic leadership, not technical implementation — don't expect code or CLI questions

The "Google Cloud gen AI options" domain has the highest relative weight on the exam (~35%)

Common trap: Vertex AI Studio (Google Cloud, enterprise) ≠ Google AI Studio (fast prototyping, lighter)

Distinguish Gemini app/Google AI Pro (consumer), Gemini Enterprise (enterprise), and Gemini in Workspace (productivity)

Grounding is the general concept of anchoring responses to real data; RAG is a specific technique for achieving it

If no training is involved, it's prompt engineering; if embeddings are trained, it's prompt tuning