Core Gen AI Concepts

Ask five people what AI means and you get five answers, so start with one picture that orders the whole vocabulary. Artificial intelligence (AI) is the broad field of building systems that reason, learn, and act. Machine learning (ML) is the part of AI where a system learns rules from data instead of being programmed with them by hand (What is machine learning?^[1]). Deep learning is a further slice of ML that uses many-layered neural networks, and generative AI (gen AI) is the family of deep-learning models that produce new content: text, images, audio, video, and code. Every ML system is AI, but not all AI is ML, and the diagram's nesting is the mental model the rest of this page reuses. One scope note before going further: this subtopic owns the vocabulary and the general decisions made with it, meaning the terms, how models learn, the ML lifecycle, and the criteria and adaptation ladder for picking a model. The sibling Data in Gen AI subtopic takes the labeled-versus-unlabeled distinction introduced below and turns it into a full assessment of whether a real dataset is usable, while Google Foundation Models names which specific Google family produces what, so meeting a term twice across those pages is deliberate reinforcement rather than a new topic.

Natural language processing (NLP) is the capability of working with human language, reading it and writing it; the language models below are how gen AI delivers NLP today.

Foundation models, LLMs, multimodal, and diffusion

A foundation model is the reusable engine under most gen AI. It is pre-trained once on a broad, massive dataset, then adapted to many tasks rather than being rebuilt for each one (What are foundation models?^[2]). That single fact drives the economics: an organization buys or calls one general model and points it at many problems, instead of funding a bespoke model per use case.

The families you will be asked to tell apart are all foundation models, specialized by the data they handle:

• A large language model (LLM) is a foundation model specialized for text. It predicts and generates language, and it powers chat, drafting, summarizing, and code.
• A multimodal foundation model works across several data types at once, for example reading an image and answering a question about it in text (Generative AI beginner's guide^[3]). Google's Gemini models are multimodal.
• A diffusion model generates images by starting from random visual noise and refining it step by step until it matches the prompt. Google's Imagen is a text-to-image model of this kind (Imagen^[4]).

The leadership takeaway is not the math. It is that 'foundation model' names a shared, reusable asset, and LLM, multimodal, and diffusion are just which content it reads and writes.

A model is only as good as what it learns from, and what it learns from is data, so the first fork is how that data is labeled. Labeled data carries the answer alongside each example: a photo tagged 'cat', a loan marked 'defaulted'. Unlabeled data is the raw example with no answer attached. That single distinction separates the two most common ways models learn (What is machine learning?^[1]).

Supervised learning trains on labeled data to map an input to a known output, which is what you want for classification and prediction: will this customer churn, is this transaction fraud. Unsupervised learning works on unlabeled data to find structure the business never defined in advance, such as natural customer segments or anomalies. Reinforcement learning uses no fixed dataset at all; an agent takes actions in an environment and learns from a reward signal which behavior pays off, the approach behind robotics and game-playing systems. Google groups ML into four types, adding semi-supervised learning, which trains on a small amount of labeled data plus a large pool of unlabeled data to get most of the benefit of labels without the full labeling cost.

Why this matters to a leader, not just a data scientist

Labeling is expensive and slow because a human usually has to do it, so the choice of approach is partly a budget decision. It also sets the ceiling on quality: a model learns whatever its data teaches it, so biased or unrepresentative training data produces a biased model no matter how good the algorithm is (What is machine learning?^[1]). More data only helps when the data is high quality and representative of the real world the model will face. Foundation models are trained with these same methods, only at enormous scale, which is why their training data and its gaps become a governance question for the whole business.

Every model, traditional or generative, moves through the same lifecycle, and Google Cloud offers a managed tool at each stage so a team is not stitching the pipeline together by hand. The five stages are data ingestion, data preparation, model training, model deployment, and model management, and Vertex AI is the unified platform that spans them (Introduction to Vertex AI^[5]). The diagram lays the stages out in order with the main tool for each.

Data ingestion is getting raw data into the cloud. Cloud Storage holds files and unstructured data as the data lake, BigQuery holds structured, analytics-ready data as the serverless warehouse (BigQuery overview^[6]), and Pub/Sub ingests real-time event streams.

Data preparation is cleaning, transforming, and labeling that raw data into something a model can learn from. Dataflow processes and transforms data at scale, BigQuery reshapes structured data with SQL, and Vertex AI managed datasets organize training data with its labels.

Model training is where the model actually learns. Vertex AI training runs custom training jobs, AutoML trains a high-quality custom model on your own labeled data with little ML expertise, and BigQuery ML builds models where the data already lives. These sit on a ladder of effort: call a pre-trained model as-is for no differentiation, use AutoML for a custom model with minimal expertise, or build fully custom on Vertex AI for the most control and differentiation.

Model deployment puts the trained model behind an endpoint so applications can send it new data and get predictions back.

Model management is the ongoing job after launch: the Vertex AI Model Registry versions each model, and monitoring watches live performance for drift, where the world shifts away from what the model was trained on and accuracy quietly decays. Management is not a one-time step; it is the reason a deployed model keeps earning its keep.

The leadership point is where budget, time, and risk actually land. Most of the effort in a real project is the first two stages, ingesting and preparing data, not the training everyone pictures.

An insurer has just approved a gen AI assistant for its claims inbox, and the project immediately hits two decisions in a row: which model to run out of the dozens in the catalog, and what to do when the model it picks is close but not quite right on its own claim letters. Those are the two halves of this section, taken in that order: first the criteria that select a model, then the ladder of ways to adapt the one you selected.

Choosing which model

There is no single best foundation model, so 'which model' is a business decision made against a short list of criteria. Weigh them together rather than optimizing one (Generative AI beginner's guide^[3]):

Criterion	The question it answers
Modality	Does it handle your data type: text, image, audio, video, or several at once?
Context window	How much input can it consider at once, a paragraph or a whole book?
Performance	Is its quality high enough for the task, judged on your own examples?
Cost	What does each call cost, and does that scale to your expected volume?
Availability and reliability	Is it offered in your region with the uptime a production workload needs?
Security and governance	Where does your data go, and does it meet your privacy and compliance rules?
Fine-tuning and customization	Can you adapt it to your domain, or is it fixed as delivered?

Vertex AI Model Garden is the catalog where you compare options side by side, distinct from the Vertex AI Model Registry in the lifecycle above that versions models you have already trained: Google's first-party models (the Gemini, Imagen, and Veo families), partner models, and open models (Model Garden^[7]). The disciplined answer is the cheapest, most controllable model that clears the quality bar on your own tests, not the largest or newest one.

Adapting the model you chose

Once you have a model, there is a ladder of ways to shape its output, and cost climbs as you move up it, so start at the bottom. Prompt engineering is crafting the input text so a frozen model returns better results; nothing inside the model changes, iterating is free and instant, and clear instructions plus a few examples solve a large share of cases (Introduction to prompting^[8]). Prompt tuning is a light training step that learns a small set of soft-prompt values from your examples, steering the model without retraining it. A soft prompt is not words you write: it is a short block of numeric values the tuning step learns and attaches ahead of your input inside the model, which is why the model's own weights stay untouched and the run stays cheap. Fine-tuning goes furthest: it updates the model itself on your labeled data for a durable change in behavior, which costs the most in data, time, and money (Introduction to tuning^[9]). The rule for a leader is to escalate only when the cheaper rung cannot meet the quality or consistency the use case needs. The diagram shows the three rungs in order of rising effort and cost.

Executives do not fund AI, they fund outcomes, so the useful move is to sort every idea into one of four things gen AI actually does: create, summarize, discover, and automate (Generative AI on Google Cloud^[10]).

• Create is generating new content: drafting marketing copy and emails (text generation), producing images and designs (image generation), writing and completing software (code generation), and generating video. This is the most visible use and the easiest to demo.
• Summarize is compressing long material into its essentials: turning a hundred-page contract, a long meeting, or a support thread into the few lines a person needs to act. The value is time returned to expensive people.
• Discover is finding insight buried in messy, mostly unstructured enterprise data: searching across documents, tickets, and knowledge bases in plain language, and pulling out answers with their sources. It turns a data pile no one reads into an answer engine.
• Automate is letting agents carry out multi-step work, not just answer a question: triaging a request, looking things up, and taking the next action across systems.

Personalized user experience cuts across all four, tailoring what is created, summarized, surfaced, or automated to the individual customer. Data analysis also spans them, since summarizing and discovering both turn raw data into decisions.

Two cautions keep these projects honest. First, gen AI is the right tool only when the task genuinely involves generating, condensing, searching, or acting on unstructured content; a deterministic, rules-based process is cheaper as ordinary software, and a structured-data prediction with a clear label is usually a job for traditional supervised ML. Second, every one of these patterns is only as trustworthy as the data behind it, which is why grounding the model in reliable enterprise data matters before you ship.

The Generative AI Leader exam tests whether you can place a scenario in this vocabulary and pick the business-sensible option, not whether you can build anything. A few patterns recur.

Definitions and the nesting

A stem that says 'a subset of AI that learns from data' wants machine learning; 'a subset of ML using layered neural networks' wants deep learning; 'creates new content' wants generative AI. An option that places gen AI outside AI, or treats ML and AI as identical, is wrong. 'One model pre-trained broadly and reused across many tasks' is a foundation model; 'specialized for text' narrows it to an LLM; 'handles text and images together' is multimodal; 'generates images from noise' is diffusion.

Which learning approach

Map the data to the approach. Labeled examples with a target answer point to supervised learning; no labels and a goal of finding structure points to unsupervised; learning from reward through trial and error points to reinforcement. A stem offering unsupervised learning for a labeled prediction task is the classic distractor.

Which model, and how to adapt it

When a question lists modality, context window, cost, and security and asks how to pick, the answer weighs them together and points to comparing options in Model Garden, not to always choosing the biggest model. When it asks how to improve output cheaply, prompt engineering is the first move; fine-tuning is the answer only when the stem stresses a durable, domain-specific behavior change that prompting cannot achieve. Reaching for fine-tuning to fix a one-off phrasing problem is the trap.

Gen AI, or not

If the scenario is a deterministic, rules-based calculation or a structured-data prediction with a clear label, the right answer is conventional software or traditional ML, and 'use a generative model' is the distractor. Gen AI fits creating, summarizing, discovering, and automating over unstructured content.