Gen AI Fundamentals

Every question this domain asks reduces to one habit: name the capability a business actually needs, then match it to the cheapest tool that delivers it. That starts with getting the words straight. Artificial intelligence (AI) is the broad field of systems that reason, learn, and act; machine learning (ML) is the subset that learns rules from data instead of being hand-programmed; deep learning is the subset of ML built on neural networks; and generative AI (gen AI) is the deep-learning family that produces new content such as text, images, audio, video, and code. These terms nest inside one another rather than compete, and a foundation model is simply one pre-trained once on broad data and then adapted to many tasks (Generative AI on Vertex AI). Hold that picture and the classic trap loses its grip: the exam rewards the fit-for-purpose choice over the newest or largest model, and it rewards knowing when the honest answer is that gen AI is the wrong tool for the job.

The page hangs together as four lenses a leader brings to any gen AI decision, in order. Core Gen AI Concepts is the vocabulary: what the terms mean, the three ways models learn (supervised, unsupervised, reinforcement, plus Google's fourth, semi-supervised), how you adapt a model (try prompt engineering before you pay to fine-tune), and the four ways gen AI creates value, namely create, summarize, discover, and automate. Data in Gen AI is the fuel: judge any dataset on quality, structure, and labeling, and recognize that reading unstructured content directly is gen AI's real unlock. Gen AI Landscape is the market map: a five-layer stack from Infrastructure up through Models, Platforms, Agents, to Applications, with the rule to operate at the highest layer that meets the requirement. Google Foundation Models then zooms into the Models layer of that stack, sorting the four families that generate content, Gemini, Gemma, Imagen, and Veo, by what each one produces; generating content is the axis that produces exactly those four, and Google's catalog is wider, holding non-generative foundation models such as Chirp for speech recognition too (Model Garden).

One instinct settles most questions where two options are both technically correct. Prefer the choice that fits the job and is no bigger than it needs to be: the cheapest model that clears your quality bar over the most capable one, the highest stack layer that meets the requirement over building lower, and prompting over tuning until prompting genuinely falls short. The same discipline knows when to say no, because a deterministic, rules-based task is cheaper and safer as ordinary software, and predicting a number or a category from tidy records where past cases already carry the right answer (a labeled structured-data prediction, the phrasing the exam uses) is usually a job for traditional ML rather than gen AI. Descend the stack, fine-tune a model, or reach for a larger tier only when a concrete requirement such as data control, differentiation, or scale forces it, not as a default.