Section 1: Fundamentals of gen AI

1.1 Core gen AI concepts and use cases

Exam Guide §1.1 Defining core gen AI concepts (AI, NLP, ML, generative AI, foundation models, multimodal foundation models, diffusion models, prompt tuning, prompt engineering, LLMs) · ML approaches (supervised, unsupervised, reinforcement) · ML lifecycle stages and Google Cloud tools · choosing a foundation model · business use cases (create, summarize, discover, automate).

Hierarchy of terms: AI → ML → Deep Learning → Gen AI

• Artificial intelligence (AI) is the overarching field: building machines that perform tasks that normally require human intelligence (learning, problem-solving, decision-making).
• Machine learning (ML) is a subset of AI: machines learn from data instead of being explicitly programmed, and use this to perform specific tasks. Consequence: different training data produces a different model. (consequence)

Foundation models, LLMs, multimodal and diffusion models

• Foundation models are large, pre-trained ML models trained on massive (mostly unlabeled) datasets. As a result they develop a broad "understanding of the world" and can be adapted to many tasks. They are the basis of gen AI.
• Three key features according to the Study Guide:
  • Trained on diverse data: learn general patterns and relationships.
  • Flexible: one model, many use cases.
  • Adaptable: can be specialized for specific domains through additional, targeted training (tuning/fine-tuning).

Prompting, prompt engineering, prompt tuning

• Prompt / prompting: prompting is the way to interact with and steer foundation models: you provide instructions or inputs to obtain the desired outputs. A prompt is this input (instruction, question, file, example); its purpose is to give the model input and trigger an output.
• Prompt engineering, according to the Study Guide, is the "art and science" of crafting effective prompts to maximize the value of gen AI models and tailor responses to specific needs. Prompt engineering does not change the model. Details on the techniques are in → 3.2.

ML approaches: supervised, unsupervised, reinforcement

• Supervised learning trains on labeled data (input + known correct output) and predicts outputs for new inputs. Typical tasks: classification (spam / not spam), regression (price), prediction (machine failure).
• Unsupervised learning uses unlabeled data to find natural groupings and patterns. Typical tasks: exploratory analysis, clustering, anomaly detection, topic modeling.

ML lifecycle and Google Cloud tools per stage

• Exam Guide wording of the stages: data ingestion → data preparation → model training → model deployment → model management. The Study Guide combines the first two stages into "Data ingestion and preparation".
• Data ingestion: collect data suited to the desired outcome and bring it into the platform (batch or streaming). Typical Google Cloud tools:
  • Pub/Sub: messaging service for real-time streaming based on the publish/subscribe principle
  • Cloud Storage: object storage, mainly for unstructured data such as images, PDFs, audio, video
  • Cloud SQL and Spanner: relational databases for structured data; Spanner is additionally globally scalable

Choosing the right foundation model

• Exam Guide factors: modality, context window, security, availability and reliability, cost, performance, fine-tuning and customization. The Study Guide explains modality, context window, performance, and availability and reliability. This knowledge base adds ease of integration.
• Token = unit of text (word or part of a word), the unit of measure for context window, traffic and billing. For Gemini, one token corresponds to about 4 characters; 100 tokens are about 60–80 English words. → 3.3 token count

Business use cases: create, summarize, discover, automate

• The four main ways to use gen AI (examples our own):
  • Create: generate new content; e.g. a publisher has blurbs and social media teasers drafted, a skilled-trades business has quotation texts drafted, a development team generates code
  • Summarize: condense information into concise summaries; e.g. a pharmaceutical company summarizes long study reports, a university summarizes minutes of committee meetings
  • Discover: find information at the right time; e.g. a logistics company searches thousands of freight documents for clauses, an energy utility identifies patterns in fault reports
  • Automate: automate previously manual tasks; e.g. an insurer automatically transfers details from claim forms into its policy system, notifications are triggered automatically
• Exam Guide example categories: text generation, image generation, code generation, video generation, data analysis, personalized user experience.

1.2 How data types are used in gen AI and the business implications

Exam Guide §1.2 Characteristics and importance of data quality and data accessibility (completeness, consistency, relevance, availability, cost, format) · structured vs. unstructured data with real-world examples · labeled vs. unlabeled data.

Data quality

• ML models analyze data, recognize patterns and use them to calculate probabilities for new cases. That is why data quality and accessibility directly determine the quality of the results; the performance of foundation models depends heavily on the data.
• According to the Study Guide, quality data is accurate, complete, consistent, relevant. The Exam Guide names completeness, consistency, relevance. This knowledge base teaches five dimensions, adding representativeness:
  • Accuracy: the data is factually correct. Faulty data leads to wrong patterns and predictions.
  • Completeness: the size of the dataset and the coverage within it; enough data without gaps for a reliable prediction.
  • Representativeness: the data reflects the target group or target situation fairly and inclusively; otherwise skewed samples and bias.
  • Consistency: uniform formats, units and labels across all sources; inconsistency confuses the model.
  • Relevance: the data fits the task the model is supposed to perform.

Data accessibility

• Accessible data: data for training must be readily available, usable and in the right format. Three dimensions according to the Exam Guide (explanation):
  • Availability: without available data there is no training and no prediction, both in training and in production.
  • Cost: obtaining, cleaning and labeling high-quality data can be a significant hurdle.
  • Format: data must be in a format the model can process. Neural networks ultimately compute only with numbers; text, images and audio are therefore translated into numerical representations (e.g. embeddings).
• Accessibility is technical (interfaces, storage location, format) and non-technical (permissions, licenses, data protection).

Structured vs. unstructured data

• Structured data: organized, easily searchable, often stored in relational databases/tables. Examples: customer ID, name, purchase date, order value, star rating 1–5, transactions, sensor readings in tables.
• Unstructured data: no predefined structure, requires sophisticated analysis methods. Examples: free-text feedback, email content, product images, audio, video, PDFs, social media posts.

Labeled vs. unlabeled data

• Labeled data: data with assigned tags (name, type, number) that give it meaning, e.g. "spam / not spam" or "sale price". Enables supervised learning. Labeling is expensive and time-consuming.
• Unlabeled data: raw, untagged information that on its own carries no meaning / "correct answer", e.g. unsorted photos or audio recordings. Basis for unsupervised learning. Foundation models are pre-trained mostly on unlabeled data.

1.3 Core layers of the gen AI landscape and business implications

Exam Guide §1.3 Infrastructure · Models · Platforms · Agents · Applications

The five-layer stack

• Applications: the application provides the user interface and overall purpose; agents are intelligent components within an application. Agents define the capabilities (search, book, recommend), the application makes them available to the user (website/app).
• Agents: a gen AI agent is an application that tries to achieve a goal by observing the world and acting upon it using the tools it has available.

1.4 Use cases and strengths of Google's foundation models

Exam Guide §1.4 Gemini · Gemma · Imagen · Veo

Other Google models from the Study Guide and documentation (not listed in the EG)

• Gemini Nano: Google's most efficient, most compact model, designed specifically to run on devices (on-device/edge). Advantages: privacy (data stays on the device), speed (no cloud round trip), offline use. On Android it runs in the AICore system service; developers use it via the ML Kit GenAI APIs (e.g. summarizing, rewriting, image description).
• Gemini variants Pro vs. Flash/Flash-Lite: Pro for complex tasks and demanding reasoning (higher latency, considerably more expensive), Flash/Flash-Lite for fast, low-cost responses under high load (chat).

Gemini app vs. Gemini model

• Gemini (model) is a gen AI model from Google that powers many different solutions. The Gemini app is Google's gen AI chatbot (an application) for writing, summarizing, translating and image generation. Gemini is also built into Google Workspace with Gemini, Gemini for Google Cloud (including data analysis in BigQuery), etc.