Fundamentals of Generative AI

What is generative AI?

Generative AI creates new content (text, code, images, audio, structured answers) from learned patterns—usually driven by large models and prompts rather than only classifying existing rows.

flowchart LR
  U["Prompt / instruction"] --> T[Tokenization]
  T --> FM["Foundation model"]
  FM --> O["Generated output"]
        

Definitions

Generative AI (GenAI)

Systems that produce new samples or completions (sequences, pixels, waveforms) conditioned on input context—versus only predicting a fixed label.

Prompt

User or system text (or multimodal input) that steers the model toward a task, tone, format, or constraints.

Completion

The model’s continuation of the prompt—output tokens generated step by step.

Token

The atomic unit a model reads/writes (often subword pieces); prompts and outputs are split into tokens for processing and billing.

Chunking · Embeddings · Vectors

Chunking splits long documents for retrieval and context limits. Embeddings map text/images into numeric vectors so “similar meaning” is nearby in vector space—used heavily in search and RAG (Domain 3 goes deeper).

flowchart TB
  DOC["Long document / corpus"] --> CH[Chunking into segments]
  CH --> E["Embedding model"]
  E --> V["Vectors stored for similarity search"]
        

Definitions

Chunking

Breaking content into smaller pieces (by size, sentences, or structure) so each piece fits model context windows and can be retrieved or cited independently.

Embedding

A learned numeric representation of data in a high-dimensional space; similar inputs get similar vectors—used for semantic search, clustering, and retrieval.

Vector

An ordered list of numbers representing a chunk, query, or item; distance metrics (e.g. cosine) measure similarity.

Context window

Maximum tokens a model can attend to in one forward pass; larger docs must be chunked, summarized, or retrieved selectively.

Transformers · LLMs · Foundation models

Most modern text GenAI uses transformer architectures. An LLM is a large language model; a foundation model (FM) is a broad base model (often pre-trained) you can adapt with prompts, tools, RAG, or fine-tuning.

flowchart TB
  TR["Transformer blocks attention FFN"] --> LLM["Large language model LLM"]
  LLM --> FM["Foundation model general capabilities"]
  FM --> USE["Prompt RAG fine-tune agents"]
        

Definitions

Transformer

Neural architecture using self-attention to relate all tokens in context—scalable basis for modern LLMs.

Large language model (LLM)

A very large transformer (or mix of architectures) trained heavily on text to model language, knowledge, and task patterns.

Foundation model (FM)

A broad, often pre-trained model that can underpin many tasks (Q&A, summarization, extraction) via prompting and customization paths.

Prompt engineering

Designing instructions, examples, and structure in prompts to improve quality, reliability, and safety—without changing model weights (in-context).

Multimodal models · Diffusion models

flowchart LR
  IMG[Image] --> MM["Multimodal FM"]
  TX[Text] --> MM
  MM --> OUT["Text caption answer or gen"]
            

flowchart LR
  Z[Random noise] --> S["Many denoising steps"]
  S --> G[Generated image or audio]
            

Definitions

Multimodal model

A model trained to align or fuse signals across modalities (text, vision, audio) for understanding or generation.

Diffusion model

A generative approach that learns to reverse a noise-corruption process—common for high-quality image generation and some audio models.

Modality

The type of data: text, image, speech, video, code, etc.—affects model choice and evaluation.

Foundation model lifecycle

From raw corpora to production: select data and base model, pre-train, optionally fine-tune, evaluate, deploy, collect feedback. Most practitioners consume FMs from providers; fewer build from scratch.

flowchart LR
  DS[Data selection] --> MS[Model selection]
  MS --> PT[Pre-training]
  PT --> FT[Fine-tuning optional]
  FT --> EV[Evaluation]
  EV --> DP[Deployment]
  DP --> FB["User feedback and monitoring"]
  FB -.-> EV
        

Definitions

Pre-training

Large-scale training (often self-supervised on broad data) to learn general representations before task-specific adaptation.

Fine-tuning

Additional training on a narrower dataset or objectives to specialize a model (domain, tone, format)—contrasts with in-context learning via prompts alone.

Continuous pre-training

Further pre-training on new corpora to refresh knowledge while retaining base capabilities—requires care and governance.

RLHF (high level)

Reinforcement learning from human feedback: align model outputs with human preferences using reward modeling—exam may name it as a fine-tuning/alignment technique.

Where GenAI creates value

Match modality and risk to the task: drafting, summarization, translation, code assist, search assistance, contact-center augmentation, content generation, personalization—always with human review where stakes are high.

flowchart TB
  subgraph val["High-value patterns"]
    D1["Draft and refine text"]
    D2["Summarize long content"]
    D3["Classify extract structure"]
    D4["Code suggestion and tests"]
    D5["Conversational assistants"]
  end
        

Definitions

Summarization

Condensing source material into shorter faithful text—may be extractive (select spans) or abstractive (new phrasing).

Translation / localization

Converting content across languages or regional style; quality and domain vocabulary matter.

Code generation

Producing or editing source from natural-language intent; must be validated with tests and review.

Conversational agent

A dialogue system using an FM (often plus tools) to hold multi-turn tasks—exam links this to agents in later domains.

Strengths vs limits

GenAI is fast and flexible but not an oracle. Expect hallucinations, nondeterminism (temperature), fairness gaps, and limited verifiability without retrieval or tools.

flowchart LR
  subgraph pros["Advantages"]
    P1["Broad generalization"]
    P2["Rapid prototyping"]
    P3["Natural language interface"]
  end
  subgraph cons["Risks and limits"]
    C1["Hallucinations"]
    C2["Nondeterminism"]
    C3["Opaque reasoning"]
    C4["Data and license sensitivity"]
  end
        

Definitions

Hallucination

Plausible-sounding but false or unsupported statements—critical risk for factual, legal, or medical use cases.

Nondeterminism

With sampling (temperature > 0), the same prompt can yield different outputs across runs—matters for testing and compliance.

Interpretability gap

Difficulty explaining why a specific token was generated; mitigated with citations (RAG), structured outputs, and audits.

Inaccuracy / staleness

Model knowledge may be outdated or wrong for niche domains—mitigate with retrieval, fine-tuning, or tool use.

Choosing models · Business metrics

Balance capability, latency, cost, context length, license, safety, and compliance. Tie pilots to business KPIs, not only novelty scores.

flowchart TB
  Q{"Model choice drivers"} --> A[Task modality]
  Q --> B[Latency and throughput]
  Q --> C[Cost tokens and hosting]
  Q --> D["Compliance and data residency"]
  Q --> E["Customization needs RAG fine-tune"]
        

Definitions

Latency

Time to first token or full response—drives UX for interactive apps.

Throughput

Tokens or requests processed per second—matters at high volume.

Token-based pricing

Billing by input/output tokens consumed; major driver of cloud FM cost on fully managed APIs.

Provisioned throughput

Reserved model capacity for predictable performance and spend—versus on-demand bursting.

Business value metrics

Examples: handle time, conversion, cost per resolution, employee productivity, error reduction, revenue per user—paired with quality checks.

Services you should recognize

Amazon Bedrock is the FM hub (multiple providers, customization paths). SageMaker JumpStart brings models and notebooks. PartyRock is a playground for quick experimentation. Amazon Q is the assistant experience across productivity and AWS.

flowchart TB
  subgraph apps["Applications and UX"]
    Q["Amazon Q"]
    PR["PartyRock playground"]
  end
  subgraph fm["Foundation models"]
    BR["Amazon Bedrock FMs tools guardrails"]
  end
  subgraph build["Build and ops"]
    JS["SageMaker JumpStart"]
    SM["SageMaker training and hosting"]
  end
  apps --> BR
  build --> BR
        

Definitions

Amazon Bedrock

Fully managed service to invoke and customize foundation models from multiple providers, often with features for knowledge bases, agents, and guardrails.

SageMaker JumpStart

Curated models, solutions, and notebooks inside SageMaker to start quickly with pre-trained models and examples.

PartyRock

Amazon’s Bedrock-based playground for low-friction GenAI app experiments and learning (naming and features evolve—verify in AWS docs).

Amazon Q

AWS’s generative assistant product family for business and builder workflows—integrates with enterprise data and AWS consoles per AWS positioning.

Guardrails (Bedrock)

Configurable policies to filter undesirable topics, PII, or toxicity—part of responsible deployment on AWS.

Why AWS for GenAI · Cost dimensions

AWS stresses security, compliance footprint, global regions, and integration with identity, storage, and data services. Costs hinge on tokens, infrastructure, customization, and reliability choices.

flowchart LR
  subgraph savings["Cost and ops levers"]
    L1["On-demand vs provisioned throughput"]
    L2["Model size vs quality"]
    L3["Caching and prompt design"]
    L4["Region and redundancy choices"]
  end
        

Definitions

Shared responsibility model

AWS secures the cloud infrastructure; customers secure what they put in the cloud (data, IAM, app config)—applies to GenAI workloads.

Regional coverage

Choosing an AWS Region for data residency, latency, and which Bedrock models/features are available there.

Redundancy / availability

Design for fault tolerance (Multi-AZ, retries); higher resilience generally increases cost.

Custom model on Bedrock

Fine-tuned or continued-pretrained variants with their own hosting and pricing profile—exam may contrast with base on-demand inference.

Master glossary (Domain 2)

Quick-review sheet aligned to Domain 2 task statements.

GenAI · Prompt · Token · Completion

Produce new content; steering text; billing/read unit; model output stream.

Chunking · Embedding · Vector

Split long content; semantic numeric representation; similarity search unit.

Transformer · LLM · Foundation model

Attention-based architecture; large text model; broad base for many tasks.

Prompt engineering · Context window

In-context control; max tokens per call.

Multimodal · Diffusion

Multiple input/output types; denoising generative process.

Pre-training · Fine-tuning · RLHF

General training; specialization; human-preference alignment.

Hallucination · Nondeterminism

False fluency; varying outputs at same prompt.

Latency · Throughput · Token pricing · Provisioned throughput

Speed; scale; pay per token; reserved capacity.

Bedrock · JumpStart · PartyRock · Amazon Q

FM service; curated SageMaker starters; playground; assistant UX.

Scroll if needed · Next: recap and Domain 3 preview

Before Domain 3 (Foundation model apps)

• Define tokens, embeddings, vectors, and chunking; explain why chunking matters
• Contrast transformer LLM vs foundation model; name two non-text modalities
• List three GenAI strengths and three risks (hallucination, nondeterminism, …)
• Name four levers for model selection (latency, cost, compliance, customization)
• Map a scenario to Bedrock vs JumpStart vs Q at a high level

Official: AIF-C01 exam guide · Domain 3 covers RAG, agents, prompts in depth