Text, images, audio, and video must become numbers before a model can work with them. You will see pixels, waveforms, tokens, vectors, and tensors as practical building blocks.

A model’s behavior depends heavily on the examples it sees. This chapter covers datasets, labels, metadata, consent, duplication, bias, and train-test splits using simple cases.

Train a tiny model, watch its loss change, and test it on examples it has not seen. You will connect training, validation, overfitting, underfitting, and generalization to real model behavior.

Generative AI grew through chatbots, statistical language models, neural networks, GANs, transformers, diffusion models, and foundation models. This history explains why today’s tools work the way they do and why scale, data, and hardware matter so much.

Current generative AI work usually mixes notebooks, Python, model hubs, APIs, GPUs, and cloud services. You will run small experiments, call a hosted model, and keep track of inputs, outputs, and costs.

Neural networks learn useful features from data instead of relying only on hand-written rules. This chapter covers layers, activations, gradients, backpropagation, optimizers, and practical training choices.

Autoencoders compress data into smaller representations that can be decoded back into usable outputs. You will use latent spaces to see how models can store style, meaning, and variation.

GANs use a generator and a discriminator in a contest that can create sharp images and realistic samples. This chapter covers adversarial training, mode collapse, stability problems, and where GANs still matter.

Autoregressive models create one piece at a time, using earlier pieces to choose the next one. You will connect n-grams, sequence models, sampling temperature, top-k, and top-p to text generation.

Embeddings place related meanings near each other in a vector space. This chapter shows how embeddings power semantic search, clustering, recommendations, duplicate detection, and grounding for generative systems.

Attention lets a model decide which earlier pieces matter most for the next step. You will trace queries, keys, values, attention weights, and why this idea changed language and multimodal AI.

Transformers combine attention, feed-forward layers, positional information, and large-scale training into a flexible architecture. This chapter covers encoder, decoder, and encoder-decoder designs and why transformers became the base for modern generative AI.

Foundation models are trained broadly and then adapted to many tasks. You will study pretraining, scale laws, emergent abilities, data mixtures, compute budgets, and the tradeoffs between closed and open models.

Tokenization and context length shape what a model can read, remember, and produce. This chapter covers subword tokens, context windows, long-context models, truncation, retrieval, and common failures with long documents.

Good prompts turn vague requests into usable work instructions. You will practice task framing, constraints, examples, output formats, style control, and iteration without treating prompting as magic.

Some tasks improve when a model is given examples, intermediate steps, or a structured way to reason. This chapter covers few-shot prompting, chain-of-thought tradeoffs, self-checking, decomposition, and when these patterns fail.

Modern models can return JSON, call tools, and connect to application logic. You will design schemas, validate outputs, handle errors, and use function calling safely in small apps.

Reasoning-focused models spend more effort at answer time to solve harder problems. This chapter covers test-time compute, planning traces, verification loops, latency tradeoffs, and tasks where reasoning models help or disappoint.

Vector indexes make semantic search fast enough for real applications. You will chunk documents, create embeddings, build an index, run similarity search, and judge whether retrieved results are actually useful.

Retrieval-augmented generation connects a model to external information instead of relying only on memory. This chapter covers retrieval, context assembly, citations, freshness, permissions, and grounded answers.

RAG systems often need more than a single vector search. You will use hybrid search, metadata filters, reranking, query rewriting, and graph-based retrieval to improve difficult knowledge tasks.

Follow a practical workflow from a user question to source selection, chunking, retrieval, answer generation, citation checks, and handoff. The project produces a small assistant that answers from trusted documents instead of guessing.

Domain adaptation starts with clean, relevant, legally usable data. This chapter covers data collection, cleaning, deduplication, formatting, redaction, quality checks, and dataset cards for generative AI projects.

Fine-tuning changes a model’s behavior by training it on task-specific examples. You will compare instruction tuning, supervised fine-tuning, continued pretraining, and the signs that fine-tuning is the wrong tool.

Parameter-efficient tuning updates small parts of a large model instead of retraining everything. This chapter covers LoRA, QLoRA, adapters, rank choices, memory savings, merge decisions, and practical limits.

Human preference training helps models follow instructions, refuse unsafe requests, and produce outputs people prefer. You will compare RLHF, reward models, DPO, constitutional methods, and the risks of over-optimization.

Diffusion models turn noise into images through repeated denoising steps. This chapter covers noise schedules, latent diffusion, prompts, negative prompts, samplers, guidance scale, and why diffusion changed visual generation.

Image generation becomes more useful when you can control layout, pose, depth, style, and edits. You will use inpainting, outpainting, ControlNet-style conditioning, image-to-image workflows, and prompt-image iteration.

Generative audio systems create speech, sound effects, and music from text, voice, or reference audio. This chapter covers spectrograms, vocoders, voice cloning risks, music generation, and quality checks for audio outputs.

Video generation adds motion, timing, continuity, and camera control to image generation. You will work with text-to-video, image-to-video, keyframes, temporal consistency, editing workflows, and current limits.

Generative models can help create 3D objects, materials, scenes, and game assets. This chapter covers text-to-3D, image-to-3D, NeRFs, Gaussian splatting, mesh cleanup, and where human artists still make key decisions.

Multimodal models can read, describe, compare, and generate across text, image, audio, and video. You will build tasks that combine modalities, such as visual question answering, document reading, captioning, and image-grounded chat.

AI coding tools generate snippets, tests, explanations, migrations, and whole files, but they also create subtle bugs. This chapter covers code prompting, repository context, review habits, test generation, and safe use in real development.

Agents use models to choose steps, call tools, inspect results, and continue toward a goal. You will build a small tool-using agent and handle planning failures, loops, permissions, and recovery.

Longer AI workflows need memory, routing, state, and coordination between components. This chapter covers agent orchestration, task queues, multi-agent patterns, human approval steps, and when simpler pipelines are better.

Synthetic data can fill gaps, protect privacy, stress-test systems, or accidentally amplify mistakes. You will create synthetic examples, filter them, measure diversity, and decide when generated data is safe enough to use.

Generative AI evaluation must judge usefulness, truthfulness, safety, style, latency, and cost. This chapter covers human review, rubrics, test sets, model-graded evaluation, benchmark limits, A/B tests, and regression testing.

Safety work reduces harmful, biased, private, or policy-violating outputs. You will use content filters, refusal design, red teaming, jailbreak tests, moderation models, and escalation paths for human review.

Generative AI systems face attacks through prompts, tools, data, models, and dependencies. This chapter covers prompt injection, data exfiltration, poisoned retrieval sources, insecure plugins, model theft, and supply-chain checks.

Serving generative models requires choices about APIs, model size, batching, caching, streaming, rate limits, and fallbacks. You will design an inference path that balances quality, speed, reliability, and cost.

Open models can run cheaper, closer to private data, or on smaller devices when optimized well. This chapter covers quantization, pruning, distillation, speculative decoding, GPU memory, and on-device deployment tradeoffs.

Production systems need continuous measurement after launch. You will track quality, drift, user feedback, safety incidents, latency, token use, cost, and version changes with dashboards and alerts.

Generative AI projects must handle privacy, copyright, licensing, consent, data retention, disclosures, and changing AI regulations. This chapter turns legal and policy duties into design decisions, records, and review checkpoints.

Some AI features should not be automated, and others need human review, transparency, or a narrow scope. This chapter covers risk tiers, user trust, accessibility, failure modes, incident response, and responsible release decisions.

Follow a full product workflow from problem framing through model choice, prototype, data preparation, evaluation, safety review, deployment, monitoring, and updates. The goal is a realistic plan for shipping a useful generative AI feature without skipping hard parts.

Generative AI work includes research, engineering, product, design, data, safety, policy, and domain-specialist roles. This closing chapter covers portfolio projects, proof-of-skill artifacts, certifications, open-source contribution paths, and habits for staying current as the field changes.