AWS AI Practitioner Notes: models, prompts, and controls

Module 1: Terms and learning types

Artificial intelligence is the broad field of systems that perform tasks associated with human judgment. Machine learning is the part that learns patterns from data instead of following only hand-written rules. Deep learning is machine learning that uses neural networks with many layers. Generative AI creates new content, such as text, images, audio, or code. Agentic AI uses a model to plan, call tools, and carry out steps. An agent is not simply a larger language model.

A model is the learned result. An algorithm is the procedure used to learn it. Training builds or updates the model. Inference is using the trained model to produce a result. Fit describes how well the model matches the data you care about. Overfitting memorizes the training set and fails on new examples. Underfitting is too simple and fails on both.

Data can be labeled or unlabeled, and structured or unstructured. Tabular rows, time series, images, and text are all in scope. Labeled text is still text. A table of numbers is structured. A contract PDF is unstructured until something extracts fields from it.

Module 3: The AI/ML lifecycle and metrics

An AI pipeline collects data, prepares it, trains or selects a model, evaluates it, deploys it, and monitors it. Version 1.1 asks you to describe and differentiate those parts. It does not ask you to perform feature engineering or hyperparameter tuning. Those tasks are out of scope for the target candidate. Treat a question that asks you to tune hyperparameters as a distractor, not as the job this exam is testing.

Foundation models can come from an open pretrained model or from training your own. In production you either call a managed API, such as Amazon Bedrock, or you host the model yourself. Services named for pipeline stages include Amazon Bedrock, Amazon Q, Amazon Quick, Kiro, and SageMaker AI.

MLOps is the set of habits that keep that pipeline repeatable: experimentation, repeatable processes, systems that scale, control of technical debt, production readiness, model monitoring, and retraining when the data drifts.

Module 5: Tokens, context, and agents

Token-based pricing charges for input tokens and output tokens. A longer prompt and a longer answer cost more and can add latency. Prompt caching reuses a repeated prefix so you do not pay the full input price every time. Provisioned throughput reserves capacity. On-demand token pricing charges for what you use. A custom model can change both the quality and the price. Regional coverage, redundancy, and availability are the other cost tradeoffs named in the guide.

Context engineering is choosing and arranging what the model sees: instructions, retrieved passages, history, and tool results. It is broader than a single user sentence. A sloppy context window is why a model answers from the wrong document or ignores the instruction.

Agentic systems break a complex job across agents. Patterns include a coordinator plus specialists, shared or separate memory, tool use, and workflow orchestration. The Model Context Protocol is a way for an agent to connect to external tools and data in a standard way. It is not a foundation model, an IAM policy, or a vector database. Multi-agent communication is how those agents pass work. Memory is how an agent keeps facts across steps. Tools are the actions it can take, such as a search or a ticket update.

Module 7: Choosing and customizing a foundation model

Selection criteria in the guide are cost, modality, latency, multilingual support, model size, complexity, customization, input and output length, and prompt caching. Inference parameters change the answer without training. Temperature makes output more varied when it is high and more stable when it is low. Input and output length change both what fits in context and what you pay.

Retrieval Augmented Generation finds relevant chunks and puts them in the prompt. Amazon Bedrock Knowledge Bases is the managed example. Use it when the answer must come from your documents and you do not want to change model weights. Embeddings for that search can live in the vector-capable databases named above.

Data for fine-tuning needs curation, governance, enough size, labels where the method requires them, and a set that represents the people the model will serve. Reinforcement learning from human feedback uses human preferences to align the model. An agent is the pattern for a multi-step business task: it keeps a goal, calls tools, and stops when the task is done. Version 1.1 asks you to define that role and its business applications, not to configure an agent console.

Module 9: Evaluating foundation models

Evaluate with people in the loop, with benchmark datasets, and with Amazon Bedrock Model Evaluation. Text metrics in the guide are ROUGE, which scores overlap for summaries, BLEU, which scores overlap for translation, BERTScore, which compares embeddings, and LLM-as-a-judge, which asks a model to score another model's output. Version 1.1 added LLM-as-a-judge.

A benchmark can look strong while the application fails. Also check productivity, user engagement, and whether the task is actually completed. For a retrieval system, an agent, or a workflow, evaluate the application, not only the base model. Business-alignment metrics named in version 1.1 are task completion rate, user satisfaction, and cost per interaction.

Module 11: Security, grounding, and governance

The candidate is not expected to implement a security program. The exam asks which control matches the risk. AWS secures the infrastructure of the managed service. You secure the data, the prompts, the identities, and how outputs are used. That is the same shared-responsibility idea as the rest of AWS, applied to AI.

Data governance covers the lifecycle, logging, residency, monitoring, observation, and retention. A prompt log that you keep forever can become the sensitive-data problem. Governance processes in the guide include policies, a review cadence, review strategies, the Generative AI Security Scoping Matrix, transparency standards, and training for the team. You are asked to recognize those processes, not to author a framework.

Extra Amazon GuardDuty, AWS Shield, and AWS WAF are out of scope for AIF-C01. Secure data engineering ideas that are in scope stay at the recognition level: data quality, privacy-enhancing technologies, access control, and integrity.

Frequently asked questions