AIF-C01 flashcards

Fundamentals of AI and ML

• AI contains ML, ML contains deep learning, and generative AI is built on deep learning; agentic AI is GenAI that plans and acts using tools.
• Predictive ML outputs a label or score for existing input; generative AI creates new content; generating one response is not agentic.
• A system is agentic when the model chooses the steps; tool calls on a fixed, code-defined path are a workflow, and an RL agent trained by rewards is a different idea.
• Deep learning learns features from raw inputs but needs more data and compute and is harder to explain; tree-based methods suit small tabular data with engineered features.
• An algorithm is the learning method, a model is the trained result, and hyperparameters are settings chosen before training. Training builds the model; inferencing uses it without changing it.
• Supervised learning needs externally labeled data; unsupervised finds structure in unlabeled data; self-supervised makes its own labels from raw data; reinforcement learning learns from rewards.
• Structured versus unstructured is about format (tabular and time-series versus image, text, and audio); labeled versus unlabeled is about whether answers are provided.
• Batch = many records offline; real-time = instant answer on provisioned capacity; asynchronous = large or slow jobs queued; serverless = quick requests, scales to zero, cold start after idle.
• Overfitting (linked to high variance) means great on training data but poor on new data; underfitting means poor on both.
• Bias is a systematic skew that disadvantages a group; fairness is the goal of equitable treatment; explainability is understanding why a model decided.
• Computer vision handles images, NLP handles language, an LLM generates text, an embedding model turns text into vectors, and a diffusion model generates images.

• AI/ML adds value when a task needs prediction or pattern-finding at scale, assists human decisions, or automates repetitive work.
• AI/ML is the wrong tool when a guaranteed exact answer is required, when a simple rule already works, or when the cost and data needed outweigh the benefit; tabular data, a human reviewer, or a team without data scientists are not reasons to avoid it.
• Regression predicts continuous numbers, classification assigns known categories, clustering groups a whole unlabeled population, and anomaly detection flags the unusual few.
• Match input and output to the service: speech to text is Transcribe, text to speech is Polly, text translation is Translate, sentiment, entities and language detection are Comprehend, conversational bots with intents and slots are Lex.
• Textract understands document structure (forms and tables); Rekognition analyzes images and can detect text in them, but not form fields.
• Audio must be transcribed before it can be translated or analyzed; Transcribe Call Analytics can return call sentiment by itself.
• Agentic AI takes actions in other systems through tools; a multi-step assistant that only searches and summarizes is a knowledge base.
• Use pre-trained managed services for common tasks; use SageMaker AI to build, train, and deploy a custom model on your own data. Amazon Forecast is closed to new customers.
• Traditional ML fits per-decision explainability, regulated automated decisions, and cheap high-volume scoring; a foundation model fits open-ended generation, scarce labels, and many tasks in one model. An LLM-written rationale or RAG citation is not a per-decision explanation.

• The AI/ML pipeline runs in order and loops: collect (and label) data, prepare it, train, evaluate and tune, deploy, monitor, then retrain — and a retrained model is evaluated before it is redeployed.
• Three ways to source a foundation model: use a provider's managed FM (Amazon Bedrock), adapt an open-source pre-trained model, or train a custom model — moving from least to most control, cost, and effort.
• A managed API service means low operational overhead but less control; a self-hosted API gives full control and data isolation but you run the infrastructure.
• Amazon S3 is the data lake; Ground Truth labels, Data Wrangler prepares, Feature Store serves the same feature values to training and inference, Model Registry versions models, and Model Monitor watches them.
• SageMaker AI builds, trains and hosts models (including JumpStart foundation models on endpoints you manage); Bedrock serves foundation models through one serverless API.
• Amazon Q Developer helps builders and operators on AWS, Q Business answers over company content, Amazon Quick is the agentic workspace for business users, and Kiro is the spec-driven development environment.
• MLOps means tracked experiments, automated pipelines, versioned data, code and models, rollback, and monitoring against a baseline — bigger instances and more tuning do not make work reproducible.
• Gradual quality decline after a good launch is drift (retrain); overfitting and bad labels show up at evaluation; latency under load is a capacity problem (scale the endpoint).
• Precision matters when false positives are costly; recall when misses are costly; F1 when both are and classes are imbalanced; accuracy alone misleads on imbalanced data.
• Business metrics — cost per user, development cost, customer feedback, and ROI — decide whether a model is worth running; a more accurate model can lose money per use.

Fundamentals of GenAI

• Generative AI creates new content; a foundation model (FM) is the broadly pre-trained base, and an LLM is a foundation model specialized for text.
• A token is the small unit of text a model reads and writes; the context window is counted in tokens, and a chat that resends its whole history grows each turn until it overflows that window.
• Chunking splits a large document into pieces; each chunk gets a focused vector that matches specific questions more precisely than one vector for a whole document.
• An embedding is a fixed-length vector that represents meaning; similarity is how close whole vectors are (distance or angle), not shared numbers.
• Prompt engineering shapes instructions, context, and examples without changing the weights; temperature and maximum output tokens control randomness and length, not knowledge, format, or tone.
• In a transformer, attention scores how much tokens influence each other and positional encoding supplies their order; diffusion models generate images from noise, and multi-modal models handle several data types.
• GenAI produces new content; classification, forecasting, and clustering are traditional ML tasks.
• The FM lifecycle runs data selection, model selection, pre-training, fine-tuning, evaluation, deployment, then feedback; evaluation measures but does not change the model, so a narrow quality gap is closed by fine-tuning.
• Token-based pricing bills input and output tokens at separate rates; a change saves tokens removed times that side's rate, unlike per-request or hourly Provisioned Throughput billing.
• Context engineering supplies the right information at the right time; a larger context window raises the limit but does not pick what is relevant.
• A supervisor agent routes variable requests to specialists at run time; a sequential pipeline suits fixed routes; hierarchical, peer-to-peer, sequential, and broadcast are the communication patterns.
• MCP is an open, vendor-neutral standard: MCP servers expose tools and data that any MCP client agent can use against the live system.

• GenAI's core advantages are adaptability, responsiveness, conversational capability, and the ability to generate original content.
• A hallucination is fabricated content stated confidently; a knowledge cutoff is once-true but outdated content, and model drift is performance that changes over time, unlike run-to-run nondeterminism.
• Grounding (RAG), automated checks such as Bedrock Guardrails contextual grounding checks, and human review reduce hallucinations; lower temperature, a larger model or fine-tuning do not eliminate them.
• Interpretability is limited, and a model's own explanation of its reasoning is generated text that may not reflect how the answer was produced.
• A model without live data or tools cannot know current figures and can get exact arithmetic wrong, so check such outputs against an authoritative source.
• Split workflows: keep exact, rule-based or must-be-justified decisions in deterministic or explainable systems, and use GenAI for drafting, summarizing and personalizing.
• Model selection balances capabilities (modality, context window), performance, cost, latency, size and compliance; compliance and data residency filter the shortlist before anything else.
• Business value is measured with ROI, efficiency, conversion rate, ARPU (per period), CLV (whole relationship), accuracy and cross-domain performance, not technical measures like throughput or perplexity.
• Good GenAI judgment weighs benefits against risks for each task rather than adopting or rejecting the technology outright.

• Amazon Bedrock is a fully managed, serverless service for accessing foundation models from multiple providers through one API; with the Converse API, switching models is a model-ID change.
• Amazon SageMaker AI builds, trains (including from scratch), fine-tunes and hosts models on instances you choose; SageMaker JumpStart starts you from pre-trained models deployed to your own SageMaker endpoints.
• Bedrock customization only adapts an existing base model; Custom Model Import serves weights trained elsewhere serverlessly; Bedrock Marketplace models run on endpoints you size and cannot be fine-tuned in Bedrock.
• Amazon Quick is the AI assistant and workspace for employees (successor to Amazon Q Business); Kiro is the agentic, spec-driven development environment for developers.
• Strands Agents is an open-source SDK to write agents; Amazon Bedrock AgentCore runs and operates agents from any framework with managed runtime, memory, identity, gateway and observability.
• Bedrock does not store, log or train on prompts and completions or share them with model providers; Guardrails add denied topics, content and word filters, and PII masking.
• On-demand is priced per input and output token; Provisioned Throughput reserves capacity billed hourly (no commitment, 1 or 6 months); custom models add training and monthly storage charges.
• Trade speed for cost with service tiers (Priority, Standard, Flex) and batch inference (about 50% cheaper, not real time); add capacity with geographic or global cross-Region inference profiles.

Applications of Foundation Models

• Select a foundation model by first checking hard limits (modality, language support, context window), then trading off cost, latency, and model size and complexity.
• The context window limits what one request can hold (prompt plus output); input that exceeds it is rejected, while the maximum output tokens setting truncates replies and bounds output cost.
• Prompt caching reuses an identical prompt prefix to cut input cost and latency on the same model; it does not change answers, does not affect output-token cost, and cached tokens still count toward the context window.
• Temperature, top-p, and top-k change word choice and do not set or cap length (only the output cap does); a low temperature makes answers repeatable, not correct, and grounding with RAG is what supplies missing facts.
• RAG retrieves relevant passages into the prompt at query time without retraining; it reduces (not eliminates) hallucination, adds input tokens, and keeps prompts small as a document set grows.
• Amazon Bedrock Knowledge Bases is the managed RAG capability and must be synced after source documents change; Guardrails' contextual grounding check validates answers but does not retrieve content.
• Embeddings can be stored in Amazon OpenSearch Service, Aurora PostgreSQL or RDS for PostgreSQL (pgvector), or Neptune Analytics.
• Customization runs cheapest to most expensive: prompt engineering / in-context learning, RAG, fine-tuning, continuous pre-training, pre-training from scratch; distillation trains a smaller student on a teacher's responses to cut inference cost.
• An AI agent is a foundation model plus tools (Bedrock Agents action groups) plus orchestration that plans steps at run time, unlike a predefined workflow such as AWS Step Functions.

• Prompt engineering shapes a model's output by improving the prompt, with no training, making it the cheapest way to change model behavior.
• The core prompt constructs are instruction (the task), context (background), input data, examples, and negative prompts; describing a situation only implies a rule, while a negative prompt excludes unwanted content explicitly.
• Techniques differ by example count: zero-shot uses no examples and the fewest tokens, single-shot uses one, and few-shot uses several to teach a pattern.
• Chain-of-thought prompting asks the model to reason step by step, fixing skipped steps in multi-step problems; prompt templates are reusable structures with placeholders.
• Temperature is an inference parameter, not prompt content: it changes how varied output is, but cannot fix reasoning, exclude an item, or supply unstated requirements.
• Best practices: be specific, state the task in plain concise wording (Bedrock guidance places it after the context), give relevant context rather than everything, use multiple short comments, and iterate.
• Prompt injection and hijacking redirect the task; jailbreaking bypasses safety rules; exposure reveals sensitive data; prompt leaking reveals the prompt; poisoning corrupts training data.
• Mitigate by risk: delimit untrusted input and use the Bedrock Guardrails prompt attack filter (jailbreaks, prompt injection, and on the Standard tier prompt leakage); keep secrets out of system prompts; keep restricted data out of the context, with the sensitive information filter as a second layer; clean poisoned training data and retrain.
• Prompt-level guardrails complement but do not replace dedicated controls such as Amazon Bedrock Guardrails, because injected or jailbreak input can override them.
• Amazon Bedrock Prompt Management uses variables, variants, a working draft, and immutable versions; switching back to an earlier version is the rollback.

• Pre-training builds broad knowledge from scratch on massive unlabeled text and is the most expensive step; you usually consume a provider's pre-trained model.
• Pre-training and continuous (continued) pre-training are self-supervised: the text supplies its own next-token labels, so they need no labeled data.
• Cost order from lowest to highest is fine-tuning, continuous pre-training, pre-training, because continuous pre-training processes far larger corpora than fine-tuning.
• Fine-tuning learns from labeled input-output pairs; continue pre-training on raw domain text first, then fine-tune, when you need both domain knowledge and task behavior.
• Distillation trains a small student on a large teacher's responses, so it needs prompts rather than labeled answers; Amazon Bedrock Model Distillation manages it.
• Good fine-tuning data is curated, governed, adequately sized, consistently labeled, and representative of live traffic; copying a minority group's few examples is not a fix.
• RLHF uses human rankings of the model's own responses to train a reward model; supervised fine-tuning uses human-written ideal responses.
• Only training updates weights; RAG, prompt engineering, and in-context (few-shot) examples act at request time and change nothing in the model.

• Human evaluation suits subjective quality but is slow and costly; benchmark datasets give fair, repeatable comparison but may not reflect your own data; Amazon Bedrock Model Evaluation runs evaluations as managed jobs.
• Bedrock automatic jobs use built-in metrics (accuracy, robustness, toxicity) on built-in or your own datasets for one model; human jobs need your own prompt dataset and can compare up to two models; judge-model and RAG jobs score with an LLM.
• Bedrock's built-in accuracy metric compares outputs with ground-truth reference responses; a retrieve-only RAG evaluation scores retrieval alone, while retrieve and generate also scores the answer.
• ROUGE is recall-oriented overlap with a reference summary; BLEU is precision-oriented n-gram overlap with reference translations; both, like exact match, penalize correct paraphrases.
• BERTScore compares meaning against the same references; LLM-as-a-judge grades against a rubric and needs no reference text, as do human evaluation and user ratings.
• Beyond metrics, judge whether the FM meets business objectives like productivity, user engagement and task engineering.
• Evaluate RAG retrieval relevance and answer grounding separately, and agents by tool choice and tool inputs as well as task completion.
• Cost per interaction is (input tokens x input price) + (output tokens x output price) per request, so a model with a lower per-token price but longer answers can cost more.

Guidelines for Responsible AI

• The features of responsible AI are bias, fairness, inclusivity, robustness, safety, and veracity — bias is the flaw, fairness the goal, and veracity (truthfulness) the feature generative AI most threatens.
• Amazon Bedrock Guardrails is the managed safety layer for generative AI: content filters (strength set separately for prompts and responses, including prompt-attack detection), denied topics, word filters (fixed lists), sensitive information filters (mask or block PII and custom regex patterns), and contextual grounding checks.
• Contextual grounding checks score two things, each with its own threshold: grounding (supported by the source) and relevance (answers the query) — the run-time hallucination control for RAG apps.
• Responsible model selection includes sustainability — bigger models use more energy, so right-size the model and reuse pretrained ones.
• Legal risks of generative AI include intellectual-property infringement claims, biased outputs, hallucinations, end-user risk, and loss of customer trust.
• Bias enters through data: deleting a protected attribute does not remove it, because proxy features carry it and the data stays unbalanced — rebalance and diversify the data instead.
• Bias underfits (model too simple) and variance overfits (model too tuned); their errors concentrate on underrepresented demographic groups, making them a fairness problem.
• High aggregate accuracy can hide poor results for small groups — subgroup analysis exposes it; conflicting labels on near-identical examples call for label-quality analysis.
• SageMaker Clarify detects bias before launch, SageMaker Model Monitor watches production (including bias drift using Clarify metrics, with CloudWatch alerts), and Amazon A2I routes low-confidence predictions to humans.

• Transparency is understanding how a model works and was built; explainability is understanding why it produced a specific output. A model card gives the first, not the second.
• Single decision trees and linear or logistic regression are interpretable; deep neural networks, LLMs and large tree ensembles are effectively black boxes, and an LLM's explanation of itself is generated text, not a trace.
• High-stakes, regulated decisions such as loans, hiring, insurance and medical use need per-decision explanations; low-stakes ones such as movie recommendations can tolerate a black box.
• SageMaker Model Cards document intended use, training, evaluation results and limitations (even for models trained elsewhere); Model Registry tracks versions and approval status.
• SageMaker Clarify explains predictions with SHAP attributions, per prediction and overall; Model Monitor alerts on feature attribution drift after deployment.
• Amazon Bedrock model evaluation compares foundation models with automatic jobs (metrics such as accuracy, robustness and toxicity, on built-in or your own prompts) or human-based jobs with your own reviewers.
• Open weights and data give inspectability, but the licence decides permitted use: Apache 2.0 allows commercial use, CC BY-NC does not.
• Releasing weights or exact filter rules aids scrutiny but also misuse; a detailed model card without the weights is a middle path.
• Measure performance against the requirement first: if only the black box meets it, add per-decision explanations; if the gap is negligible, prefer the interpretable model.
• Human-centered design: disclose AI use, give plain-language reasons, collect in-product feedback and offer human review.

Security, Compliance, and Governance for AI Solutions

• Under the AWS shared responsibility model, AWS secures the cloud infrastructure while you secure your data, access and configuration — that split does not change for AI workloads.
• Use IAM roles with least privilege so each AI component gets only the permissions it needs and receives temporary credentials instead of long-lived keys.
• Encrypt AI data both at rest (with AWS KMS) and in transit (with TLS); a customer managed KMS key adds control of the key policy and CloudTrail-audited key use.
• Amazon Macie finds sensitive data such as PII in S3, continuously with automated discovery or with targeted jobs; AWS PrivateLink interface endpoints keep traffic to Bedrock off the public internet, and an endpoint policy limits what can be called through them.
• Bedrock model invocation logging (off by default) captures prompt and response text; CloudTrail records who called which API but not the bodies.
• Knowledge base citations show where each answer came from; data lineage, data catalogs and SageMaker Model Cards document the data and the model.
• RAG grounds answers in approved sources; output validation such as the Guardrails contextual grounding check and confidence scoring catch what grounding misses — a "decline when unsure" prompt is not a control.
• AgentCore Identity gives agents their own identity and vaulted OAuth tokens or API keys (user-consented or machine-to-machine); Policy in AgentCore enforces Cedar rules at AgentCore Gateway, outside the model, so prompt manipulation cannot bypass them.

• Governance and compliance prove an AI system is run responsibly and meets the rules — think of the chain regulation to control to evidence, run continuously.
• AWS Config tracks resource configuration and compliance over time (and conformance packs deploy a framework's rule set in one step); AWS CloudTrail records who called which API and when, including Amazon Bedrock model invocations.
• AWS Audit Manager collects evidence against a framework for existing customers but has been closed to new accounts since 30 April 2026 — AWS points new users to AWS Config conformance packs. AWS Artifact provides AWS's own compliance reports and agreements such as the HIPAA BAA.
• Amazon Inspector continually scans EC2 instances, ECR images and Lambda functions for known vulnerabilities and unintended network exposure; AWS Trusted Advisor runs automated account-wide best-practice checks, including service quotas, unlike the questionnaire-driven Well-Architected Tool.
• Data-governance strategies cover lifecycles and retention (enforced with S3 Lifecycle rules), residency (geographic, not global, Bedrock inference profiles keep processing in a geography), and logging and monitoring.
• CloudTrail shows who invoked a model; Bedrock model invocation logging captures the prompt and response; CloudWatch metrics show aggregate usage, latency and errors.
• Governance processes include written policies, periodic plus event-triggered reviews, transparency standards, and team training requirements.
• The Generative AI Security Scoping Matrix has five scopes — consumer app, enterprise app, pre-trained models, fine-tuned models, self-trained models; RAG on an unmodified model stays Scope 3, and fine-tuning moves you to Scope 4.

Yes — the whole AIF-C01 deck is free and needs no account, as is every AIF-C01 revision lesson. Create a free account to save progress and drill practice questions; full-length mock exams are a one-time unlock per certification.