What AI, ML and deep learning actually are

Artificial Intelligence (AI) is the broad field of building systems that can perform tasks normally needing human intelligence - reasoning, perception, language, decision-making.

Machine Learning (ML) sits inside AI: it is the practice of training algorithms on data so they learn patterns and improve at a task without being explicitly programmed with rules for every case.

Deep learning sits inside ML: it uses artificial neural networks with many layers (hence 'deep') to learn complex patterns, especially from unstructured data like images, audio and text.

Generative AI sits inside deep learning: models (often built on transformer architecture) that generate new content - text, images, code, audio - rather than just classifying or predicting a number.

The three main ML learning types

• Supervised learning: trains on labelled data (input plus known correct output), e.g. predicting house price from features. Used for classification and regression.
• Unsupervised learning: trains on unlabelled data to find structure, e.g. clustering customers into segments. No 'correct answer' given.
• Reinforcement learning: an agent learns by taking actions in an environment and receiving rewards or penalties, aiming to maximise cumulative reward over time, e.g. game-playing agents or robotics.

Key building blocks of neural networks

• Neurons (nodes) organised in layers: input layer, one or more hidden layers, output layer.
• Weights and biases: learned numeric values that determine how strongly inputs influence outputs.
• Activation functions (e.g. ReLU, sigmoid): introduce non-linearity so networks can learn complex patterns, not just straight lines.
• Training uses forward propagation (making a prediction) and backpropagation (adjusting weights based on the error, using gradient descent) across many epochs.

Common exam trip-ups

• Do not confuse AI (broadest), ML (subset), deep learning (subset of ML) and generative AI (subset of deep learning) - the exam loves testing this nesting.
• Labelled data is the defining feature of supervised learning, not unsupervised.
• Reinforcement learning is defined by reward signals and sequential decisions, not by having labelled training examples.
• A model 'training' on data is different from a model 'performing inference' (making predictions on new, unseen data) - know the split between the training phase and the inference phase.
• Overfitting means a model performs well on training data but poorly on new data - it has memorised rather than generalised.

What generative AI actually is

Generative AI is a subset of deep learning that creates new content - text, images, audio, video, or code - rather than just predicting a label or a number.

It learns patterns from huge training datasets and uses that learned distribution to produce novel, often unique, outputs each time you ask.

Most modern generative AI for text and code runs on transformer-based large language models (LLMs), which use a self-attention mechanism to weigh the relevance of each word against every other word in the input.

Foundation models (FMs)

A foundation model is a very large model pre-trained on broad, unlabelled data that can be adapted to many downstream tasks without retraining from scratch.

Key traits examiners test: FMs are multi-purpose (not built for one narrow job), they need adaptation (prompting, fine-tuning, or RAG) to shine at a specific task, and they get better with more data and compute (scale).

Amazon Bedrock is AWS's managed service for accessing FMs from multiple providers (Anthropic, Meta, Mistral, Cohere, Stability AI, Amazon Titan/Nova) through one API - no infrastructure to manage.

Core generative AI concepts

• Tokens: the chunks (words or word-pieces) a model reads and generates; pricing and context limits are measured in tokens, not words.
• Embeddings: numerical vector representations of meaning, used for search, clustering, and RAG retrieval.
• Prompt: the input instruction; prompt engineering is refining wording to get better outputs without changing the model.
• Context window: the maximum amount of text (in tokens) a model can consider at once - too long and older content gets truncated or ignored.
• Hallucination: when a model generates plausible-sounding but false or unsupported content - a known limitation, not a bug you can fully eliminate.
• Temperature: a setting controlling randomness/creativity of output - low temperature = more deterministic, high = more varied.

Common exam trap

Generative AI is NOT the same as traditional/discriminative ML. Discriminative models classify or predict from existing data (is this email spam?); generative models create brand-new content that did not exist before. Know which use case fits which.

Business use cases (exam favourites)

• Content creation: marketing copy, images, video scripts, summarisation.
• Chatbots and virtual assistants (e.g. Amazon Q).
• Code generation and review (e.g. Amazon Q Developer).
• Search and personalisation, using embeddings/RAG.
• Summarising long documents and translating text.

Limitations to remember

High compute cost, interpretability challenges (it's hard to explain exactly why an output was produced), potential for bias inherited from training data, and hallucination risk - always mentioned as trade-offs against generative AI's flexibility.

What is a foundation model?

A foundation model (FM) is a large, pre-trained model trained on huge, broad datasets that can be adapted to many downstream tasks (text, image, code, audio) rather than one narrow job.

Amazon Bedrock is the AWS managed service that gives serverless, API access to FMs from multiple providers (Amazon, Anthropic, Meta, Mistral, Cohere, Stability AI, AI21) without you managing infrastructure.

Key foundation model concepts

• Pre-training: the model learns general patterns from massive unlabelled data; this is expensive and done once by the provider.
• Fine-tuning: you adapt a pre-trained FM to your own labelled data for a specific task; cheaper and faster than training from scratch.
• Continued pre-training (domain adaptation): further training on unlabelled domain data to improve vocabulary/context understanding.
• Retrieval Augmented Generation (RAG): the model retrieves relevant external data at inference time to ground its answer, reducing hallucination without retraining.
• Amazon Bedrock Knowledge Bases is AWS's managed RAG feature, using vector embeddings and a vector store.

Inference parameters (exam favourites)

• Temperature (0 to 1): controls randomness/creativity. Low = deterministic, focused, repeatable. High = diverse, creative, riskier.
• Top P: controls diversity by only sampling from the smallest set of tokens whose cumulative probability meets P. Lower Top P = narrower, more predictable choices.
• Top K: limits sampling to the K most likely next tokens.
• Max tokens / response length: caps the length of the generated output and therefore cost.
• These parameters affect output quality and cost but do NOT change the model's underlying knowledge.

Prompt engineering essentials

Prompt engineering is crafting inputs to get better, more reliable outputs WITHOUT changing model weights - it's the cheapest and fastest way to improve results.

• Zero-shot prompting: no examples given, just the instruction.
• Few-shot prompting: a small number of examples included in the prompt to show the desired pattern.
• Chain-of-thought (CoT) prompting: asking the model to reason step by step, improving performance on multi-step or logic problems.
• Prompt templates: reusable structured prompts with placeholders, useful for consistency at scale.
• Negative prompting: telling the model what NOT to do or include, common in image generation.

Prompt risks and mitigation

• Prompt injection: malicious input trying to override the system prompt or instructions - a security concern for AI Practitioner exam.
• Jailbreaking: attempts to bypass safety guardrails.
• Hallucination: the model confidently generates false or unsupported information - mitigate with RAG, fine-tuning, or lower temperature.
• Amazon Bedrock Guardrails helps filter harmful content, block denied topics, and reduce hallucinations at the platform level.

Common mistakes

• Confusing fine-tuning (changes model weights) with prompt engineering (changes only the input, no retraining).
• Thinking temperature and Top P do the same thing - they are separate, complementary controls.
• Forgetting RAG does NOT retrain the model; it augments the prompt with retrieved context at runtime.
• Assuming more examples always help - few-shot only needs enough examples to demonstrate the pattern, not a full dataset.

Amazon Bedrock

Bedrock is AWS's fully managed service for building generative AI apps using foundation models (FMs) from multiple providers — Amazon (Titan, Nova), Anthropic (Claude), Meta (Llama), Mistral, Cohere, AI21, and Stability AI — all through one API.

• It is serverless: no infrastructure to provision or manage.
• Bedrock lets you customise FMs via fine-tuning and continued pre-training on your own labelled data, stored privately.
• Knowledge Bases add Retrieval Augmented Generation (RAG) by connecting FMs to your own data sources (S3, etc.) without retraining.
• Agents for Bedrock let a model orchestrate multi-step tasks and call external APIs/Lambda functions.
• Guardrails let you filter harmful content, block topics, redact PII, and reduce hallucinations — configurable independent of the model.
• Your data is never used to train the underlying FMs, and stays encrypted in your VPC boundary — a common exam point on data privacy.

Amazon SageMaker

SageMaker is the end-to-end platform for building, training, and deploying your OWN custom machine learning models (not just using pre-trained FMs).

• SageMaker Studio is the unified IDE for the whole ML lifecycle.
• SageMaker Canvas gives a no-code visual interface for building ML models — key for the 'no coding required' exam clue.
• SageMaker JumpStart provides pre-built, pre-trained models and solution templates you can deploy quickly.
• SageMaker Data Wrangler simplifies data prep and feature engineering.
• SageMaker Clarify detects bias in data/models and adds explainability (SHAP values).
• SageMaker Model Monitor watches deployed models for data drift and quality degradation in production.
• Ground Truth is for building high-quality labelled training datasets, including human-in-the-loop labelling.

Common mistake

Students mix up Bedrock (use/customise existing FMs) with SageMaker (build custom models from scratch). If the scenario says 'quickly build a generative AI app with minimal ML expertise', think Bedrock. If it says 'train a custom model on our own algorithm/data with full control', think SageMaker.

Pre-built AI services

AWS also offers task-specific AI services needing zero ML knowledge:

• Amazon Rekognition — image and video analysis (object/face detection).
• Amazon Textract — extracts text and data from scanned documents.
• Amazon Comprehend — natural language processing, sentiment analysis, entity extraction.
• Amazon Transcribe — speech to text.
• Amazon Polly — text to speech.
• Amazon Translate — language translation.
• Amazon Lex — builds conversational chatbots (same tech as Alexa).
• Amazon Kendra — intelligent enterprise search using natural language.
• Amazon Personalize — real-time personalised recommendations.

Know which service maps to which use case — the exam loves matching questions.

Why responsible AI matters for the exam

AWS AI Practitioner tests whether you understand the risks generative AI and ML introduce, and which AWS tools/features address each risk. Think in categories: bias, fairness, explainability, safety/toxicity, transparency, privacy, and governance.

Core concepts

• Bias: systematic skew in a model's outputs, usually because training data under-represents or misrepresents a group. Can enter at data collection, labelling, or model design stages.
• Fairness: outputs should not disadvantage protected groups (age, gender, ethnicity, disability, etc). There is no single fairness metric - AWS SageMaker Clarify offers several (e.g. class imbalance, difference in positive proportions).
• Explainability: the ability to say why a model produced a result. SHAP values (used inside Clarify) are the standard technique tested.
• Transparency: giving users and stakeholders visibility into how a system works, its limitations, and intended use - AWS publishes 'service cards' and model cards for this.
• Toxicity and safety: filtering harmful, hateful, or unsafe content from generative outputs.
• Hallucination: a generative model producing confident but false or fabricated information - a generative-AI-specific risk, not really a 'bias' issue.

Key AWS tools (exam favourites)

• Amazon SageMaker Clarify: detects bias in training data and model predictions (pre-training and post-training bias metrics) and provides explainability via SHAP.
• Amazon SageMaker Model Monitor: watches deployed models for data drift, quality drift, and bias drift over time.
• Amazon Augmented AI (A2I): inserts human review into ML predictions, useful for low-confidence or high-stakes decisions.
• Guardrails for Amazon Bedrock: configurable filters for denied topics, harmful content, PII redaction, word filters, and grounding/relevance checks to reduce hallucination.
• AWS AI Service Cards: AWS's own transparency documents describing intended use, limitations, and performance of AWS AI services.

AWS's Responsible AI pillars

AWS names eight core dimensions: fairness, explainability, privacy and security, safety, controllability, veracity and robustness, governance, and transparency. You do not need to recite all eight verbatim, but recognise them when listed as answer options.

Common mistakes

• Confusing bias (a data/model property) with fairness (an outcome/impact property) - bias causes unfairness, but they are not the same word for the exam's purposes.
• Thinking more training data automatically fixes bias - it only helps if the new data addresses the specific under-representation.
• Forgetting that human-in-the-loop (A2I) is a control, not an automatic fix - it needs correct escalation thresholds.
• Assuming Bedrock Guardrails stops all hallucination - it reduces risk (via grounding checks) but cannot guarantee zero hallucinations.
• Mixing up SageMaker Clarify (bias/explainability) with SageMaker Model Monitor (ongoing drift detection) - Clarify is more about the 'why', Monitor is about 'is it changing over time'.

Why this topic matters

AI Practitioner leans heavily on responsible AI, security and governance - expect several questions on shared responsibility, data protection, and AWS's governance tooling.

Shared Responsibility Model for AI

• AWS secures the infrastructure ('security OF the cloud') - hardware, host OS, physical facilities, the managed service itself.
• You secure 'IN the cloud' - your data, IAM permissions, model fine-tuning data, prompts, encryption choices, and how you configure guardrails.
• For Bedrock and SageMaker, AWS never uses your prompts or outputs to train its base models, and your data is not shared across customers - this is a favourite exam point.

Data protection basics

• Encryption at rest uses AWS KMS by default across Bedrock, SageMaker, and S3-backed datasets; you can bring your own customer-managed key (CMK).
• Encryption in transit uses TLS.
• VPC endpoints (via AWS PrivateLink) let you call Bedrock/SageMaker without traffic crossing the public internet.
• Amazon Macie discovers and classifies sensitive data (like PII) stored in S3 - useful for auditing training data before it touches a model.

Governance and compliance tooling

• AWS Artifact - self-service portal for AWS's own compliance reports (SOC, ISO, PCI) and to accept agreements.
• AWS Audit Manager - continuously collects evidence to assess risk and compliance against frameworks (e.g. GDPR, HIPAA).
• AWS Config - tracks resource configuration changes and checks them against rules (compliance drift).
• AWS CloudTrail - logs every API call for auditing 'who did what, when' - critical for AI governance trails.
• Amazon SageMaker Model Cards and Model Monitor document model intent, risk, performance and detect drift in production - core to AI governance.
• SageMaker Clarify detects bias in data/models and explains predictions (explainability), supporting fairness requirements.

Guardrails for Amazon Bedrock

• A managed feature to filter harmful content, block denied topics, redact PII, and reduce hallucinations by grounding checks - configurable per use case, reusable across models.

Common mistakes

• Confusing 'security of the cloud' (AWS) with 'security in the cloud' (you) - IAM misconfiguration is always your responsibility.
• Assuming Bedrock trains on your prompts by default - it does not.
• Mixing up Audit Manager (continuous compliance evidence) with Artifact (static compliance documents/reports).
• Forgetting that CloudTrail is about API activity logging, not resource configuration - that's Config's job.
• Thinking Clarify is only for bias - it also covers explainability of predictions.

Exam tip

When a question mentions 'who is responsible for X', default to the Shared Responsibility Model split. When it mentions bias/explainability, think Clarify; when it mentions drift/monitoring after deployment, think Model Monitor.