AI that creates new content — text, images, audio, or code — rather than just classifying it.
Generative AI is artificial intelligence that produces new content — text, images, audio, video, or code — by learning statistical patterns from massive datasets. Unlike discriminative models that label existing inputs, generative models synthesize novel outputs. On AWS, Amazon Bedrock is the managed gateway for generative AI, letting builders invoke foundation models from providers like Anthropic, Meta, and Amazon without managing infrastructure. The key exam distinction: a sentiment classifier is traditional ML, while a chatbot drafting a reply is generative AI. Don’t conflate Amazon SageMaker (training and deploying custom ML models) with Amazon Bedrock (accessing pre-built foundation models).
The basic unit of text a language model processes — often a word piece or subword.
A token is the smallest chunk of text a language model reads and generates — a subword unit, not a full word, so “unbelievable” may split into three tokens while “cat” is one. Inference is billed in tokens, so Amazon Bedrock pricing is quoted per thousand input and output tokens and context windows are sized in tokens, not words. The key exam distinction is input tokens (the whole prompt, including instructions and history) versus output tokens (the response), often priced differently. As a rough rule, one token is about four characters, roughly three-quarters of a word.
A numeric vector representation of data that captures semantic meaning.
An embedding is a fixed-length numeric vector that a model learns to represent text, images, or other data so that semantically related items cluster close together in vector space. On AWS, Amazon Bedrock provides embedding models (such as Amazon Titan Embeddings) that convert input into these vectors, which are then stored in a vector database to power semantic search and retrieval-augmented generation (RAG) pipelines.
The key exam distinction is between an embedding model and a generative model: embedding models produce vectors for comparison and retrieval, not human-readable text. Confusing the two roles is a common trap — a generative model answers questions, while the embedding model finds the relevant context to feed it.
An array of numbers representing data in a multi-dimensional space.
A vector is an ordered list of floating-point numbers that represents an object — a word, sentence, image, or document — as a point in high-dimensional space. Models produce these numeric representations so semantically similar items land close together, enabling search without exact keyword matching. The exam distinction: an embedding is the model output that converts raw data into a vector, while the vector is simply the resulting array. Amazon Bedrock Knowledge Bases store vectors to power retrieval-augmented generation (RAG), comparing a query vector against document vectors using metrics like cosine similarity.
The neural network architecture, based on attention, behind modern large language models.
The transformer architecture introduced self-attention, which lets the model weigh the relevance of every token in an input sequence against every other token simultaneously rather than one by one. This parallel processing made it practical to train on massive datasets and gave rise to large language models like those available through Amazon Bedrock. The encoder portion learns contextual representations useful for tasks such as classification, while the decoder generates new tokens one at a time, and many LLMs use a decoder-only design. A key exam nuance: transformers replaced recurrent networks (RNNs and LSTMs), which processed tokens sequentially and struggled with long-range dependencies.
A large model pre-trained on broad data that can be adapted to many downstream tasks.
A foundation model (FM) is a large neural network pre-trained on massive, diverse data — text, images, or both — giving it broad knowledge applicable to many tasks through prompting or fine-tuning, without training from scratch. Its scale (billions of parameters) enables emergent capabilities like reasoning, translation, and code generation. On AIF-C01, distinguish using an FM from customizing one: accessing an FM via Amazon Bedrock needs no ML expertise, while fine-tuning it on domain data is a separate, more involved Bedrock step. Remember that pre-training is done by the model provider, not the customer.
A foundation model trained on huge text corpora to understand and generate language.
Large language models are deep neural networks built on the transformer architecture, trained on vast text corpora to learn statistical patterns across words and documents. This lets them handle many language tasks — summarization, translation, question answering, and code generation — without task-specific retraining. On AWS, Amazon Bedrock provides managed access to leading LLMs such as Anthropic Claude, Meta Llama, and Amazon Nova. The key exam distinction: all LLMs are foundation models, but not all foundation models are LLMs — image generators like Stability AI’s Stable Diffusion are also on Bedrock yet are not LLMs.
A model that handles more than one data type, such as text plus images.
A multimodal model processes and generates content across more than one data type, such as accepting both text and images as input, then producing a text response that reasons over both. On the AIF-C01 exam, this capability is associated with Amazon Bedrock, which hosts foundation models like Anthropic Claude and Amazon Titan that accept image-plus-text input, enabling visual question answering, chart interpretation, and document understanding. Note the distinction between multimodal input (a model that reads images) and multimodal output (a model that also generates images); not every model does both, so read whether the scenario needs understanding or generation before answering.
A generative model that creates images by iteratively removing noise from random input.
A diffusion model trains in two phases: a forward pass that gradually adds Gaussian noise to a training image until only noise remains, and a reverse pass where the model learns to denoise step-by-step. At inference only the reverse process runs, conditioned on a text prompt, enabling text-to-image generation. On the AIF-C01 exam, Amazon Bedrock hosts diffusion models such as Stability AI’s Stable Diffusion and Amazon Nova Canvas. Unlike GANs, which use a generator-discriminator adversarial loop, diffusion relies on iterative denoising, generally giving higher diversity and fewer mode-collapse artifacts.
When a model produces plausible-sounding but false or fabricated output.
Hallucination occurs when a large language model generates output that sounds confident and coherent but is factually incorrect, invented, or unsupported by its training data. The model has no internal truth-checker — it predicts plausible token sequences, which can produce fabricated citations, wrong dates, or nonexistent service features. On the AIF-C01 exam, hallucination is the primary motivation for Retrieval-Augmented Generation (RAG), which grounds responses in verified external documents before generation. Amazon Bedrock supports RAG natively through its Knowledge Bases feature, letting applications anchor model output to real source content and reduce unsupported claims.