AIAI Agent Dev Labaiagentdev-lab.com
Sign in to save progressSign in
Menu

AIF-C01 · LESSON 2

Fundamentals of Generative AI24% · 3 topics

Learn the basic concepts of generative AI, its possibilities and limits in business, and AWS infrastructure and technologies for generative AI.

2.1Basic concepts of generative AI

2.1 Explain the basic concepts of generative AI (GenAI).

Concept

Generative AI creates new text, images, audio, code and more. A foundation model (FM) is a large model pre-trained on large amounts of data, and LLMs (large language models) are a representative type. Text is split into tokens, and embeddings (vectors) represent meaning numerically. There are Transformer architectures, diffusion models (image generation), multimodal models and more.

How it works / steps

  1. Understand the meaning of tokens, embeddings and the context window
  2. Distinguish model types (text, image, multimodal, embedding)
  3. Understand the foundation model lifecycle (data selection → pre-training → adaptation → evaluation → deployment → feedback)
  4. Choose a model that fits the use case

Workplace example

For searching internal documents, convert documents into vectors with an embedding model to find semantically similar ones, then have an LLM summarize them to answer.

Common confusions and exam tips

Mini quiz

Q1. Which is the most appropriate characteristic of the Transformer (the mechanism underlying LLMs)?

  1. It is a mechanism only for images and cannot process text
  2. It looks only at the previous character, one at a time, and picks the next character from a dictionary in order
  3. Through its attention mechanism, it can process text while considering relationships even between distant words in a sentence
  4. People write in all grammar rules, and it produces sentences exactly according to those rules
Show answer and explanation

Answer: C Through its attention mechanism, it can process text while considering relationships even between distant words in a sentence

Through self-attention, the Transformer captures which words each word is most related to across the whole input. It is easy to process in parallel and suits large-scale training.

Q2. Which is the most appropriate description of a foundation model (FM)?

  1. A general-purpose model pre-trained on large amounts of data that can be used for many purposes such as summarization, question answering and classification
  2. A foundational database service for storing data used to train AI
  3. A dedicated model built from scratch with little data for just one task at one company, unusable for other purposes
  4. A rules engine that just collects business rules written by people and decides according to them
Show answer and explanation

Answer: A A general-purpose model pre-trained on large amounts of data that can be used for many purposes such as summarization, question answering and classification

An FM is a general-purpose model pre-trained on large-scale data and can be used for many purposes through prompt engineering, RAG and fine-tuning. LLMs are a type of FM.

Reference links

2.2Possibilities and limits of generative AI

2.2 Understand the capabilities and limitations of GenAI for solving business problems.

Concept

The advantages of generative AI are adaptability (diverse tasks), fast responses, natural conversation and task automation. Its limits include hallucinations (plausible errors), nondeterminism (different output each time), difficulty of interpretation, knowledge cutoff, bias and cost. Business value is measured with metrics such as efficiency, conversion rate and customer satisfaction.

How it works / steps

  1. Check whether it is work generative AI is good at (summarizing, drafting, classifying, conversing)
  2. For processing where errors are not allowed, combine human review and grounding (RAG)
  3. Choose models based on accuracy, cost, speed and regulatory requirements
  4. Measure impact with business metrics (ROI, processing time)

Workplace example

Leave drafting sales emails to generative AI with the rep checking them, but calculate contract amounts in conventional systems.

Common confusions and exam tips

Mini quiz

Q1. You introduced a generative AI assistant for inquiry handling. Which is the most appropriate metric for measuring its business impact?

  1. Changes in the number of model parameters and the size of the model file
  2. Monthly changes in the average character count of answers and the share of kanji used in them
  3. The share of inquiries resolved without handing off to a person, and the average handling time per inquiry
  4. The number of employees who attended the rollout briefing and the number of handouts distributed
Show answer and explanation

Answer: C The share of inquiries resolved without handing off to a person, and the average handling time per inquiry

Measure impact with business metrics such as resolution rate, handling time, customer satisfaction and cost. Technical figures alone cannot show impact.

Q2. Which is the most appropriate advantage of generative AI?

  1. Once introduced, there will never again be any cost for human review or operations
  2. It always returns factually correct answers in the same wording to any question
  3. One model can be used for many tasks such as translation, summarization and drafting just by changing instructions
  4. It always automatically knows the latest internal information even if it was not in the training data
Show answer and explanation

Answer: C One model can be used for many tasks such as translation, summarization and drafting just by changing instructions

Generative AI is highly adaptable and handles diverse tasks through prompts. On the other hand, it has limits on accuracy and up-to-date information.

Reference links

2.3AWS infrastructure and technologies for generative AI

2.3 Describe AWS infrastructure and technologies for building GenAI applications.

Concept

Amazon Bedrock is a fully managed service for using foundation models from multiple companies through an API (including Knowledge Bases, Agents, Guardrails, evaluation and customization). Amazon Q Business is an AI assistant grounded in company data, Amazon Q Developer supports development, SageMaker AI JumpStart deploys pre-trained models, and PartyRock is a playground for trying Bedrock. Pricing includes per-token on-demand, batch and Provisioned Throughput.

How it works / steps

  1. If you want a ready-to-use business assistant, use Amazon Q Business
  2. To embed generative AI in your own app, use Amazon Bedrock
  3. To host and tune models yourself, use SageMaker AI JumpStart
  4. Choose the pricing model (on-demand, batch, provisioned) by usage

Workplace example

Adopt Amazon Q Business to provide Q&A based on internal policies and procedure manuals without development, and build your own app's summarization feature with the Bedrock API.

Common confusions and exam tips

Mini quiz

Q1. In an internal chat, every prompt includes the entire 50-page policy handbook along with the question. Cost and response time are too high. Which is the most appropriate improvement?

  1. Expand the handbook to 100 pages so it can answer any question
  2. Use RAG to retrieve only the parts relevant to the question and add them to the prompt, reducing input tokens
  3. Insert the handbook three times into each prompt to emphasize its content
  4. Set maximum output tokens to the maximum and have answers written as long and detailed as possible
Show answer and explanation

Answer: B Use RAG to retrieve only the parts relevant to the question and add them to the prompt, reducing input tokens

Pricing and response time grow in proportion to input and output tokens. Passing only relevant parts with RAG lowers cost while keeping answers grounded.

Q2. Which is the most appropriate description of how prompts entered into Amazon Bedrock and the model's outputs are handled?

  1. They are shared with all providers and used to train the next version of the models
  2. They are made public, and other AWS customers can search and see the content
  3. AWS staff read every prompt entered to check quality
  4. They are not used to improve foundation models and are not shared with model providers
Show answer and explanation

Answer: D They are not used to improve foundation models and are not shared with model providers

In Bedrock, customer content is not used to improve foundation models and is not shared with model providers.

Reference links

Lessons are written by this site based on public information and do not guarantee exam content. Product specifications change, so check the latest information at the reference links.

Notes (unofficial, original questions, no pass guarantee)