Day 1 - Domain Review 1: AI/ML Fundamentals + Generative AI Fundamentals: Critical Summary

📌 Key Takeaways

• AI ⊃ ML ⊃ Deep Learning ⊃ Generative AI. If a human writes the rules as code, it is a rule-based system, not ML.
• Learning methods split on whether answers (labels) exist: supervised (classification, regression), unsupervised (clustering, dimensionality reduction), reinforcement (reward).
• If the output is a category, it is classification; if it is a continuous number, it is regression. Pick an evaluation metric by asking "which kind of mistake is more costly?"
• Generative AI predicts the probability of the next token, one token at a time. Cost, limits, non-determinism, and hallucination all come from this.
• The order for trying to change model behavior is prompt engineering → RAG → fine-tuning, cheapest and fastest first.

AI · ML · Deep Learning · Generative AI: Re-Drawing the Inclusion Relationships

This is the last week. The goal of a review week is not "learning something new" but "re-drawing the boundaries that blur together." Today we combine Domain 1 (AI/ML Fundamentals, 20%), the starting point of the exam, with Domain 2 (Generative AI Fundamentals, 24%). Together the two domains make up 44% of the scored questions, and the questions in the remaining domains are built on the vocabulary used here.

The first trap to flag at the fundamentals level is the inclusion relationship among four terms. Small circles sit inside bigger circles.

• "Generative AI is a kind of ML" is true, and "all ML is deep learning" is false.
• If a human writes the rules directly as code, it is not ML but a rule-based system.
• The current exam guide adds agentic AI to the comparison. Where generative AI "produces the content you ask for," agentic AI is where a model receives a goal, plans the steps itself, and calls tools to finish the job. It is less a new circle in the inclusion diagram than a way of using a generative AI model as an "acting subject."

💡 Related Theory: What fundamentally separates ML from traditional programming is "do humans write the rules, or are the rules extracted from data?" Traditional programming is input + rules → output, but ML learns it backwards: input + output (answers) → rules (the model). That is why ML works well only when there is enough "data with answers attached," and if the data is biased, the model is biased too. This is why ML is the answer when the exam presents a situation like "spam rules keep changing and people cannot keep up by hand."

Three Learning Methods: Supervised · Unsupervised · Reinforcement

ML learning methods split on "are there answers (labels)?"

• Supervised learning splits again. If the output is a category (spam/normal, dog/cat), it is classification; if the output is a continuous number (price, temperature), it is regression. This distinction decides the answer on the exam.
• Not everything that outputs a number is regression. Choosing from fixed slots, such as "one of grades 1 to 5," is classification. Conversely, "tomorrow's visitor count" is a countable value, but because it predicts a continuous magnitude, it is regression.

Saying that unsupervised learning "learns without answers" means it finds structure within the data (distance, density, distribution) on its own. Clustering, for example, builds groups using only the notion of distance, "group the close ones together." That is why unsupervised learning is strong in exploratory situations where "you don't know in advance what to look for" (detecting abnormal transactions, discovering customer groups). Conversely, if you "already know what the right answer is," supervised learning is the fit.

Four Things Easy to Miss in Domain 1

These are items you skip if you stop at inclusion relationships and learning methods. All of them are in the exam guide's Domain 1 tasks.

① When not to use ML

② Traditional ML model or foundation model: which to use

• In a scenario like "decide whether to approve a loan and explain the reason for rejection," picking the newest large model by reflex is wrong. When you see regulation and explanation requirements, look at traditional ML first.

③ Four inference methods

④ Evaluation metrics

• Accuracy is the share of correct predictions out of all predictions. When classes are heavily imbalanced (1% fraud), answering "normal" for everything still gives 99%, so it cannot be trusted.
• Precision is the share of true cases among those flagged (it reduces false alarms), and recall is the share of true cases that were caught (it reduces misses). For a spam filter, precision comes first; for cancer diagnosis, recall comes first.
• The F1 score looks at precision and recall together.
• Separately from model metrics, the exam guide also lists business metrics (cost per user, development cost, customer feedback, return on investment) as evaluation criteria. Even with high accuracy, a model is not a good one if its cost exceeds its revenue.

Generative AI Fundamentals: Tokens · Embeddings · Foundation Models

• Generative AI handles text in tokens, not words. A token is a word fragment, and both model cost and context limits are counted in tokens. Input tokens + output tokens is the bill.
• An embedding converts text or images into a numeric vector that carries meaning. If the meanings are similar, the vectors are close. It is the basis of search, recommendation, and RAG.
• A foundation model (FM) is a large general-purpose model pretrained on massive data, and one model serves many tasks such as summarization, translation, and classification. The giant text-centered ones are LLMs (large language models).

The principle by which an LLM writes is "next-token probability prediction." Looking at the tokens so far, it picks the most plausible next token one at a time and appends it. Because of this probabilistic nature, the answer varies slightly even for the same question (non-determinism), and the model also makes up things it does not know in a plausible way, which is hallucination. Tokens, probability, and hallucination are concepts from the same root, so understand them together.

Here are the basic terms from the exam guide that are not in the table above, one line each.

Tokens Determine Cost and Performance

• Pricing is per input tokens + output tokens. A longer prompt raises input cost, and a longer answer raises output cost.
• A chatbot feeds the previous conversation back in as input on every turn. The longer the conversation, the more a single question costs, and once it exceeds the context window, the earliest part is cut off.
• The more tokens there are to process, the slower the response. Cost and latency usually move together.
• That is why choosing what goes into the context matters. The exam guide calls this context engineering. Of the instructions, retrieved documents, conversation history, and tool results, you include only what is needed right now, which raises quality and saves tokens.
• Ways to reduce: an output length cap (max tokens), retrieving and including only the necessary document pieces, and choosing a smaller model suited to the task.

Inference Parameters and Prompts in One Page

The key parameters that control generation results and the prompting techniques are also in scope for the exam.

• Temperature: higher is more diverse and creative, lower is more consistent and conservative. Lower it when accuracy matters.
• Top-p / Top-k: sampling controls that narrow the range of next-token candidates.
• Max tokens: the output length cap (cost control).
• Zero-shot: instructions only, no examples. Few-shot: show the pattern with a few examples (no weight change, In-Context Learning).
• The order for trying to change model behavior is prompt engineering → RAG → fine-tuning, cheapest and fastest first.

💡 Related Theory: When temperature is close to 0, the model picks only the "highest-probability token" every time, so it behaves almost deterministically. That is why a low temperature suits tasks with a single right answer such as classification and extraction, and a high temperature suits tasks that need diversity such as brainstorming and copywriting. Remember it as a dial between "creativity and consistency."

• Parameters cannot change what the model knows. No matter how you tune temperature, information the model does not have does not appear.

Strengths and Weaknesses of Generative AI, and Choosing a Model

• When choosing a model, look together at model type, performance requirements, capabilities, constraints, regulatory compliance, cost, latency, and model complexity. It is rare that "the biggest model" is the answer. The smallest, cheapest model that meets the requirements is usually the right one.
• Business value is measured with metrics such as efficiency, conversion rate, average revenue per user, accuracy, customer lifetime value, and return on investment.

Basic Terms of Agentic AI

The current exam guide puts the basic concepts of agentic AI in Domain 2. Because this is an entry-level exam, being able to say each term in one line is enough.

• What separates a chatbot from an agent is action. "Tell me the refund policy" is answering; "look up the order and file the refund" is an agent's job.

Six Common Misconceptions

Final Checklist for Domains 1 and 2

If you can complete the sentence aloud, that line passes.

[Domain 1 — AI/ML Fundamentals]
□ What is the inclusion relationship of the four terms? (→ AI ⊃ ML ⊃ Deep Learning ⊃ Generative AI)
□ What separates classification from regression? (→ Is the output a category or a continuous number)
□ Which task finds groups without labels? (→ Clustering in unsupervised learning)
□ Two situations where ML is not suitable? (→ The rules are clear; a single fixed outcome is required)
□ Which metric takes priority when misses are unacceptable? (→ Recall)
□ Which inference method predicts in bulk overnight? (→ Batch inference)
□ Which model to consider first for a regulated task that requires explanation? (→ A traditional ML model)
 
[Domain 2 — Generative AI Fundamentals]
□ What is the unit of cost and limits? (→ Tokens, input + output)
□ Which numeric representation gets closer when meanings are similar? (→ Embedding)
□ Why does hallucination occur at root? (→ It predicts next-token probability rather than verifying facts)
□ What should temperature be when consistent answers are needed? (→ Lower it)
□ What does not change in few-shot? (→ The model weights)
□ In what order do you try to change model behavior? (→ Prompt → RAG → Fine-tuning)
□ What separates an answer from an agent? (→ Whether it actually acts by using tools)

In the next post, we cover AWS's AI services (foundation model applications) in one sweep on top of this foundation. The key is practicing quickly recalling "which service for which situation."

📖 Glossary

• Rule-based system: A system in which a person writes conditions and rules directly as code to make it work. It does not learn from data.
• Label: The correct answer attached to training data.
• Classification / Regression: A task that predicts a category / a task that predicts a continuous number.
• Clustering: An unsupervised learning task that groups similar things together without answers.
• Token: A word fragment, the unit in which a model handles text. It is the basis of cost and limits.
• Embedding: A numeric vector that converts text or images so that it carries meaning.
• Foundation model (FM): A large model pretrained on massive data and reused for many tasks.
• Hallucination: The phenomenon of plausibly producing content that is not true.
• In-Context Learning: Getting the model to follow a pattern using only examples inside the prompt, without changing the weights.
• Agentic AI: AI in which a model plans on its own toward a goal and acts by using tools.