Chapter 11: Must-Know Concepts

AI becomes easier to understand — and far easier to use — when you know a few key concepts. These ideas explain how AI learns, how it generates outputs, and why it sometimes makes mistakes.

This chapter breaks down the essential building blocks of modern AI in simple, practical terms.

Training is the process of teaching an AI model. It involves exposing a neural network to massive amounts of data and adjusting billions of parameters so it learns patterns.

Inference is what happens when you use the model — generating text, interpreting an image, answering a question, or making a prediction.

Training is expensive and done rarely. Inference is fast and happens every time you interact with AI.

Parameters are the internal numerical values a model adjusts during training. They encode everything the AI has learned from its data.

Modern frontier models contain hundreds of billions to over a trillion parameters, while smaller on-device models may contain only a few billion or less.

More parameters generally mean more capacity — but not always better performance. Training quality, data, and architecture also matter.

A prompt is the input you give an AI — a question, an instruction, a block of text, an image, or a task description.

Prompts shape the output. Clear, specific prompts produce clearer, more useful results.

Prompting is becoming an essential skill for working with AI — similar to learning how to search effectively on the internet.

AI doesn’t read text one word at a time. Instead, it breaks language into tokens, which may represent whole words, word fragments, numbers, or symbols.

A longer prompt equals more tokens. A longer response also means more tokens. This affects speed, memory usage, and sometimes cost.

Modern AI models can process contexts of hundreds of thousands to over a million tokens, enabling them to analyze long documents, codebases, and transcripts.

A hallucination occurs when an AI produces incorrect, invented, or misleading information with confidence.

Hallucinations happen because the model is predicting likely sequences — not verifying facts. Without retrieval or external grounding, it may fill gaps with plausible but false content.

Human judgment and fact-checking are essential when accuracy matters.

AI systems learn from real-world data — which means they can inherit real-world biases.

AI bias can appear in:

• Recommendations
• Language patterns
• Image classifications
• Hiring or evaluation tools
• Search results

Bias is not always intentional — but it must be recognized and mitigated. Diverse training data and human oversight are crucial.

Alignment refers to ensuring AI systems behave in ways that match human values, safety standards, and ethical expectations.

Research in alignment focuses on preventing harmful outputs, ensuring truthfulness, reducing bias, and maintaining user control.

Better alignment leads to safer, more reliable AI systems.

Multimodal AI models can understand and generate multiple types of data — such as text, images, audio, video, and even sensor data.

For example, you can upload a picture, ask a question about it, and receive a text response. Or you can describe an image and ask an AI to generate it.

Multimodality brings AI closer to how humans process the world: through multiple senses working together.

RAG is a powerful technique where the AI model retrieves relevant information from trusted sources — such as documents, databases, or knowledge bases — and uses that information to generate more accurate, grounded answers.

RAG greatly reduces hallucinations because the model is no longer relying only on its training data — it is actively pulling in real facts.

Many enterprise systems and applications use RAG to connect AI to reliable, up-to-date information.

These concepts — training, inference, tokens, prompts, bias, alignment, multimodality, and RAG — provide the foundation for understanding modern AI.

The more you understand these ideas, the more effectively and safely you can use AI in real-world situations.