AI has developed its own dialect — and everyone assumes you already speak it. One minute you’re hearing about LLMs, the next it’s RAG, agents, vector databases, and MCP. Most explanations fail you in one of three ways: they’re too academic, too vague, or written for people who already know what they’re looking up.
This guide does something different. It gives you the mental model behind each term — the intuition that actually sticks. No equations. No research-paper abstractions. Just clear thinking and real analogies you can use in actual conversations.
By the end, you’ll have a solid reference you can return to before interviews, while building, or whenever a blog post leaves you nodding along but not quite following.
1. LLM (Large Language Model)
An LLM is a neural network trained on a massive amount of text. It doesn’t think or reason the way humans do — it learns statistical patterns in language: which words tend to follow other words, at enormous scale. Every response it generates is essentially a very sophisticated prediction of what should come next.
Think of it as a well-read autocomplete. It has absorbed so much human writing that its predictions become eerily coherent — but there’s no understanding underneath, only pattern.
2. Tokenization
AI doesn’t read words the way you do. It reads tokens — chunks of text that might be a whole word, a fragment, or a punctuation mark. “Hello, this is Mark” becomes four tokens. Why does this matter? Because more tokens means more cost and slower responses. Efficient prompts are economical prompts.
3. Transformer
The transformer is the architectural breakthrough behind virtually every modern AI model. Its key innovation is the attention mechanism: instead of reading a sentence one word at a time, a transformer processes all words simultaneously, weighing how each one relates to every other. This is how AI grasps context over long passages rather than losing the thread halfway through.
Older models read like slow typists — one character at a time, context fading fast. Transformers read like a skilled editor who sees the whole page at once.
4. Embeddings
AI operates on numbers, not words. Embeddings are the translation layer: they convert text into lists of numbers (called vectors) that encode meaning. The crucial property is that similar meanings produce similar numbers. “King” and “Prince” end up numerically close to each other. “King” and “Pizza” end up far apart.
Think of it as plotting every word as a point in space. Related words cluster together. Embeddings are the coordinates of meaning.
5. Vector Database
A vector database stores embeddings and retrieves them by similarity — not by exact keyword match. Ask it a question and it finds documents with related meaning, even if they share no words with your query. This makes vector databases far more powerful than traditional databases for AI-powered search, and they’re the backbone of most modern AI applications.
6. RAG (Retrieval-Augmented Generation)
RAG gives an LLM access to external information before it answers. When you ask a question, the system first searches a vector database for relevant documents, then passes those documents to the model as context. The model answers based on retrieved facts rather than just its training data.
Without RAG, an AI answers from memory alone. With RAG, it looks things up first. The difference between a closed-book and an open-book exam.
7. Hallucination
Hallucination is what happens when an LLM produces a confident, fluent, completely false answer. It’s not lying — it has no concept of truth. It’s pattern-matching its way into plausible-sounding nonsense, which is most dangerous when the output looks authoritative. RAG reduces hallucination by grounding answers in real retrieved data. Never trust AI output blindly — especially on facts.
8. Prompt Engineering
The quality of AI output is directly proportional to the quality of the instructions it receives. Prompt engineering is the practice of crafting inputs that specify tone, format, role, and constraints precisely. Good prompts produce sharper, faster, cheaper results. Vague prompts produce vague answers.
Think of it like briefing a contractor. “Build me something nice” yields chaos. “Build a 12×10 deck with cedar planking, no nails visible” gets you what you wanted. Specificity is the skill.
9. Fine-Tuning
Fine-tuning takes a general-purpose model and trains it further on domain-specific data — your company’s documents, a particular writing style, a specialized vocabulary. The result is a model that behaves more reliably in your context. It’s more powerful than prompting, but more expensive and complex. Most small applications don’t need it; it’s what you reach for when prompts hit their ceiling.
10. Latency
Latency is the time between sending a prompt and receiving a response. It’s shaped by model size, token count, and infrastructure. Lower latency makes AI feel snappy and alive. High latency kills user experience — even when the answer is good. For real-time applications, latency is often more important than raw accuracy.
11. Agent
An agent is an LLM given the ability to take actions: calling APIs, running code, browsing the web, reading files, or triggering other models. Given a goal, it plans steps, executes them, observes results, and adapts. Agents are what transform AI from a sophisticated chatbot into an autonomous system that actually gets things done.
The difference: asking a chatbot “find me flights under $400 to Tokyo” gets you advice. An agent opens the booking site, filters results, and brings you back an answer.
12. Evaluation
Evaluation is how you measure whether your AI is actually working. Without it, you’re shipping blind. Good evaluation catches regressions, surfaces failure modes, and builds the feedback loop that makes systems improve over time. There are four main approaches:
Human evaluation — A person reads and scores outputs. Slow and expensive at scale, but the gold standard for nuanced quality. Still essential for high-stakes decisions.
BLEU / ROUGE — Automated metrics that measure word overlap between AI output and a reference answer. Useful for summarization and translation, but blind to meaning. A perfectly valid paraphrase can score poorly.
BERTScore — Uses embeddings to compare meaning rather than exact words. A significant improvement over BLEU/ROUGE, and better at catching semantic equivalence — though slower to compute.
LLM-as-a-Judge — One AI model evaluates the outputs of another. Scales easily and is surprisingly effective. The risk: the judge model carries its own biases and blind spots, so it needs careful calibration.
Leave a Reply