Glossary

Tokens are how language models read and price text. What they are, why they drive cost and context limits, and what that means for building with LLMs.

A context window is the model's working memory—finite and token-measured. What it is, why it constrains your app, and how RAG works around its limits.

Fine-tuning trains a model on your examples to change its behavior. What it is, when it helps, and why to try prompting and RAG first.

An AI hallucination is a confident, plausible, wrong answer. Why models do it, why it's dangerous, and how grounding and evaluation keep it in check.

RLHF is how raw language models learn to be helpful and aligned. What it is, how it works at a high level, and why it shapes the AI you use every day.

Temperature is the dial between predictable and creative AI output. What it does, and how to set it right for reliability or for ideation.

Neural networks are the engine behind modern AI. What they are in plain terms, how they learn, and why they power everything from vision to language.

NLP is how computers work with human language. What it is, what it does, and how large language models transformed what's possible with text.

Computer vision is how machines 'see.' What it is, what it can do, and why messy real-world images—not benchmarks—decide whether it works.

Reinforcement learning teaches AI through reward and consequence, not labeled examples. What it is, where it shines, and where it doesn't fit.

Diffusion models are the tech behind AI image generation. What they are, how they work at a high level, and where they create business value.

A knowledge graph maps how things relate. What it is, why it matters, and how pairing it with LLMs can make AI answers more accurate and explainable.

Function calling lets an LLM do more than talk—it can call tools, fetch data, and take actions. What it is, and why it's the bridge to real AI agents.

MCP is an open standard for connecting AI to your tools and data. What it is, why a standard matters, and how it simplifies building integrated AI.

Guardrails keep AI inside safe, useful boundaries. What they are, the kinds that matter, and why no production AI should run without them.

Transfer learning lets you build on what a model already knows. What it is, why it slashes data needs, and why nearly all modern AI relies on it.

Few-shot learning teaches an LLM by example—right in the prompt. What it is, when it beats zero-shot, and when to reach for fine-tuning instead.

Zero-shot learning is when an LLM does a task it was never explicitly trained for—just from instructions. What it is, why it works, and its limits.

Chain-of-thought prompting asks a model to 'think step by step'—and it measurably improves reasoning. What it is, why it works, and its trade-offs.

Prompt injection is the AI equivalent of an injection attack—hijacking a model through crafted input. What it is, why it's dangerous, and how to defend.

Distillation shrinks a big model into a small, fast one that keeps most of the smarts. What it is, why it cuts cost, and where it fits.

Quantization makes AI models lighter by using lower-precision numbers. What it is, why it speeds up and shrinks models, and what you trade for it.

Models don't stay accurate forever—the world shifts under them. What model drift is, why it's dangerous, and how monitoring keeps AI reliable.

Inference is what happens every time an AI answers—and it's the cost that never stops. What it is, why it drives economics, and how to optimize it.