Every term, in plain English

An LLM that can plan, use tools, and take multi-step actions toward a goal.

Mass-produced, low-effort AI content that looks fine but adds little value.

Software that performs tasks we'd normally call intelligent - the broad umbrella over ML, deep learning, and GenAI.

Skew a model absorbs from human-written training data (stereotypes, skewed defaults).

Prompting a model to reason step by step, improving logic and math.

How much text a model can consider at once; long inputs can get truncated or lost.

ML using many-layered neural networks that learn features directly from raw data (images, audio, text).

A numeric vector capturing the meaning of text, so machines can compare similarity.

Humans hand-picking the useful columns/inputs a traditional ML model learns from.

Deep learning trained to create new content (text, images, audio, code), not just classify or predict.

When a model states something false as if true, because it generates plausible text without a fact-check.

The date a model's training data ends; it can't know newer events without tools/search.

A large neural network that generates text by predicting the next token, trained on huge amounts of text.

Algorithms that learn patterns from data instead of following hand-written rules.

Model Context Protocol - a standard way to connect models to tools and data sources.

How an LLM generates: it assigns probabilities to possible next tokens and samples one, repeatedly.

One of the model's billions of adjustable 'dials', set during training.

Personally Identifiable Information - names, emails, IDs, health, payment data.

Later training (fine-tuning + RLHF) that makes a base model helpful, honest, and safe.

The first training phase: predict the next token across internet-scale text (self-supervised) to learn language and knowledge.

The instruction/input you give an AI model.

Retrieval-Augmented Generation - fetch relevant documents and feed them to the model so it answers from real sources.

An agent pattern that interleaves reasoning and actions (tool calls).

Reinforcement Learning from Human Feedback - humans rank answers and the model learns what people prefer.

A dial for randomness: low = focused/predictable, high = creative/riskier.

The chunk of text a model reads/writes - roughly 4 characters or 0.75 of a word.

Splitting text into tokens before the model processes it.