AI DICTIONARY

How often a model gets the right answer compared to the truth. In classification, it's simply "correct predictions ÷ total predictions." It can be misleading when one class shows up far more than others, so it's usually paired with precision and recall.

A function inside a neural network that decides whether a neuron "fires." It adds non-linearity, which is what lets deep models learn complex patterns. Without activation functions, a neural network would act like a basic linear model.

A training method where the model flags the next data points that would be most valuable to label. That way, humans spend time labeling the highest-impact examples instead of everything. It's especially useful when labels are expensive or slow to get.

An ensemble technique that turns a bunch of "okay" models into one stronger model. It keeps paying extra attention to the examples earlier models got wrong, improving step by step.

An input that's been intentionally tweaked—sometimes in a way humans barely notice—to make a model fail. In images, tiny pixel changes can flip a label. In text, carefully engineered prompts can push a model into wrong or unsafe outputs.

A training setup where models compete in a way that improves results. The classic example is GANs: one model generates outputs, another tries to spot what's fake. That pressure makes the generator produce more realistic results over time.

An AI system that can plan and take actions to reach a goal. Instead of replying once like a chatbot, an agent can use tools, check its work, and keep going until the task is done. It's the shift from "AI answers" to "AI execution."

The layer that coordinates multiple agents, tools, and systems. It decides who does what, in what order, what happens when something fails, and what permissions apply. Think of it as air traffic control for multi-step AI work.

AI designed to operate more like an ongoing worker: it plans, executes, monitors outcomes, and adjusts as things change. In business terms, this is the move from a content helper to an ops helper.

A set of steps or rules for solving a problem. In AI, algorithms shape how models learn during training and how they produce outputs during inference. A model is basically algorithms + data + compute, packaged into something usable.

Making sure an AI system behaves the way people actually want it to—safely and reliably. That includes guardrails, refusal rules, and "don't do harm" constraints. It also includes practical alignment: getting the model to consistently follow your business goals and policies.

Labeling data so models can learn from it—tagging images, marking intent in emails, labeling outcomes, and so on. Great annotations usually mean better models. Messy labels often lead to confident, wrong behavior.