AI Terminology

A core subset of AI where computational models are trained on data to detect complex patterns, make predictions, and adapt without being explicitly programmed for individual edge cases.

A mathematical framework inspired by biological neural structures. It consists of layers of nodes (neurons) that process inputs and dynamically adjust connection weights during training to minimize errors.

A training paradigm where a model is fed paired inputs and outputs (labeled dataset). The model learns to map inputs to targets by comparing its predictions to the ground-truth answers.

An approach where models process raw, unlabeled data to discover organic groupings, clusters, anomalies, or underlying distributions on their own.

An optimization method where a software agent learns to execute actions in an environment to maximize cumulative rewards, utilizing trial-and-error reinforced loops.

The state-of-the-art neural architecture (introduced in 2017) that underpins modern large language models. It replaces recurrence with 'self-attention' mechanisms, allowing models to weigh the importance of all words in a sentence simultaneously.

The process of parsing text blocks into smaller numeric identifiers (tokens) that an LLM can ingest. A token can represent a whole word, character, or sub-word (approximately 4 characters of English text).

The exact memory boundary (measured in tokens) that a model can read, process, and refer back to during a single API call or chat interaction. Modern models feature context sizes scaling into millions of tokens.

High-dimensional vector outputs that represent the semantic meaning of text or media. By converting words or sentences into mathematical coordinates, models can instantly compute conceptual similarities.

A class of AI systems designed to produce fresh, original outputs (such as text, software code, vector assets, and synthesized media) based on distributions learned from historical training corpuses.

A phenomenon where an LLM generates factually incorrect, logically inconsistent, or fabricated claims with absolute confidence, often due to statistical patterns mismatching training sources.

A foundational context block injected before user input to define the model's persona, operating rules, constraints, API access guidelines, and target safety guardrails.

A scaling parameter that adjusts the probability distribution of predicted next-tokens. Setting it low forces deterministic, factual results; setting it high increases output diversity, creativity, and entropy.

The process of adapting a pre-trained base model to target domains or specialized formats by performing additional training epochs on a highly curated, smaller dataset.

An autonomous loop system that uses an LLM as its central brain. The agent acts, evaluates results, accesses tools (like calculators, search engines, or database connections), and iterates until it achieves its objective.

An architecture that queries external vector databases to retrieve relevant context snippets based on a user's prompt. This content is then injected into the prompt, grounding the LLM in real-time factual data.

A database optimized for storing, indexing, and calculating distance metrics between high-dimensional vector embeddings, enabling semantic searches in milliseconds.

A native model capability that detects when a prompt requires external logic (e.g., executing a web search) and outputs a clean payload (like JSON) specifying the function name and arguments to run.