AI Glossary
Simple and accurate meanings of AI terms in Persian: LLM, token, context window, RAG, fine-tuning, hallucination, intelligent agent, MCP, and dozens of other words.
Alignment AI Alignment
The field that tries to align the behavior of AI systems with human goals and values; meaning the model is helpful, honest, and harmless.
API API
A way for applications to communicate directly with a service (e.g., an AI model). Developers use APIs to employ models in their own products.
Artificial Intelligence Artificial Intelligence (AI)
A branch of computer science aimed at building systems that perform tasks requiring human intelligence, such as language understanding, image recognition, decision-making, and problem-solving.
Attention mechanism Attention Mechanism
The part of the transformer that determines how much "attention" the model should pay to other words when processing each word; for example, in the sentence "Ali gave the book to Mary because she asked for it," identifying who "she" refers to.
Benchmark Benchmark
A standard test to measure and compare model capabilities in tasks like math, programming, or reading comprehension. Benchmark results are useful but do not always reflect performance in your actual work.
Chain of thought Chain of Thought (CoT)
A method where the model solves a problem step-by-step before the final response. This increases accuracy in math and logical problems and is the basis of reasoning models.
Context window Context Window
The maximum amount of text (in tokens) that the model can consider at once; including your messages, previous responses, and attached files. The larger it is, the longer documents the model can read at once.
Deep learning Deep Learning
A subset of machine learning that uses neural networks with many layers. Almost all recent advancements in artificial intelligence, from ChatGPT to facial recognition, are built on deep learning.
Expert combination Mixture of Experts (MoE)
An architecture that divides the model into several "experts", and for each input, only a few relevant experts are activated. The result is a large model that is faster and cheaper to run.
Few-shot learning Few-shot Prompting
A method where you put a few examples of desired input and output inside the prompt so the model understands the pattern. Giving no examples is called Zero-shot.
Fine-tuning Fine-tuning
Retraining a pre-trained model on your specialized data so it performs better on a specific task or with a specific style; for example, responding in your brand's tone.
Freemium; Free with paid features Freemium
A service delivery model where basic features are free, and a fee is charged for certain capabilities or higher usage limits. A free basic version does not mean all features or APIs are free. This differs from a trial period that expires after some time.
General Artificial Intelligence Artificial General Intelligence (AGI)
A hypothetical AI that can perform almost any intellectual task at or above human level. The precise definition and timeline for reaching it are subjects of serious debate among experts.
Graphics Processing Unit GPU
A processor designed to perform a large number of simple calculations simultaneously, making it ideal for training and running artificial intelligence models. Nvidia is the largest manufacturer in this field.
Hallucination Hallucination
When a model states incorrect or fabricated information with complete confidence; for example, a source that does not exist or a number that is not correct. For this reason, important claims should always be verified.
Inference Inference
The stage of using a trained model to generate responses. Every time you ask ChatGPT a question, an inference operation is performed.
Intelligent agent AI Agent
A language model-based system that plans by itself to achieve a goal, uses tools (search, code execution, email sending, etc.), and executes multiple steps sequentially.
Knowledge distillation Knowledge Distillation
Training a small model (student) using the outputs of a large model (teacher), so the small model achieves a large part of the large model's capability at a lower cost.
Large Language Model Large Language Model (LLM)
A model trained on a massive volume of text to predict the next word (more precisely: token). This simple capability at scale enables conversation, writing, translation, and programming. GPT, Claude, and Gemini are examples of LLMs.
Machine learning Machine Learning
A method in artificial intelligence where instead of writing manual rules, the system learns patterns from data. For example, by seeing thousands of emails, it learns to detect spam.
Model Context Protocol Model Context Protocol (MCP)
An open standard introduced by Anthropic that is now supported by many companies. It allows AI assistants to connect to tools and data sources (like Google Drive, GitHub, or databases) in a uniform way.
Multimodal Multimodal
A model that, in addition to text, understands or generates other types of data such as images, audio, and video.
Neural network Neural Network
A mathematical model inspired by the brain, made of a large number of connected artificial "neurons". Each connection has a weight adjusted during the training process.
Open weights Open Weights
A model whose parameter files have been published and can be downloaded and run on personal hardware; like Llama, Qwen, and DeepSeek. It is completely different from "open source", because the training data and code are usually not released.
Open-weight model Diffusion Model
A type of image generation model that learns to remove noise step by step from a noisy image to generate the desired image. Stable Diffusion and many modern image generators are based on this.
Parameter Parameters
The numbers (weights) that a model learns during training. Model size is usually measured by the number of parameters (e.g., 7 billion or 70 billion). More parameters do not necessarily mean a better model.
Pre-training Pre-training
The first and costly phase of building a language model where it is trained on a massive volume of internet text and books to learn general language and knowledge.
Prompt Prompt
The instruction or text you give to the AI model to generate a response. Prompt quality has a direct impact on output quality.
Prompt engineering Prompt Engineering
The skill of writing clear and structured prompts to get the best results from the model; includes role definition, context, output format, and examples.
Prompt injection Prompt Injection
An attack in which malicious instructions are hidden inside text, a webpage, or a document so that the model reading it does something contrary to the user's wishes. This is one of the most important security challenges of intelligent agents.
Quantization Quantization
Reducing the numerical precision of model parameters (e.g., from 16-bit to 4-bit) so the model uses less memory and runs on weaker hardware; with a slight drop in quality.
Reasoning model Reasoning Model
A model that spends time "thinking" before answering and examines the problem in multiple steps. More accurate for math, programming, and complex analysis, but slower and more expensive.
Reinforcement learning from human feedback RLHF
A method where humans compare and rank model responses, and the model learns to provide responses that are more useful, accurate, and safe.
Retrieval-Augmented Generation Retrieval-Augmented Generation (RAG)
A method where relevant information from a source (company documents, web, or database) is searched and given to the model before answering. This makes responses more accurate, up-to-date, and documented.
Semantic Vector Embedding
A numerical representation (vector of numbers) of a text or image that shows its meaning. Texts with similar meanings have vectors close to each other; this feature is the basis of semantic search and RAG.
Synthetic data Synthetic Data
Data that is generated with the help of AI models themselves rather than being collected from the real world, and is used to train or test other models.
System prompt System Prompt
The general instruction given to the model before a conversation, which sets its behavior, tone, and limitations throughout the conversation; usually written by the application creator.
Temperature Temperature
A setting that controls the randomness of the response. Low temperature yields more predictable responses, and high temperature yields more creative and diverse responses.
Token Token
The basic unit of text for a language model; it can be a word, part of a word, or a sign. API usage costs and input length limits are calculated based on the number of tokens. Persian text usually consumes more tokens than its English equivalent.
Tool calling Tool Use / Function Calling
A feature that allows the model to request the execution of a function or tool (such as getting currency rates or searching a database) instead of a text response, and use the result in its answer.
Transformer Transformer
The neural network architecture introduced by Google in 2017, which forms the basis of almost all modern language models. Its main innovation is the "attention" mechanism, which allows the model to gauge the relationship of each word with the rest of the words in the text.
Vector database Vector Database
A database designed for fast storage and retrieval of embeddings, finding texts that are "synonymous", not just texts with identical words.