AI Glossary — 7min.ai

Key terms and concepts in artificial intelligence, machine learning, and the AI industry. A comprehensive AI glossary by 7min.ai.

Artificial Intelligence (AI): The broad field of computer science focused on creating systems that can perform tasks typically requiring human intelligence, such as understanding language, recognizing patterns, and making decisions.
Machine Learning (ML): A subset of AI where systems learn patterns from data rather than being explicitly programmed. Includes supervised, unsupervised, and reinforcement learning approaches.
Deep Learning: A subset of machine learning using neural networks with many layers (hence 'deep') to learn representations of data. Powers most modern AI breakthroughs in language, vision, and generation.
Neural Network: A computing system inspired by biological neural networks, consisting of interconnected nodes (neurons) organized in layers that process information and learn patterns from data.
Generative AI: AI systems that can create new content — text, images, audio, video, or code — based on patterns learned from training data. Includes large language models and diffusion models.
Large Language Model (LLM): A neural network trained on massive text datasets to understand and generate human language. Examples include GPT-4, Claude, Gemini, and Llama. Typically built on the transformer architecture.
Transformer: A neural network architecture introduced in 2017 that uses self-attention mechanisms to process sequential data. The foundation of nearly all modern LLMs and many vision models.
Diffusion Model: A type of generative model that learns to create data by reversing a gradual noising process. Used in image generators like DALL-E, Midjourney, and Stable Diffusion.
Foundation Model: A large AI model trained on broad data that can be adapted to many downstream tasks. The base layer that powers chatbots, code assistants, image generators, and other applications.
Multimodal: AI models that can process and generate multiple types of data — text, images, audio, video — within a single system. Examples include GPT-4o and Gemini.
Fine-tuning: The process of further training a pre-trained model on a specific dataset to improve its performance on a particular task or domain.
RAG (Retrieval-Augmented Generation): A technique that enhances LLM responses by first retrieving relevant documents from an external knowledge base, then using them as context for generation. Reduces hallucinations and keeps answers grounded in source material.
Mixture of Experts (MoE): A model architecture where multiple specialized sub-networks (experts) handle different parts of the input, with a gating mechanism routing each input to the most relevant experts. Allows larger models with less computation per inference.
Chain of Thought (CoT): A prompting technique where the model is encouraged to show its reasoning step by step before giving a final answer. Improves performance on math, logic, and complex reasoning tasks.
Embedding: A numerical representation of data (text, images, etc.) as a vector in a high-dimensional space. Similar items have similar embeddings, enabling semantic search and clustering.
Inference: The process of running a trained model to generate predictions or outputs from new inputs. Distinct from training, inference is what happens when you use an AI model.
RLHF (Reinforcement Learning from Human Feedback): A training technique where human preferences are used to fine-tune AI models. Humans rank model outputs, and a reward model learns from those rankings to guide further training. Key to making LLMs helpful and safe.
Prompt Engineering: The practice of crafting effective inputs (prompts) to get desired outputs from AI models. Includes techniques like few-shot examples, system prompts, and structured instructions.
Context Window: The maximum amount of text (measured in tokens) a language model can process in a single interaction. Larger context windows allow the model to consider more information at once.
Token: The basic unit of text that language models process. Roughly equivalent to 3/4 of a word in English. Models have limits on tokens they can process (context window) and generate.
Alignment: The challenge of ensuring AI systems behave in ways that are consistent with human values, intentions, and safety requirements. A major focus of AI safety research.
AGI (Artificial General Intelligence): A hypothetical AI system with human-level cognitive abilities across a wide range of tasks. Unlike current narrow AI, AGI would be able to learn and reason about any domain. No consensus exists on its timeline or definition.
Hallucination: When an AI model generates plausible-sounding but factually incorrect or fabricated information. A known limitation of current LLMs that makes fact-checking AI outputs important.
Benchmark: A standardized test used to evaluate and compare AI model performance. Common benchmarks include MMLU (knowledge), HumanEval (coding), and GSM8K (math reasoning).
Synthetic Data: Data generated by AI models rather than collected from real-world sources. Used to augment training datasets, create privacy-safe alternatives, and bootstrap AI training where real data is scarce.
LoRA (Low-Rank Adaptation): An efficient fine-tuning technique that adds small trainable matrices to a frozen pre-trained model, drastically reducing the compute and memory needed for adaptation.
Quantization: Reducing the numerical precision of a model's weights (e.g., from 32-bit to 4-bit) to make it smaller and faster, often with minimal quality loss. Enables running large models on consumer hardware.
AI Agent: An AI system that can autonomously plan and execute multi-step tasks, use tools, browse the web, write code, and interact with external systems to achieve a goal.
MCP (Model Context Protocol): An open protocol that standardizes how AI applications connect to external tools and data sources. Lets AI agents interact with APIs, databases, and services through a unified interface.
Agentic AI: AI systems designed to operate with a degree of autonomy — planning actions, using tools, and making decisions to accomplish complex tasks with minimal human intervention.
GPU (Graphics Processing Unit): A processor originally designed for graphics rendering, now essential for AI training and inference due to its ability to perform many parallel computations. NVIDIA dominates the AI GPU market.
TPU (Tensor Processing Unit): Google's custom AI accelerator chip, designed specifically for machine learning workloads. Used to train and serve Google's own AI models including Gemini.
Training Run: The process of training an AI model on data, which can take weeks or months on thousands of GPUs and cost millions of dollars for frontier models.
Open Source Model: An AI model whose weights are publicly released, allowing anyone to download, run, fine-tune, and modify it. Examples include Llama, Mistral, and Qwen. Licensing terms vary.
Reinforcement Learning: A machine learning approach where an agent learns by interacting with an environment, receiving rewards or penalties for its actions. Used in game-playing AI, robotics, and RLHF.
Natural Language Processing (NLP): The field of AI focused on enabling computers to understand, interpret, and generate human language. LLMs are the latest major advancement in NLP.
Computer Vision: The field of AI that enables machines to interpret and understand visual information from images and video. Applications include object detection, image generation, and autonomous driving.
Reasoning Model: LLMs specifically trained to perform step-by-step logical reasoning before answering. Examples include OpenAI's o1/o3 and DeepSeek R1. Use chain-of-thought internally to improve accuracy on complex problems.
Superintelligence: A hypothetical AI system that vastly exceeds human cognitive abilities in all domains. A central topic in AI safety discussions about long-term existential risk.
Distillation: A technique where a smaller 'student' model is trained to replicate the behavior of a larger 'teacher' model. Produces compact models that retain much of the original's capability at lower cost.
Artificial Analysis: An independent platform that benchmarks and compares AI models on quality, speed, and pricing. Maintains leaderboards including the AA-Omniscience Index for factual accuracy. Widely referenced for neutral, third-party model evaluations.
GPQA Diamond: A graduate-level science benchmark with questions written and validated by domain experts. The Diamond subset contains the hardest questions, covering physics, chemistry, and biology. Used to test whether AI models can reason about advanced scientific concepts.
LiveCodeBench: A coding benchmark that uses continuously updated competitive programming problems published after model training cutoffs. Prevents data contamination by only including new problems, providing a more honest measure of coding ability than static benchmarks.
OmniDocBench: A benchmark for evaluating AI models on document understanding tasks including OCR, table extraction, layout analysis, and information retrieval across diverse document types like PDFs, scanned images, and forms.
Codeforces: A competitive programming platform with an Elo-style rating system. AI models are evaluated by solving Codeforces problems and receiving a rating — scores above 2000 indicate expert-level competitive programming ability.
ICPC: The International Collegiate Programming Contest, the oldest and most prestigious competitive programming competition. AI models earning 'gold medal' performance solve problems at a level comparable to top university teams worldwide.
IMO: The International Mathematical Olympiad, the world's most prestigious pre-university math competition. AI models achieving IMO-level scores can solve multi-step proof-based problems requiring deep mathematical reasoning.
BrowseComp: An OpenAI benchmark that tests AI models on autonomous web browsing and information retrieval. Tasks require navigating multiple websites, following links, and synthesizing information to answer complex questions.
SWE-bench: A benchmark that evaluates AI models on real-world software engineering tasks from GitHub issues. Models must understand codebases, locate bugs, and produce working patches. SWE-bench Verified is a human-validated subset with confirmed solvable tasks.
OSWorld-Verified: A benchmark that tests AI agents on real desktop operating system tasks — file management, web browsing, office apps — inside full virtual machines. Measures whether an agent can complete multi-step computer workflows end to end.
AndroidWorld: A benchmark that evaluates AI agents on real Android device tasks across multiple apps. Tests whether an agent can navigate mobile interfaces, tap buttons, fill forms, and complete multi-step workflows on a phone.
AIME 2025: The 2025 American Invitational Mathematics Examination, a challenging math competition widely used as a benchmark for AI mathematical reasoning. Scores indicate how well a model handles multi-step problem solving.
MIT License: A permissive open-source license that lets anyone use, modify, and distribute software (or model weights) for any purpose, including commercial use, with minimal restrictions. Only requires preserving the copyright notice.
Apache 2.0: A permissive open-source license similar to MIT but with an explicit patent grant — users get a license to any patents the contributors hold on the code. Widely used for AI model weights, including Qwen and Llama.
Open Weights: When an AI company publishes the trained parameters of a model so anyone can download and run it. Separate from open source, which refers to releasing the code. Licensing terms vary — some allow commercial use, others restrict it.

Back to 7min.ai