[
{"term":"Large language model (LLM)","slug":"large-language-model","short":"A neural network trained on large amounts of text to predict the next token, which lets it generate and analyze language.","long":"LLMs are usually transformer-based models with billions of parameters. They are pretrained on broad text data, then often fine-tuned with instructions and human feedback so they follow requests. Chat assistants, coding assistants and many AI agents are built on LLMs.","related":["transformer","token","fine-tuning"]},
{"term":"Transformer","slug":"transformer","short":"The neural-network architecture, introduced in 2017, that uses attention to relate every part of an input to every other part.","long":"The transformer was described in the 2017 paper “Attention Is All You Need.” Its self-attention mechanism processes sequences in parallel, which made very large language, image and audio models practical to train.","related":["large-language-model","attention"]},
{"term":"Attention","slug":"attention","short":"A mechanism that lets a model weigh how relevant each part of its input is when producing each part of its output.","long":"Self-attention computes, for every token, a weighted mix of all other tokens in the context. It is the core operation of transformers.","related":["transformer","context-window"]},
{"term":"Token","slug":"token","short":"The unit of text a model reads and writes, often a word or a piece of a word.","long":"Text is split into tokens by a tokenizer before a model processes it. Model limits and API prices are usually measured in tokens.","related":["context-window","large-language-model"]},
{"term":"Context window","slug":"context-window","short":"The maximum number of tokens a model can consider at once, including both your input and its output.","long":"Anything outside the context window is not seen by the model for that request. Larger windows let a model work with longer documents or conversations.","related":["token","retrieval-augmented-generation"]},
{"term":"Fine-tuning","slug":"fine-tuning","short":"Further training a pretrained model on a smaller, specific dataset to change its behavior or specialize it.","long":"Fine-tuning adapts a general model to a task, style or domain. Instruction tuning and reinforcement learning from human feedback are common forms.","related":["large-language-model","rlhf"]},
{"term":"RLHF","slug":"rlhf","short":"Reinforcement learning from human feedback: training a model using human preference ratings of its outputs.","long":"People compare model responses; those preferences train a reward model, which is then used to optimize the language model toward responses people prefer.","related":["fine-tuning","alignment"]},
{"term":"Retrieval-augmented generation (RAG)","slug":"retrieval-augmented-generation","short":"A technique where relevant documents are fetched and added to the model's input so its answer can draw on them.","long":"RAG systems search a knowledge base (often with embeddings), place the best matches in the prompt, and ask the model to answer from them. It helps keep answers current and lets them cite sources.","related":["embedding","context-window"]},
{"term":"Embedding","slug":"embedding","short":"A list of numbers that represents the meaning of text, images or audio so similar items sit close together.","long":"Embeddings power semantic search, recommendations, clustering and retrieval-augmented generation.","related":["retrieval-augmented-generation"]},
{"term":"Hallucination","slug":"hallucination","short":"When an AI model states something false or unsupported as if it were true.","long":"Hallucinations come from models generating plausible text rather than looking facts up. Grounding with sources, retrieval and verification reduce but do not eliminate them.","related":["retrieval-augmented-generation","large-language-model"]},
{"term":"AI agent","slug":"ai-agent","short":"A system that uses a model to plan and take multi-step actions with tools, such as browsing, running code or calling APIs.","long":"Agents loop: decide on an action, call a tool, read the result, and continue until a goal is met. Their usefulness and risk both come from acting in real systems.","related":["large-language-model","alignment"]},
{"term":"Alignment","slug":"alignment","short":"The work of making AI systems pursue the goals and values their developers and users intend.","long":"Alignment research covers training methods, evaluations, interpretability and oversight so that capable systems behave safely and as intended.","related":["rlhf","agi"]},
{"term":"AGI","slug":"agi","short":"Artificial general intelligence: a hypothetical AI able to match or exceed humans across most cognitive tasks.","long":"There is no single agreed definition or test for AGI. Forecasting platforms and researchers publish differing estimates of when, or whether, it will arrive.","related":["alignment","ai-agent"]},
{"term":"Diffusion model","slug":"diffusion-model","short":"A generative model that creates images, audio or video by learning to remove noise step by step.","long":"Diffusion models start from random noise and iteratively denoise it toward an output that matches a prompt. Many AI image and music generators use them.","related":["multimodal"]},
{"term":"Multimodal","slug":"multimodal","short":"Able to take in or produce more than one kind of data, such as text, images, audio and video.","long":"Multimodal models can, for example, describe a photo, answer questions about a chart, or speak and listen in real time.","related":["diffusion-model","large-language-model"]},
{"term":"Open-weight model","slug":"open-weight-model","short":"A model whose trained parameters are published so anyone can download and run it, subject to its license.","long":"Open weights are not always fully open source: training data and code may be withheld, and licenses can restrict use.","related":["large-language-model","fine-tuning"]}
]