AI Terms, Explained Simply

Artificial General Intelligence — AI that can perform any intellectual task a human can. Not yet achieved.

An AI system that autonomously takes actions, uses tools, and makes decisions to complete multi-step goals.

Application Programming Interface — a way for software to communicate. AI APIs let you send prompts and receive responses.

The neural network component that allows models to focus on relevant parts of input — the core of Transformers.

A model that generates output one token at a time, each token conditioned on all previous tokens.

A standardized test used to compare AI model performance (e.g. MMLU, HumanEval, HellaSwag).

Contrastive Language-Image Pre-training — OpenAI model connecting images and text, used in many image AI tools.

A prompting technique where you ask the model to reason step-by-step before giving the final answer.

The maximum amount of text (tokens) an LLM can process at once. Larger = more text it can "remember" per conversation.

Anthropic's method of training AI to be helpful, harmless, and honest using a set of principles.

A type of generative AI that creates images by learning to reverse a noise-adding process. Powers Stable Diffusion, DALL-E.

OpenAI's image generation model. DALL-E 3 is integrated into ChatGPT Plus.

A numerical vector representation of text/images that captures semantic meaning. Used in search and RAG systems.

Abilities that appear in large models that were not explicitly trained — like multi-step reasoning, coding, or translation.

Providing a few examples in the prompt to guide the model's output style or format without retraining.

Further training a pre-trained model on a specific dataset to specialize it for a particular task or style.

A large AI model trained on broad data that can be adapted for many tasks (GPT-4, Claude, Gemini, Llama).

Generative Adversarial Network — two networks (generator vs discriminator) competing to create realistic outputs.

Generative Pre-trained Transformer — OpenAI's architecture. GPT-4 powers ChatGPT.

When an AI confidently generates false information. A key challenge with current LLMs.

The process of running a trained model to generate outputs. Opposite of training.

A model fine-tuned to follow instructions rather than just predict the next token (e.g. GPT-4-turbo-instruct).

The date after which an LLM has no training data. Events after this date are unknown to the model.

The compressed mathematical space where AI models represent learned concepts and relationships.

Large Language Model — an AI trained on massive text datasets to understand and generate human language.

Low-Rank Adaptation — efficient fine-tuning method that adds small trainable matrices without modifying the full model.

Mixture of Experts — architecture where only a subset of model parameters activate per token, enabling larger models cheaply.

AI that processes multiple types of data — text + images + audio + video (e.g. GPT-4o, Gemini Ultra).

The numerical parameters of a trained neural network. "Open weights" means you can download and run the model.

Layers of interconnected mathematical nodes inspired by the brain. The foundation of modern AI.

Natural Language Processing — AI field focused on understanding and generating human language.

Providing exactly one example in a prompt to guide model output.

A measure of how well a language model predicts a text sample. Lower perplexity = better model.

The input text you give to an AI model. Quality of output heavily depends on prompt quality.

The practice of crafting prompts to get optimal outputs from AI models.

Reducing model precision (e.g. 16-bit to 4-bit) to run large models on consumer hardware with minimal quality loss.

Reinforcement Learning from Human Feedback — training technique where humans rate outputs to guide model behavior.

Retrieval-Augmented Generation — combining an LLM with external knowledge retrieval to reduce hallucinations.

Techniques to make AI behave safely and according to human values (RLHF, Constitutional AI, etc.).

Search using meaning/embeddings rather than exact keyword matching. Enables much better relevance.

Instructions given to an LLM before the user message — sets behavior, persona, and constraints.

Open-source image generation model by Stability AI. Runs locally, highly customizable.

A parameter controlling output randomness. Low (0.1) = focused/deterministic. High (1.0) = creative/random.

The basic unit LLMs process — roughly 4 characters or 0.75 words. Pricing is usually per 1000 tokens.

The neural network architecture (from "Attention Is All You Need" 2017) powering most modern LLMs.

The component that converts text to tokens before feeding to an LLM. Different models use different tokenizers.

A database optimized for storing and querying embeddings. Powers RAG systems (Pinecone, Weaviate, Chroma).

A model that understands both images and text — can describe images, answer visual questions, read documents.

OpenAI's open-source speech-to-text model. Powers many transcription tools.

Asking a model to perform a task with no examples in the prompt — just instructions.