AI Glossary

An AI agent is an autonomous program capable of perceiving its environment, making decisions, and executing actions to achieve a goal. It often combines an LLM with external tools.

Application Programming Interface allowing software to communicate with each other. AI APIs (OpenAI, Anthropic) allow you to integrate models into your applications.

Core mechanism in Transformers that allows the model to weigh the relative importance of each element in a sequence. It is the foundation of the architecture behind GPT and BERT.

Use of AI and software tools to perform repetitive tasks without human intervention. Automated workflows often combine multiple APIs and models.

Processing data in batches rather than in real-time. In AI, this allows processing large volumes of requests simultaneously to optimize costs and performance.

Bidirectional Encoder Representations from Transformers. A Google model that understands word context in both directions of a sentence, revolutionizing NLP in 2018.

Systematic distortion in AI model outputs caused by unrepresentative training data or design choices. It is a major ethical concern in AI development.

OpenAI's conversational interface based on GPT models. Launched in late 2022, it democratized access to large language models for the general public.

Machine learning task of assigning a predefined category to input data. Examples include spam detection, sentiment analysis, and image sorting.

AI assistant developed by Anthropic, known for its safety-focused approach (Constitutional AI). It excels at analyzing long documents and reasoning tasks.

Unsupervised learning technique that automatically groups similar data into clusters. Used for customer segmentation, market analysis, and more.

Field of AI enabling machines to interpret and analyze images and videos. Applications include facial recognition, autonomous vehicles, and quality control.

Maximum amount of text (in tokens) that an LLM can process in a single request. GPT-4 offers 128K tokens, Claude up to 200K tokens.

Technique of creating new training data from existing data through transformations (rotation, noise, paraphrasing). Improves model generalization.

Structured collection of data used to train, validate, or test an AI model. Dataset quality is a key determinant of model performance.

Subfield of machine learning using multi-layered neural networks. It drives breakthroughs in computer vision, NLP, and content generation.

Type of generative model that learns to create images by reversing a progressive noising process. Stable Diffusion and DALL-E use this approach.

Dense numerical representation of a word, sentence, or document in a vector space. Allows measuring semantic similarity between texts.

One complete pass through the entire training dataset. The number of epochs is a key hyperparameter that influences model convergence.

Quantitative measures used to assess model performance: accuracy, recall, F1-score, perplexity, BLEU, and more.

A model's ability to learn a new task from just a few examples provided in the prompt. It sits between zero-shot learning and fine-tuning.

Process of adapting a pre-trained model on a specific dataset for a particular task. Allows specializing an LLM without retraining from scratch.

Large model pre-trained on vast unlabeled data, serving as a base for multiple tasks. GPT-4, Claude, and LLaMA are foundation models.

Generative Adversarial Networks: architecture where two networks (generator and discriminator) compete to produce realistic content. Pioneers of image generation.

Generative Pre-trained Transformer, a family of models by OpenAI. GPT-4 is one of the most advanced LLMs, capable of understanding and generating text, code, and analyzing images.

Fundamental optimization algorithm in deep learning that adjusts model weights by following the direction of the loss function gradient descent.

Technique of anchoring LLM responses in factual, verifiable data. RAG is a common grounding method.

Phenomenon where a language model confidently generates false or fabricated information. It is one of the main challenges of current LLMs.

Parameter set before model training (learning rate, number of layers, batch size). Its tuning directly impacts model performance.

Phase of using a trained model to make predictions on new data. This is when the model 'works' to answer your queries.

Field of computer science aimed at creating systems capable of simulating human cognitive abilities: learning, reasoning, perception, and decision-making.

Large Language Model. AI model trained on massive text corpora, capable of understanding and generating natural language. Examples: GPT-4, Claude, Mistral.

Low-Rank Adaptation. Efficient fine-tuning technique that adds small adaptation matrices to the model, allowing specialization with far fewer resources.

Mathematical function measuring the gap between model predictions and expected values. The goal of training is to minimize this function.

Subfield of AI where algorithms learn automatically from data without being explicitly programmed. Includes supervised, unsupervised, and reinforcement learning.

AI image generation tool accessible via Discord. Known for the artistic quality of its images, it is widely used in design and visual creation.

Statistical or neural model trained to predict and generate text. Modern language models (LLMs) are based on the Transformer architecture.

Describes a model capable of processing multiple data types: text, image, audio, video. GPT-4V and Gemini are examples of multi-modal models.

Open-source workflow automation platform that connects applications and AI models through a visual no-code interface.

Artificial neural network: model inspired by the human brain, composed of layers of interconnected neurons. The foundation of deep learning.

Natural Language Processing. Branch of AI dedicated to understanding and generating human language by machines.

Internal model values adjusted during training (neuron weights). A model like GPT-4 has hundreds of billions of parameters.

Natural language instruction or question sent to an AI model to get a response. Prompt quality largely determines response quality.

Art and science of designing optimal prompts to get the best results from an LLM. Includes techniques like chain-of-thought, few-shot, and role-playing.

Retrieval-Augmented Generation. Technique combining an LLM with an external knowledge base to generate more accurate and up-to-date responses.

Learning method where an agent learns by interacting with an environment, receiving rewards or penalties to guide its behavior.

Reinforcement Learning from Human Feedback. Technique used to align LLMs with human preferences using feedback from human evaluators.

Recurrent Neural Network. Deep learning architecture that processes sequences step by step with internal memory. Largely replaced by Transformers.

Open-source image generation diffusion model developed by Stability AI. It can be run locally and customized with LoRA models.

Training method where the model learns from labeled data (input-output pairs). It is the most common method in ML.

Technology for generating artificial voice from text (TTS). Modern models produce voices nearly indistinguishable from human speech.

Hyperparameter controlling the degree of randomness in LLM responses. Low temperature (0) gives more deterministic responses, high temperature gives more creative ones.

Basic processing unit of an LLM. A token represents roughly 3/4 of a word in English. API usage cost is typically charged per million tokens.

Technique of reusing a model trained on one task to adapt it to another similar task. Fundamental principle of foundation models.

Neural network architecture introduced in 2017 in the paper "Attention Is All You Need". Foundation of all modern LLMs thanks to its attention mechanism.

Training method where the model discovers patterns in data without labels. Used for clustering, dimensionality reduction, and anomaly detection.

Database specialized in storing and searching vectors (embeddings). Essential for RAG, semantic search, and recommendation systems.

See Computer Vision. Field of AI that enables machines to analyze and understand visual content (images, videos).

Vector representation of a word capturing its semantic meaning. Word2Vec and GloVe are classic methods, now integrated into Transformers.

Automated sequence of steps and actions using AI tools. Allows automating complex processes by connecting multiple services and models.

A model's ability to perform a task without any prior examples. Modern LLMs excel at zero-shot thanks to their massive training.