Artificial Intelligence

A function that transforms a neural network unit’s input into its output, typically introducing nonlinearity or constraining output values.

Adam is a gradient-based optimization algorithm that combines momentum with adaptive, coordinate-wise scaling based on estimates of gradient moments.

Artificial intelligence systems that pursue goals through planning, tool use, action, and feedback with varying degrees of autonomy.

An artificial intelligence system that perceives its environment and selects actions to pursue goals with some degree of autonomy.

AI hallucination is the generation of seemingly credible content that is false, unsupported by evidence, or inconsistent with a system’s input.

AI slop is a disparaging term for low-quality AI-generated content, often mass-produced and distributed with little editorial oversight.

AlexNet is a deep convolutional neural network whose victory in the 2012 ImageNet competition helped establish deep learning in computer vision.

AlphaFold is a family of artificial intelligence systems for predicting three-dimensional protein structures and biomolecular complexes.

AlphaGo is DeepMind’s Go-playing artificial intelligence system, combining neural networks, reinforcement learning, and tree search to achieve professional and superhuman performance.

Anomaly detection identifies observations, events, or patterns that depart from an established model of expected behavior.

Arthur Samuel was an American computer scientist whose learning checkers programs helped establish machine learning as a field of research.

Artificial intelligence is the field of computer science that builds systems able to do tasks linked to human intelligence, such as reasoning, learning, perception and language.

A computational model of interconnected units that learns patterns by adjusting numerical parameters, supporting prediction, representation learning, and data generation.

An attention mechanism dynamically weights input representations, enabling neural networks to select and combine information relevant to a particular computation.

An autoencoder is a neural network that learns data representations by encoding inputs and reconstructing them under architectural or training constraints.

Backpropagation efficiently computes gradients through a computational graph, enabling the training of neural networks and other differentiable models.

A text representation that records word occurrences while disregarding their order, enabling numerical analysis, document retrieval, and machine learning.

Batch normalization standardizes neural-network activations using batch statistics and learned scale and shift parameters to support efficient training.

BERT is a pretrained Transformer encoder that learns bidirectional text representations and can be adapted to many natural-language processing tasks.

Bootstrap aggregating, or bagging, combines models trained on resampled data to stabilize predictions and reduce variance.

Chain-of-thought prompting elicits intermediate reasoning steps from language models before a final answer, sometimes improving performance on tasks requiring multiple steps.

ChatGPT is OpenAI’s conversational AI service for generating content, answering questions, and performing tool-assisted tasks.

Anthropic's agentic coding tool, which runs in a terminal, IDE, desktop app or browser to read, edit and test code.

Claude Fable is Anthropic’s family of generally available AI models for complex coding, research, and long-running knowledge work.

Claude Mythos is Anthropic’s restricted-access language-model series, developed for advanced work including cybersecurity and life-sciences research.

Claude Opus 5.5 is an Anthropic AI model released in September 2026 for complex coding, knowledge work, and tool-assisted tasks.

Computer vision develops computational methods for extracting information about objects, scenes, and activities from images, videos, and other visual measurements.

A context window is the bounded amount of tokenized information a language model can use during a generation request.

A representation-learning approach that learns similarities and distinctions by comparing related and unrelated examples.

A neural network architecture that uses shared, locally applied filters to learn patterns in images and other grid-structured data.

Cross-attention lets one set of neural representations selectively retrieve information from another through learned query–key matching.

American artificial intelligence researcher and entrepreneur, co-founder and chief executive officer of Anthropic.

Data augmentation creates additional training examples through transformations or combinations of existing data, helping models learn useful invariances and generalize beyond their training samples.

Data leakage occurs when information unavailable under the intended prediction or evaluation conditions influences model development, making reported performance misleading.

Data mining uses computational and statistical methods to discover patterns, relationships, and predictive models in datasets.

A supervised learning method that constructs branching rules from labeled examples to predict categories or numerical values.

Deep Blue was an IBM chess-playing supercomputer that defeated reigning world champion Garry Kasparov in a six-game match in 1997.

Deep learning is a branch of machine learning that learns layered representations of data, typically using neural networks with multiple processing layers.

DeepSeek is a Chinese artificial intelligence company based in Hangzhou, known for releasing large language models with open weights, especially the reasoning model R1.

DeepSeek 4.1 is a generation of multimodal artificial intelligence models, introduced with DeepSeek-V4.1-Flash in September 2026.

DeepSeek 4.1 Flash (DeepSeek-V4.1-Flash) is a multimodal Mixture-of-Experts language model released in September 2026, with a 552-billion-parameter backbone and a one-million-token context window.

DeepSeek-R1 is an open-weight language-model family developed to improve reasoning through reinforcement learning and distillation.

A generative machine-learning model that learns to produce data by reversing a gradual noise-corruption process.

Distribution shift is a difference between the data distributions used to develop a predictive model and those encountered during evaluation or deployment.

Dropout is a neural-network regularization technique that randomly suppresses activations during training to reduce overfitting.

Early stopping limits iterative model training by selecting a stopping point, often using validation performance, to control overfitting and computational cost.

Empirical risk minimization selects a predictive model by minimizing its average loss on observed data, using that loss as a proxy for population risk.

A neural-network design in which an encoder represents input data and a decoder uses that representation to produce an output.

Ensemble learning combines multiple predictive models to improve accuracy, stability, or uncertainty estimation through methods such as bagging, boosting, voting, and stacking.

Explainable artificial intelligence develops methods for making AI behavior and outputs understandable to people while assessing whether explanations faithfully represent the underlying system.

Feature engineering is the creation, transformation, and selection of data representations used as inputs to machine-learning models.

Feature extraction transforms raw observations into representations suitable for machine learning and other forms of computational analysis.

Feature scaling transforms numerical input variables to comparable ranges or statistical scales, affecting distances, optimization, and regularization in machine learning.

Feature selection identifies a subset of input variables to improve predictive modeling, reduce computational or measurement costs, and support interpretation.

Computer scientist known for ImageNet, visual recognition research, and initiatives in human-centered artificial intelligence and inclusive AI education.

Fine-tuning adapts a pretrained neural network to a task or domain through additional training of selected parameters.

FlashAttention is a family of GPU attention algorithms that reduces memory traffic through tiled computation, accelerating Transformers without sparsifying dense attention.

A foundation model is a broadly trained machine-learning model that can be adapted to many downstream tasks, providing a shared basis for specialized artificial-intelligence systems.

American psychologist who developed the perceptron, an early learning neural network that helped establish the foundations of machine learning.

Gemini 3 is Google DeepMind’s third-generation family of multimodal artificial intelligence models, introduced in November 2025.