Artificial Intelligence

Gemini Agent is Google’s task-execution AI, introduced in the Gemini app in 2025 and also used as the name of an enterprise agent announced in 2026.

Gemini CLI is Google’s open-source terminal-based AI agent for coding, information retrieval, and tool-assisted task automation.

Generalization is a machine-learning model’s ability to perform effectively on previously unseen data, rather than merely fitting its training examples.

A generative adversarial network learns to synthesize data through competition between a generator and a discriminator.

Generative artificial intelligence produces text, images, audio, and other content by learning patterns from data.

A generative pre-trained transformer is a transformer-based model trained to predict and generate sequences, forming the basis of the GPT language-model family.

Genie 3 is a Google DeepMind world model that generates interactive virtual environments and responds to users’ actions in real time.

GPT-5 is an OpenAI model family combining language generation, multimodal understanding, reasoning, and tool use.

GPT-6 is a family of proprietary large language models from OpenAI, first released in September 2026 as the Astra, Sol, and Luna variants.

GPT-6.1 is an OpenAI model family introduced in September 2026, represented by GPT-6.1 Sol for coding, computer use, and professional work.

Gradient boosting builds predictive models by sequentially adding learners that approximate loss-reducing changes to an existing ensemble.

Grok 4 is a reasoning-focused language model developed by xAI, released in July 2025 with integrated tool use and search capabilities.

A hyperparameter governs a model’s structure, learning procedure, or higher-level probability distribution, rather than serving as an ordinary fitted parameter.

Image classification is the computational assignment of one or more category labels to an image according to its visual content.

ImageNet is a large, hierarchically organized image dataset that became a foundational benchmark for visual recognition and deep learning.

In-context learning is a model’s ability to adapt its predictions using examples or instructions in its input, without updating its parameters.

John McCarthy was an American computer scientist who helped establish artificial intelligence, created Lisp, and pioneered logical approaches to machine reasoning.

K-means clustering partitions numerical observations into a specified number of groups by minimizing their squared distances from cluster means.

A nonparametric learning algorithm that predicts a sample’s class or numerical value from its nearest labeled examples.

A family of mathematical learning techniques that use kernel functions to perform computations in implicit feature spaces.

Kimi K2 is an open-weight language-model family developed by Moonshot AI, emphasizing sparse computation, coding, reasoning, and tool-using agents.

Kling AI is a family of video and image generation models developed by Kuaishou, a Chinese short-video company, first released in 2024.

The study of how computational systems encode knowledge and use it to derive conclusions, explain observations, and guide action.

A language model learns statistical patterns in text to estimate the likelihood of linguistic sequences, predict missing or subsequent tokens, and support language-processing tasks.

A large language model is a neural language model trained at scale to process and generate text, often supporting many tasks through a shared model.

A latent space is a model-defined space of unobserved variables or learned representations used to describe, reconstruct, or generate data.

Layer normalization standardizes neural-network activations within each example, using feature-wise statistics and usually learned scale and bias parameters.

A learning curve plots model performance against training-set size or training progress, helping characterize generalization, data requirements, and optimization behavior.

A learning rate controls the scale of parameter updates during iterative training, influencing optimization speed, stability, and model performance.

Chinese entrepreneur who co-founded the quantitative hedge fund High-Flyer and founded and leads the AI company DeepSeek.

Long short-term memory is a gated recurrent neural network architecture designed to learn dependencies across sequences while reducing vanishing-gradient difficulties.

Machine learning is the study of algorithms that improve their performance on a task by learning patterns from data, rather than following explicitly programmed rules.

Machine translation uses computational systems to render text or speech from one natural language into another.

Manifold learning identifies low-dimensional structure in high-dimensional data through nonlinear representations that preserve selected geometric or neighborhood relationships.

Manus is a general-purpose autonomous AI agent developed by Butterfly Effect, launched in 2025, whose 2025 acquisition by Meta was later unwound under Chinese regulatory pressure.

Marvin Minsky was an American artificial intelligence pioneer known for neural-network research, frame-based knowledge representation, and the Society of Mind theory.

Model Context Protocol is an open protocol that standardizes how artificial intelligence applications discover and use external tools, data sources, and reusable prompts.

Moonshot AI is a Beijing-based artificial intelligence company founded in 2023 that develops the Kimi chatbot and a family of large language models, several released with open weights.

Multi-head attention combines parallel attention computations with distinct learned projections to represent multiple relationships within or between sequences.

A multilayer perceptron is a feedforward neural network that learns nonlinear mappings through successive layers of weighted connections and activation functions.

Multimodal learning develops machine-learning systems that relate or combine information from different forms of data, such as text, images, audio, and video.

A family of probabilistic classifiers that applies Bayes’ theorem using the assumption that features are conditionally independent given the class.

Natural language processing studies computational methods for analyzing, interpreting, and generating human language in text and speech.

Nested cross-validation separates model selection from performance estimation by placing an inner validation procedure inside an outer evaluation loop.

Neural network inference is the execution of a trained neural network to produce predictions or generated outputs from input data, usually without updating its parameters.

An empirical relationship describing how neural-network performance changes with model size, training data, and computational resources.

Nonmonotonic reasoning allows conclusions drawn from incomplete information to be withdrawn when additional information becomes available.

Object detection identifies individual objects in images or video and estimates their locations, usually through labeled bounding boxes.

One-hot encoding represents each category as a binary vector with exactly one active component, enabling categorical data to be processed numerically.

OpenAI's AI coding agent family, from the 2021 code model to the 2025 Codex CLI, cloud agent and desktop app.

Overfitting occurs when a model learns sample-specific patterns that impair its performance on previously unseen data.

A perceptron is a linear classification model and learning algorithm that adjusts weighted inputs to distinguish between classes.

Perplexity measures predictive uncertainty by exponentiating entropy or average negative log-likelihood, especially in language-model evaluation.

Physical AI is artificial intelligence that perceives, reasons about and acts in the physical world through robots, vehicles and other embodied systems.

A policy is an agent’s rule for selecting actions from states, observations, or interaction histories in reinforcement learning.

Positional encoding supplies neural networks with information about the order or spatial location of elements in an input.

Prompt engineering is the design, testing, and refinement of inputs that guide generative artificial intelligence models toward specified outputs.

A security vulnerability in which untrusted inputs redirect a language-model application away from its authorized instructions.

An open-source machine learning framework providing GPU-accelerated tensor computations and dynamic neural network training.

Q-learning is a model-free reinforcement learning algorithm that estimates optimal action values from experience to learn decisions maximizing expected cumulative reward.