Artificial Intelligence

A family of language models developed by Alibaba Cloud, combining multilingual generation with configurable reasoning.

A machine-learning method that combines randomized decision trees to make classification or regression predictions.

A rectified linear unit is a neural-network activation function that outputs zero for negative inputs and leaves positive inputs unchanged.

A recurrent neural network processes sequential data by repeatedly updating an internal state that carries information between sequence positions.

Recursive self-improvement (RSI) is a process in which an AI system contributes to improving AI systems, including its own successors, so that each gain makes the next gain easier.

Reinforcement learning is a machine-learning approach in which agents learn decision-making policies by optimizing cumulative rewards from interaction or recorded experience.

Reinforcement learning from human feedback trains models using human judgments as a source of reward, often to improve instruction following and other desired behaviors.

Representation learning discovers useful features from data, enabling machine-learning systems to organize information and transfer it across tasks.

Retrieval-augmented generation combines information retrieval with generative models to produce responses informed by external sources.

American entrepreneur and investor, co-founder of OpenAI and former president of Y Combinator.

Self-attention is a neural-network mechanism that contextualizes elements of an input by weighting and combining information from other elements of the same input.

A machine-learning approach that derives training targets from the data itself, enabling learning without manually supplied task labels.

Semantic segmentation assigns a category label to each image pixel, producing a detailed map of objects, materials, and scene regions.

A machine-learning framework that combines labeled and unlabeled examples to learn predictive models when task-specific annotations are limited.

OpenAI’s second-generation video-and-audio generation model, introduced in 2025 and discontinued in 2026.

Speech recognition is the computational conversion of spoken language into text, using acoustic analysis and statistical or neural models.

Stargate is an OpenAI-led infrastructure initiative announced in 2025 to expand large-scale artificial intelligence computing capacity through data centers, energy infrastructure, and technology partnerships.

Chinese artificial intelligence company based in Shanghai, known for the multimodal Step series of foundation models.

Supervised learning trains predictive models from examples that pair inputs with known target outputs.

A support vector machine is a learning model that uses margin-based optimization and optional kernels for classification, regression, and related tasks.

Symbolic artificial intelligence represents knowledge explicitly and uses rules, logic, and search to derive conclusions or select actions.

A family of learning methods that updates predictions using rewards and differences between successive estimates of future outcomes.

A test set is data reserved for evaluating a trained machine-learning model independently of the decisions used to develop it.

Tokenization divides text into computational units, such as words, subwords, characters, or bytes, for language processing and modeling.

Training data consists of examples used to fit machine-learning models, shaping their learned patterns, capabilities, and limitations.

A machine-learning approach that reuses knowledge from one task or domain to improve learning in another.

A neural-network architecture that uses attention to process sequences, supporting language understanding, text generation, and image recognition.

Underfitting occurs when a learned model fails to capture important patterns in its training data, limiting predictive performance on both familiar and unseen examples.

A machine-learning paradigm that discovers patterns, representations, or probability distributions in data without externally supplied target labels.

A validation set is data reserved for evaluating and selecting machine-learning models during development, distinct from training data and final test data.

The vanishing gradient problem occurs when derivatives shrink during backpropagation, weakening learning signals across many neural-network layers or time steps.

A variational autoencoder is a probabilistic generative model that learns latent representations and approximate inference jointly through a variational training objective.

Veo 3 is Google DeepMind’s generative video model, introduced in 2025, that creates short video clips with synchronized speech, sound effects, and ambient audio.

A neural-network architecture that represents images as sequences of patches and processes them with Transformer layers for visual recognition and representation learning.

A word embedding represents a word as a numerical vector whose learned relationships capture aspects of linguistic usage and meaning.

Word2vec is a family of predictive models that learns dense word vectors from patterns of neighboring words in text.

An internal model of an environment that enables an AI system to predict events and anticipate the consequences of actions.

xAI is an AI company founded by Elon Musk in 2023, known for the Grok models; it absorbed the social platform X in 2025, was acquired by SpaceX in 2026, and reportedly rebranded as SpaceXAI.

Yang Zhilin is a Chinese artificial intelligence researcher and entrepreneur, co-founder of Moonshot AI and co-author of Transformer-XL and XLNet.