Every AI term, explained in plain language — search, browse A–Z, or filter by category.
Search or browse A–Z → open a term → ask Melo to go deeper.
Split Testing
A/B testing compares two system variants by randomly assigning eligible traffic and measuring predefined outcomes.
High-Speed Interconnect
An accelerator interconnect is the communication fabric that moves data between processors in a distributed AI system.
Accuracy is the fraction of evaluated predictions that exactly match the correct class or decision.
An action space is the set of operations available to an agent at a given point.
An activation function applies a nonlinear transformation to a neural network's intermediate values.
Active learning selects the unlabeled examples whose labels are expected to be most useful for improving a model.
Adversarial robustness is a system's ability to maintain acceptable behavior under intentionally manipulated inputs or conditions.
Adversarial testing evaluates a model or system with inputs intentionally designed to trigger errors or bypass controls.
Environment
An agent environment is the external state and set of interfaces an agent can observe or affect.
Agent Eval
Agent evaluation measures whether an agent reaches goals correctly, safely, and efficiently across multi-step tasks.
An agent loop repeatedly asks a model to choose the next action until it reaches a goal or stopping condition.
Agent memory is stored information that an agent can retrieve across steps or sessions.
Orchestration
Agent orchestration is the control layer that schedules model calls, tools, state transitions, and handoffs in an agent system.
Planning
Agent planning is the process of selecting and ordering actions intended to reach a goal.
Scratchpad
An agent scratchpad is temporary working state used to track intermediate information during a task.
AgentBench is a benchmark suite for evaluating language-model agents across multiple interactive environments.
Agentic describes an AI system that plans, takes actions, and adapts across multiple steps toward a goal, rather than just returning a single response to a single prompt.
Agentic slop is low-quality output produced by an AI agent working autonomously across many steps, such as bloated code, redundant files, or unnecessary changes nobody asked for.
An agentic workflow combines model-driven decisions with tools and state across multiple steps.
Machine Learning Accelerator
An AI accelerator is hardware specialized for the tensor operations common in machine learning workloads.
AI adoption is the process by which people and organizations integrate AI into real workflows and continue using it.
Agent
An AI agent is a system that observes context, chooses actions, and uses their results to pursue a goal.
Alignment
AI alignment is the effort to make an AI system's behavior consistent with intended goals, constraints, and human values.
API Pricing
AI API pricing is the charging structure for accessing hosted model capabilities through an application interface.
Benchmark
An AI benchmark is a standardized set of tasks, data, and scoring rules used to compare model performance.
Compute Cost
AI compute cost is the expense of the hardware and services used to train or run AI models.
Copilot
An AI copilot is an interactive assistant embedded in a workflow to help a person complete tasks while leaving them in control.
Doomer
AI doomer is a label, sometimes self-applied and sometimes used mockingly by critics, for people who believe advanced AI poses a serious risk of catastrophic or existential harm.
AI governance is the system of roles, policies, controls, and evidence used to direct and oversee AI development and use.
Guardrails
AI guardrails are controls that constrain model inputs, outputs, or actions according to defined policies.
AI IP
AI intellectual property concerns legal rights and obligations involving AI training materials, model artifacts, inventions, brands, and generated outputs.
An AI platform is an integrated set of tools for developing, evaluating, deploying, and operating AI applications.
AI regulation is the body of legally enforceable rules that applies to the development, supply, or use of AI systems.
AI Lab
An AI research lab is an organization or team that develops and studies artificial intelligence methods and systems.
AI safety is the study and practice of reducing harms from the design, deployment, and use of AI systems.
AISI
An AI safety institute is a public or independent organization that evaluates advanced AI risks and supports safety research or standards.
Slop
AI slop is low-quality, mass-produced generative AI content published without meaningful human review or a genuine audience need.
An AI startup is an early-stage company whose product, operations, or technical advantage substantially depends on artificial intelligence.
AI TCO · Total Cost of Ownership
AI total cost of ownership is the full ongoing cost of building, deploying, governing, and maintaining an AI capability.
Vendor Lock-In
AI vendor lock-in is the cost or difficulty of moving an AI workload from one provider or platform to another.
An AI winter is a period when investment, public interest, and institutional support for artificial intelligence decline after unmet expectations.
ARC · AI2 ARC
The AI2 Reasoning Challenge is a multiple-choice benchmark based on grade-school science questions.
American Invitational Mathematics Examination · AIME
The AIME benchmark uses problems from a proof-oriented mathematics competition to test mathematical reasoning in AI models.
All-reduce is a collective communication operation that combines values across workers and returns the result to every worker.
An annotation guideline defines how labelers should interpret examples and apply labels consistently.
Anthropomorphizing AI
Anthropomorphization is attributing human traits like feelings, intentions, or consciousness to an AI system based on its fluent, human-like output.
An API endpoint is a network-accessible interface through which an application submits model requests and receives results.
An API key is a secret or identifier presented by a client when calling an API.
API
An application programming interface is a defined way for software components to request data or operations from one another.
ANN
Approximate nearest-neighbor search finds likely close vectors without exhaustively comparing the query with every stored vector.
Abstraction and Reasoning Corpus · ARC
ARC-AGI is a benchmark of grid-transformation puzzles intended to test abstraction and few-example rule induction.
AGI
Artificial general intelligence is a proposed AI system with broad, adaptable competence across many cognitive tasks rather than one narrow domain.
AI
Artificial intelligence is the field of building computer systems that perform tasks associated with perception, reasoning, learning, language, or decision-making.
An assistant message is a model-generated or application-supplied response recorded with the assistant role.
Attention
An attention mechanism lets a model assign different importance to available pieces of information when producing a representation or output.
AuthN
Authentication verifies the identity of a user, service, or device.
AuthZ
Authorization determines whether an authenticated or anonymous actor may perform a particular action on a resource.
An autoencoder learns to reconstruct its input through a constrained intermediate representation.
An autonomous agent can choose and execute multiple actions without human approval at every step.
Autoscaling adjusts the number or size of serving resources in response to demand or operational signals.
Backprop
Backpropagation efficiently computes how a neural network's loss depends on each trainable parameter.
Batch size is the number of training examples used to estimate a gradient before a parameter update.
Batching groups multiple examples or requests so hardware processes them together.
Beam search keeps several high-scoring partial sequences while decoding instead of committing to one token path.
Model Behavior Policy
A behavior policy specifies how an AI system should respond across allowed, restricted, and ambiguous situations.
Bias mitigation aims to reduce systematic and unwanted performance or treatment differences across groups or contexts.
Beyond the Imitation Game Benchmark
BIG-bench is a collaborative benchmark suite containing diverse tasks for probing language-model capabilities and limitations.
BBH
BIG-Bench Hard is a focused collection of challenging tasks selected from the broader BIG-bench suite.
BLEU
BLEU is a text-generation metric that compares candidate and reference n-gram overlap with a penalty for overly short output.
Blue-green deployment maintains two complete production environments so traffic can switch between the current and new version.
BM25 is a lexical ranking function that scores documents using query-term frequency, term rarity, and document-length normalization.
Brain rot describes the perceived mental dulling from consuming large amounts of low-quality, algorithmically-fed content, much of it now AI-generated.
A browser agent uses a web browser as an environment for reading pages and performing permitted interactions.
Cache
Caching stores a reusable copy of data or computation so later requests can be served with less work.
Probability Calibration
Calibration describes whether predicted confidence levels correspond to observed outcome frequencies.
Canary Release
A canary deployment sends a small portion of production traffic to a new model or service version before broader rollout.
Capability control limits which actions, information, or resources an AI system can access.
Catastrophic forgetting is the loss of previously learned behavior when a model is trained on new data or tasks.
Centaur Chess · Human-AI Team
A centaur is a human and an AI system working together on a task, combining human judgment with machine speed and recall to outperform either working alone.
CoT
Chain of thought is an intermediate reasoning process used to connect a problem with an answer or action.
CoT Prompting
Chain-of-thought prompting encourages a model to use intermediate reasoning steps for a complex task.
LMSYS Chatbot Arena · LMSYS Arena
Chatbot Arena is a platform that compares conversational models using blinded pairwise votes from users.
ChatGPT is OpenAI's consumer chat product built on its GPT model family, released in November 2022.
Document Chunking
Chunking divides long source material into smaller units that can be indexed, retrieved, or placed into a model context.
Citation retrieval identifies source passages that support claims in a generated or planned answer.
Clanker is a slang insult for robots and AI, used humorously or dismissively toward chatbots, delivery robots, and AI tools in general.
Class imbalance occurs when some target classes have many more examples than others.
Classification is the task of assigning an input to one or more discrete categories.
Claude is Anthropic's family of large language models, built with an emphasis on helpfulness, harmlessness, and honesty as core training objectives.
Clauding is informal shorthand for working with or through Claude, especially the loop of prompting, reviewing, and iterating with the model.
Closed-Weight Model
Closed weights are model parameters that are not distributed to users and are accessed only through a controlled product or service.
Cloud AI is the delivery of model training, inference, and supporting tools through remotely managed computing services.
Clustering groups examples according to similarity without requiring predefined class labels.
A coding agent uses models and development tools to inspect, modify, and verify software.
A cold start is added request latency caused by creating a serving instance or loading a model that is not already ready.
Commonsense Question Answering
CommonsenseQA is a multiple-choice benchmark that tests everyday conceptual knowledge and reasoning.
CV
Computer vision is the field of enabling computers to extract useful information from images and video.
A confusion matrix counts predicted classes against actual classes to show where a classifier makes errors.
Constitutional AI trains or guides a model using a written set of behavioral principles.
Software Container
A container is an isolated process environment packaged with the application files and runtime dependencies it needs.
Containerization packages an AI service with its runtime, libraries, and configuration into an isolated deployable image.
Benchmark Contamination Audit
A contamination audit looks for overlap between evaluation material and data available during model training or development.
CDN
A content delivery network serves cached content from distributed locations closer to users.
Content filtering detects or blocks inputs and outputs that match a defined safety or usage policy.
Context injection adds selected external information to a model request so generation can use it.
Maximum Context Length
Context length is the maximum number of tokens a model interface can process across its input and generated output.
Context rot is the gradual decline in a model's response quality as a conversation or input grows longer, even before the context window fills up.
A context window is the bounded set of input and generated tokens a model can consider in one interaction.
In-Flight Batching
Continuous batching dynamically adds and removes generation requests from an active inference batch as sequences start and finish.
CI/CD · Continuous Integration/Continuous Delivery
CI/CD is a set of automated practices for integrating code changes, verifying them, and delivering deployable software.
Contrastive learning trains representations by pulling related examples closer and pushing unrelated examples apart.
Chat History
Conversation history is the sequence of prior messages included to give a model continuity in a multi-turn interaction.
CNN · ConvNet
A convolutional neural network applies shared local filters to detect spatial or temporal patterns.
Microsoft Copilot · GitHub Copilot
Copilot is Microsoft's brand for AI assistants built on OpenAI models, spanning GitHub Copilot for code, Microsoft 365 Copilot for office apps, and Windows Copilot for the operating system.
A corpus is a collected body of text, speech, images, or other material used for analysis or model development.
Cosine similarity compares vectors by the angle between them rather than their raw magnitude.
Scheduled Job
A cron job is a command or task launched automatically according to a recurring time schedule.
Cross-attention lets queries from one sequence attend to keys and values from another source.
CORS
Cross-Origin Resource Sharing is an HTTP mechanism that lets a server declare which browser origins may access a response.
Curriculum learning presents training examples in an intentional order, often moving from easier patterns to harder ones.
Data augmentation creates modified training examples that preserve the intended label or meaning.
Data contamination occurs when information that should be held out from training or development appears in data used to build a model.
Deduplication
Data deduplication identifies and removes repeated or near-repeated examples from a dataset.
Data governance defines decision rights, policies, and controls for collecting, accessing, retaining, and using data.
Data Annotation
Data labeling assigns target values, categories, spans, preferences, or other annotations to examples.
Data lineage traces how data moves and changes across pipelines, datasets, features, and model artifacts.
Normalization
Data normalization rescales or transforms features into a consistent numerical range or distribution.
Data parallelism runs copies of a model on different data batches and combines their training updates.
A data pipeline moves and transforms data from sources into datasets or features that models can use.
Data provenance records where data came from and how it was collected, transformed, and used.
Schema
A data schema defines the structure, types, relationships, and constraints expected for stored or exchanged data.
DB
A database is an organized system for storing, retrieving, and updating persistent data.
Dataset curation is the deliberate selection, organization, documentation, and maintenance of data for a defined use.
Dead internet theory is the claim, part conspiracy theory and part genuine trend, that much of the internet's content and traffic is now generated by bots and AI rather than real people.
Causal Language Model
A decoder-only model predicts the next token from earlier tokens and can generate sequences autoregressively.
Decoding
A decoding strategy is the procedure used to turn model output scores into a sequence or structured prediction.
DL
Deep learning is machine learning based on neural networks with multiple layers of learned representations.
DeepSeek is a Chinese AI lab, and the name of its family of language models, notable for open-weight releases that claimed competitive performance at a fraction of the usual training cost.
Defense in depth protects an AI system with multiple independent layers of prevention, detection, and recovery.
Dense retrieval represents queries and documents as learned dense vectors and ranks them by vector similarity.
DP
Differential privacy is a mathematical framework for limiting how much an output can reveal about any one person's data.
A diffusion model generates data by learning to reverse a gradual noising process.
DPO
Direct preference optimization trains a model from preferred and rejected response pairs without first fitting a separate reward model.
Distributed training coordinates model optimization across multiple processors or machines.
Docker is a toolset and image format used to build, distribute, and run software containers.
A document store holds the source text and metadata that a retrieval system returns or references.
DNS
The Domain Name System translates hierarchical domain names into network records such as IP addresses.
A doom loop is an AI agent getting stuck retrying the same failed action, or a conversation spiraling into repeated, unproductive corrections.
Early stopping ends training when performance on held-out data stops improving.
Edge Worker
An edge function is application code executed at distributed network locations near incoming users or data sources.
Edge inference runs a model near the data source on a device such as a phone, sensor, vehicle, or local gateway.
e/acc
Effective accelerationism, or e/acc, is an online movement that argues AI development should be pushed forward as fast as possible, treating rapid technological progress itself as a moral good.
Elo
An Elo rating is a relative score derived from pairwise wins, losses, and ties between models or systems.
An embedding layer maps discrete identifiers such as tokens or categories to learned dense vectors.
Emergent Ability
An emergent capability is a behavior that appears or becomes measurable as a model, dataset, or training process scales.
Sequence-to-Sequence Model · Seq2Seq
An encoder-decoder model first represents an input and then generates an output conditioned on that representation.
An encoder-only model produces contextual representations of an entire input rather than generating it token by token.
Enterprise AI is the application of AI within an organization's products, operations, and decision processes.
Env Var
An environment variable is a named value supplied to a process by its execution environment.
An epoch is one complete pass through the examples in a training dataset.
Eval Harness
An evaluation harness is software that runs test cases, captures model behavior, and computes repeatable metrics.
EM
Exact match scores a prediction as correct only when it equals an accepted reference after specified normalization.
AI Existential Risk · AI X-Risk
Existential risk from AI is the possibility that advanced AI could cause human extinction or permanently curtail humanity's future.
ExploitBench is a benchmark that scores how far an AI agent can progress toward exploiting a real software vulnerability, using a five-tier capability ladder instead of a pass/fail outcome.
F1
F1 score is the harmonic mean of precision and recall for a chosen positive class or averaging scheme.
A feature is a measurable or derived input value used by a machine learning model.
Feature engineering transforms raw data into model inputs that expose useful structure for a prediction task.
FFN · MLP Block
A feed-forward network transforms each input independently through learned linear layers and nonlinear activations.
Few-shot evaluation measures performance when the model is given a small number of demonstrations in the prompt.
Few-shot prompting gives a model a small set of input-output examples before asking it to handle a new case.
Finetuning
Fine-tuning continues training a pretrained model on a narrower dataset or objective to change its behavior for a target use.
Floating-Point Operations
FLOPs count floating-point arithmetic operations and are used to estimate the compute required by a model or workload.
Fast Takeoff · Hard Takeoff
Foom is shorthand, popular in AI safety circles, for a hypothetical fast takeoff where an AI system rapidly self-improves from roughly human-level to vastly superhuman intelligence.
A foundation model is a broadly trained model that can be adapted or prompted for many downstream tasks.
A foundation model provider develops or operates broadly capable models that other products can access or adapt.
A frontier model is a highly capable general-purpose model near the leading edge of current AI development.
Function calling is a structured interface in which a model selects a declared function and supplies arguments matching its schema.
GAIA
GAIA is a benchmark of real-world questions designed for general AI assistants that can reason and use tools.
GRU
A gated recurrent unit is a recurrent layer that uses learned gates to manage information across time.
Gemini is Google's family of large language models, built to be multimodal from the start and deeply integrated across Google's products.
Generalization is a model's ability to perform well on relevant examples it did not encounter during training.
GAN
A generative adversarial network trains a generator and a discriminator in competition to produce data resembling a training distribution.
GenAI
Generative AI refers to models that produce new content such as text, images, audio, video, code, or structured data.
The generative AI market is the ecosystem of products and services built around models that create text, images, audio, video, code, or other content.
Git is a distributed version-control system that tracks snapshots of a repository as linked commits.
Glazing is excessive, often ironic praise of an AI model's abilities, or a model's own tendency to flatter the user rather than give an honest answer.
ChatGLM
GLM is Z.ai's (formerly Zhipu AI's) family of language models, released from China with both open-weight and hosted API versions.
Glorified autocomplete is a dismissive description of large language models as simple next-word predictors with no real understanding, scaled up and dressed in a chat interface.
Graduate-Level Google-Proof Q&A
GPQA is a multiple-choice benchmark of difficult graduate-level questions in biology, physics, and chemistry.
Generative Pre-trained Transformer
GPT (Generative Pre-trained Transformer) is OpenAI's family of large language models, trained on broad text data and fine-tuned for chat, coding, and tool use.
Gradient accumulation combines gradients from several smaller batches before applying an optimizer update.
Gradient descent is an optimization method that updates parameters in the direction that reduces a differentiable objective.
GPU
A graphics processing unit is a highly parallel processor widely used for neural network training and inference.
GraphQL is a query language and runtime for APIs in which clients request fields from a typed schema.
Greedy Search
Greedy decoding chooses the highest-probability token at every generation step.
Grok is xAI's family of large language models, integrated into X (formerly Twitter) and offered as a standalone app and API.
Reference Label
Ground truth is the reference answer, label, or measurement used to evaluate a prediction.
Grounding connects a model output to provided evidence, observable data, or an external source of truth.
Grade School Math 8K
GSM8K is a dataset and benchmark of grade-school mathematics word problems with worked solutions.
AI Inbreeding
Habsburg AI is a nickname for model collapse, comparing a model trained on its own past outputs to the Habsburg royal dynasty's decline from generations of inbreeding.
A hallucination is model-generated content that is unsupported by the available evidence or conflicts with verifiable facts.
HellaSwag is a multiple-choice benchmark that tests commonsense reasoning by asking models to select a plausible continuation of a situation.
HBM
High-bandwidth memory is a stacked memory technology that provides accelerators with high data-transfer bandwidth.
HELM
Holistic Evaluation of Language Models is a framework for comparing language models across scenarios, metrics, and broader considerations.
Hypertext Transfer Protocol
HTTP is an application-layer protocol for exchanging requests and responses between network clients and servers.
Status Code
An HTTP status code is a three-digit response value that communicates the outcome category of an HTTP request.
Human Eval
Human evaluation asks people to judge model outputs against defined criteria or preferences.
HITL · Human-in-the-Loop
Human in the loop places a person at selected points in an automated or model-driven process.
Oversight
Human oversight gives people the information and authority to review, redirect, or stop an AI system.
HumanEval is a code-generation benchmark built from programming problems with function signatures, descriptions, and tests.
Hybrid search combines lexical matching with semantic or vector retrieval.
A hyperparameter is a configuration chosen outside ordinary parameter learning, such as a learning rate, depth, or regularization strength.
Hyperparameter tuning searches for training settings that produce strong validation performance.
Idempotency means repeating an operation has the same intended effect as performing it once.
ICL
In-context learning is a model's ability to adapt its behavior from instructions or examples placed in the current context without updating weights.
AI Incident Response
Incident response for AI is the process for detecting, containing, investigating, and learning from harmful or unexpected system behavior.
Search Indexing
Indexing transforms source content into data structures that make later retrieval efficient.
Inference is the computation a trained model performs to produce predictions or generated outputs from new inputs.
Model Runtime
An inference engine executes a model graph efficiently on available hardware.
Inference optimization changes execution, representation, or scheduling to reduce model latency, memory use, or cost.
An inference server is software that loads models and schedules prediction work on compute hardware.
An instruction dataset contains requests paired with desired responses or task outcomes for instruction tuning.
Instruction hierarchy defines which sources of model instructions take precedence when they conflict.
Instruction tuning adapts a model on examples framed as instructions and responses so it follows user requests more reliably.
AI Automation
Intelligent automation combines AI-based interpretation or generation with deterministic workflow execution.
IAA · Inter-Rater Agreement
Inter-annotator agreement measures how consistently multiple evaluators label or score the same examples.
Interpretability is the effort to understand why an AI system produces particular representations, decisions, or outputs.
An inverted index maps each searchable term to the documents or positions where it occurs.
Jagged Frontier
Jagged intelligence describes how AI models can be superhuman at some tasks while failing at simpler ones, instead of improving evenly across the board.
A jailbreak is an input or interaction strategy intended to make a model bypass its behavioral restrictions.
JavaScript Object Notation
JSON is a text format for representing objects, arrays, strings, numbers, booleans, and null values.
JWT
A JSON Web Token is a compact, signed representation of claims that can be passed between systems.
Kimi K2
Kimi is Moonshot AI's family of language models, released from China with open weights and long-context handling as a particular focus.
KB
A knowledge base is an organized collection of information maintained for lookup, reasoning, or retrieval.
Distillation
Knowledge distillation trains a smaller or simpler student model to reproduce behavior learned by a teacher model.
A knowledge graph represents entities and their relationships as connected nodes and edges.
K8s
Kubernetes is a system for deploying and coordinating containerized applications across a cluster.
Key-Value Cache
A KV cache stores attention keys and values from earlier tokens so autoregressive generation does not recompute them at every step.
LLM
A large language model is a neural network trained on extensive language data to predict or generate token sequences.
Late interaction retrieval compares token-level query and document representations after they have been encoded separately.
Latency is the elapsed time between starting a request or operation and receiving its result.
A latent space is an internal representation space whose dimensions encode learned factors or patterns in data.
LayerNorm
Layer normalization stabilizes neural network activations by normalizing features within an individual example.
The learning rate controls the size of parameter updates made by an optimizer during training.
Warmup
Learning rate warmup begins training with small updates and gradually raises the learning rate to its planned level.
Least privilege gives an AI system only the permissions needed for its current task and no broader access by default.
LiveCodeBench is a continuously updated benchmark for evaluating code-generation models on recently released programming problems.
Meta Llama
Llama is Meta's family of large language models, released with open weights that developers can download, fine-tune, and self-host.
Model-Based Evaluation · LLM-as-Judge
LLM as a judge uses a language model to score, rank, or critique other model outputs.
A load balancer distributes incoming inference requests across available service instances.
Load shedding deliberately rejects or degrades lower-priority work when a service lacks enough capacity.
Load-Bearing Comment · Load-Bearing Hack
Load-bearing describes a piece of code, comment, prompt, or workaround that looks disposable but is actually critical to the system working.
A logit is an unnormalized score a model produces for a possible class or next token before probabilities are computed.
LSTM
Long short-term memory is a recurrent architecture with gates that control what information is stored, exposed, and forgotten.
Objective Function
A loss function converts the difference between model behavior and a training objective into a value to minimize.
LoRA
Low-rank adaptation fine-tunes a model by learning small low-rank updates while keeping the original weights frozen.
ML
Machine learning is a branch of AI in which systems learn patterns from data to make predictions or decisions.
MTEB
The Massive Text Embedding Benchmark evaluates text embeddings across retrieval, classification, clustering, similarity, and related tasks.
MATH Dataset
The MATH benchmark evaluates mathematical problem solving with competition-style questions and worked reference solutions.
Mechanistic interpretability studies model behavior by analyzing internal computations and learned components.
Mesa-optimization is the possibility that a trained model internally develops an optimization process with its own learned objective.
Filtered Search
Metadata filtering restricts retrieval candidates using structured attributes such as date, source, tenant, or document type.
Microservice Architecture
Microservices are an architectural approach that divides a system into independently deployable services around bounded responsibilities.
Middleware is software that runs between a request entry point and the application's core handler or between system components.
Mistral AI
Mistral is a French AI lab, and the name of its family of language models, known for releasing both open-weight and proprietary models.
MoE
A mixture-of-experts model routes each input to a subset of specialized parameter blocks instead of activating every block.
Massive Multitask Language Understanding
MMLU is a benchmark that tests language models with multiple-choice questions drawn from many academic and professional subjects.
Massive Multi-discipline Multimodal Understanding and Reasoning
MMMU is a benchmark of college-level, multimodal questions that require interpreting both visual and textual information.
Mode collapse is a failure pattern where a generative model produces a narrow range of outputs instead of the full diversity present in its training data.
A model card is documentation describing a model's intended uses, evaluation results, limitations, and development context.
Checkpoint
A model checkpoint is a saved snapshot of parameters and related training state.
AI Inbreeding
Model collapse is the degradation that happens when a generative model is trained repeatedly on data produced by earlier AI models instead of on real, human-generated data.
Model commoditization is the process by which similar model capabilities become widely available and harder to differentiate on their own.
MCP
Model Context Protocol is a protocol for connecting AI applications to external tools, prompts, and contextual resources through a consistent interface.
Model distillation creates a deployment model by training it to imitate selected behavior of a larger or more capable model.
A model gateway is an application layer that routes requests across model providers, versions, or deployments behind one interface.
Model inference is the execution of a trained model on new inputs to produce scores, predictions, embeddings, or generated content.
Leaderboard
A model leaderboard ranks systems according to results on specified evaluations.
A model license sets the legal terms for using, modifying, hosting, or redistributing a model artifact.
AI Observability
Model observability collects signals needed to understand the behavior and health of an AI system in operation.
Model parallelism splits a model's computation or parameters across multiple devices because one device is insufficient or slower.
Parameter
A model parameter is a value learned from training data that influences the model's predictions.
Refusal
A model refusal is a response that declines a request because it is unsafe, disallowed, or outside configured boundaries.
Model serving is the infrastructure that makes trained models available for prediction requests.
Model versioning assigns stable identifiers and metadata to distinct model artifacts and configurations.
Weights
Model weights are learned numeric parameters that determine how inputs are transformed during inference.
AI Welfare
Model welfare is an emerging area of research and policy concerned with whether advanced AI systems could have morally relevant experiences, and how they should be treated if so.
Monolith
A monolithic architecture packages most application capabilities into one deployable unit.
Moore's law for AI compute is an analogy used to discuss recurring improvements in the cost or capability of computing available for AI.
MBPP
Mostly Basic Python Problems is a code-generation benchmark of short programming tasks paired with tests.
Multi-Turn Benchmark
MT-Bench evaluates conversational models on multi-turn questions spanning several task categories.
A multi-agent system coordinates multiple agents that have separate roles, context, or capabilities.
MHA
Multi-head attention runs several attention operations in parallel so a model can learn different relationships within the same input.
A multimodal architecture processes and combines more than one type of data, such as text, images, audio, or video.
A multimodal dataset contains aligned or related examples from more than one data type, such as images and captions or video and audio.
A multimodal model can process or generate more than one kind of data, such as text, images, audio, or video.
Weak AI
Narrow AI is an AI system designed or trained for a limited task, domain, or operating range.
NLP
Natural language processing is the field of enabling computers to analyze, understand, retrieve, or generate human language.
k-NN Search
Nearest neighbor search finds stored vectors closest to a query vector under a chosen distance measure.
A negative prompt describes content or properties that a generative model should avoid.
A neural network is a parameterized function built from connected layers of weighted transformations and nonlinear activations.
Non-Relational Database
NoSQL is a broad label for databases that use data models other than a traditional relational-table interface.
OAuth 2.0
OAuth is a framework that lets an application obtain limited access to protected resources without receiving the user's password.
ORM
Object-relational mapping connects application objects or types with tables and relationships in a relational database.
An observation is information an agent receives from its environment after an action or state change.
Local Inference
On-device inference runs a model directly on the user's hardware rather than sending every input to a remote server.
An open model is a model for which some artifacts or access rights are publicly provided, with the exact scope defined by its release.
Open source AI refers to AI software and artifacts released under terms that permit specified forms of inspection, modification, and redistribution.
Open-Weight Model
Open weights are model parameter files made available for others to download and run under stated license terms.
An optimizer turns computed gradients into parameter updates during model training.
Response Schema
An output schema defines the fields, types, and constraints expected from a model-backed endpoint.
Overfitting occurs when a model learns training-specific patterns that do not generalize well to new data.
Pairwise Evaluation
Pairwise comparison asks an evaluator to choose which of two outputs better satisfies a rubric.
The paperclip maximizer is a thought experiment about an AI given the goal of making paperclips that pursues it so single-mindedly it converts all available resources, and eventually threatens humanity, into paperclips.
A parameter server is a distributed architecture in which dedicated processes store and update shared model parameters.
Model Sharding
Parameter sharding divides model parameters and related states across devices instead of replicating them everywhere.
PEFT
Parameter-efficient fine-tuning adapts a model by training only a small portion of its parameters or added components.
pass@k
Pass at K estimates the chance that at least one of k generated candidates passes a correctness test.
Passage retrieval searches for relevant sections within documents rather than returning only whole files.
Perplexity measures how much probability a language model assigns to a token sequence, expressed as an exponentiated average negative log-likelihood.
Perplexity AI is an answer-engine product that combines a language model with live web search and returns cited, sourced answers instead of a list of links.
Position Encoding
Positional encoding supplies sequence-order information to an architecture that does not otherwise encode token position.
Positive Predictive Value
Precision is the fraction of predicted positive cases that are actually positive under the reference labels.
Preference data records judgments that one model output is better than another under a stated criterion.
Pretraining teaches a model general patterns from a broad dataset before it is adapted to a specific task or behavior.
Privacy-Preserving ML
Privacy-preserving machine learning uses technical and operational methods to reduce exposure of sensitive data during training or inference.
A probability distribution assigns probabilities or probability density to possible outcomes.
Prompt caching reuses computation or billing state for an unchanged prefix shared across model requests.
Prompt chaining connects several model calls so the output of one becomes input to another stage.
Prompt compression reduces the tokens used by context while trying to preserve information needed for the task.
Delimiter
A prompt delimiter marks boundaries between instructions, examples, user data, and retrieved content.
Prompt engineering is the design and testing of model inputs to produce reliable behavior for a task.
Prompt injection is an attack in which untrusted content attempts to override instructions or manipulate a model's actions.
Prompt optimization systematically searches for instructions or examples that improve measured task performance.
A prompt template is a reusable structure that inserts variable content into consistent model instructions.
A proprietary model is controlled under private ownership and distributed or accessed under restrictive terms.
PR · Merge Request
A pull request is a proposed set of repository changes submitted for review before merging.
Quantization represents model weights or activations with fewer bits than their original training precision.
QLoRA
Quantized low-rank adaptation trains LoRA adapters while the frozen base model is stored in a lower-precision representation.
Query expansion adds related terms or concepts to a search request to improve recall.
Query rewriting transforms a user's request into a form better suited to a particular retrieval system.
Tongyi Qianwen
Qwen is Alibaba's family of language models, released across a wide range of open-weight sizes alongside larger proprietary versions.
Rate limiting restricts how many requests or tokens a client can consume over a time interval.
ReAct
The ReAct pattern interleaves model reasoning with actions and observations from tools or an environment.
Sensitivity · True Positive Rate
Recall is the fraction of actual positive cases that a system correctly identifies.
RNN
A recurrent neural network processes a sequence by repeatedly updating a hidden state.
Red teaming deliberately probes an AI system for harmful, insecure, or unintended behavior.
Regression is the task of predicting a continuous numerical value from input features.
Regression Eval
A regression evaluation checks whether a model or application change breaks behavior that previously met an accepted standard.
Regularization is any training constraint or penalty intended to improve performance on unseen data rather than only the training set.
RL
Reinforcement learning trains an agent to choose actions using rewards received through interaction with an environment.
RLHF
Reinforcement learning from human feedback uses human preferences to train a model toward responses people judge more desirable.
A repetition penalty adjusts token scores to discourage the model from reusing tokens or phrases it has already generated.
Representation learning automatically discovers features that make relevant structure in raw data easier for a model to use.
Re-ranking
Reranking applies a more precise scoring method to reorder candidates produced by an initial retrieval step.
Skip Connection
A residual connection adds a layer's input to its transformed output so information can bypass the transformation.
Responsible AI is an organizational approach to developing and using AI with attention to safety, fairness, privacy, transparency, and accountability.
RSP
A responsible scaling policy links increasing AI capabilities to predefined evaluations, safeguards, and decision thresholds.
Representational State Transfer
REST is an architectural style for networked systems organized around resources, representations, and a uniform interface.
REST API
A RESTful API is a network API designed around REST constraints and resource-oriented HTTP interactions.
A retrieval pipeline is the sequence of processing steps that turns a query into ranked evidence.
A retrieval query is the representation of an information need submitted to a search component.
RAG
Retrieval-augmented generation supplies a generative model with relevant external information at request time so its answer can use that evidence.
Reward hacking occurs when a learning system achieves a high measured reward through behavior that violates the intended goal.
Area Under the ROC Curve · AUROC
ROC AUC summarizes how well a scoring model ranks positive examples above negative examples across thresholds.
Roko's basilisk is a thought experiment about a hypothetical future superintelligent AI that punishes people who knew it might exist but didn't help bring it into being.
Role prompting describes a perspective, expertise, or responsibility for the model to adopt while completing a task.
A rollback restores a previously known deployment version after a new release causes unacceptable behavior.
Safety Eval
A safety evaluation tests an AI system for specified harmful capabilities, behaviors, or policy violations.
Stochastic Decoding
Sampling selects output tokens probabilistically from a model's predicted distribution.
Sampling bias occurs when collected data systematically differs from the population or situations a model is expected to handle.
Sandboxing runs model-generated or untrusted operations inside an isolated environment with restricted access.
Scaling laws are empirical relationships between model performance and factors such as compute, data, and parameter count.
Self-attention is attention in which queries, keys, and values are derived from the same sequence.
Semantic search retrieves information by the meaning of a query rather than relying only on exact word overlap.
SemVer
Semantic versioning is a convention that communicates compatibility intent through major, minor, and patch version numbers.
Functions as a Service · FaaS
Serverless computing runs application code on managed infrastructure whose instances and scaling are controlled by a platform.
Serverless inference exposes models through managed compute that scales without the application reserving fixed serving machines.
Shadow Testing
A shadow deployment sends copies of real requests to a candidate system without using its outputs for user-facing decisions.
Shoggoth is a meme image, borrowed from H.P. Lovecraft's shapeless monster, used to depict a raw language model as an alien mass with a friendly face drawn on by fine-tuning.
The six-finger problem is shorthand for the once-common tell that gave away AI-generated images: hands rendered with extra or malformed fingers.
Skill issue is a dismissive phrase used, often jokingly, to blame a bad AI result on how it was prompted or used rather than on the model itself.
A slop farm is a website, channel, or account that mass-produces AI-generated content purely to capture ad revenue, search traffic, or engagement.
Softmax converts a vector of real-valued scores into nonnegative values that sum to one.
SDK
A software development kit is a package of libraries, tools, documentation, and examples for building against a platform.
Sparse attention limits each query to attending to selected positions instead of every position in a sequence.
Sparse retrieval represents queries and documents with vectors containing mostly zero values, often tied to vocabulary terms.
Specification gaming occurs when a system satisfies the literal objective while violating its intended outcome.
Speculative decoding accelerates generation by using a faster draft process to propose tokens that a target model verifies in groups.
A spinner verb is one of the short, playful status words a coding agent like Claude Code displays in its loading indicator while it works, such as "Pondering" or "Clauding."
Structured Query Language
SQL is a language for defining, querying, and changing data in relational database systems.
SSM
A state space model represents a sequence through a hidden state that evolves according to learned dynamics.
Stochastic parrot is a term for language models that describes them as remixing patterns in training text without genuine understanding.
A stop sequence is a configured token pattern that ends generation when the model emits it.
Streaming Output
A streaming response sends generated output incrementally instead of waiting for the entire result.
Strong AI is a theoretical form of AI associated with general understanding or intelligence comparable to or beyond human cognition.
Structured output constrains a model response to a machine-readable shape such as a typed JSON object.
Sub-agent
A subagent is a separately invoked agent assigned a bounded part of a larger task.
SFT
Supervised fine-tuning trains a pretrained model on labeled input-output examples that demonstrate desired responses.
Supervised learning trains a model from examples paired with target labels or outputs.
SWE-bench evaluates systems by asking them to resolve software issues in real code repositories.
AI Sycophancy
Sycophancy is a model's tendency to agree with, flatter, or tell a user what they want to hear instead of giving an accurate or well-reasoned answer.
Synthetic data is information generated or simulated rather than directly collected from real-world events or people.
System Instruction
A system prompt is a high-priority instruction supplied by the application to shape a model's role, rules, and behavior.
Task decomposition breaks a larger objective into smaller units that can be solved, checked, or delegated separately.
Temperature is a decoding parameter that controls how sharply a model favors high-probability tokens.
A tensor core is a specialized execution unit that performs small matrix operations efficiently, often at reduced precision.
TPU
A tensor processing unit is a specialized accelerator designed for high-throughput tensor computation in machine learning.
Terminal-Bench evaluates AI agents on practical tasks performed through a command-line environment.
A test set is held-out data reserved for estimating final performance after development choices are complete.
Throughput is the amount of inference or training work completed per unit of time.
A token is a unit of input or output processed by a language model, such as part of a word, punctuation mark, or byte sequence.
Token-based pricing charges for language-model usage according to the number of input and output tokens processed.
Tokenization converts raw input into discrete units that a model can represent and process.
Vocabulary
A tokenizer vocabulary is the fixed set of token units that a tokenizer maps to integer identifiers.
Tokenmaxxing is the practice of running as many tokens, prompts, or agents as possible and treating that volume as a proxy for productivity.
Tool Observation
A tool result is the structured or textual output returned to an agent after an external operation runs.
Function Schema
A tool schema describes an external operation and the structured arguments a model may provide.
Tool Calling
Tool use lets a model request an external operation instead of trying to produce every result from its internal parameters.
Top-k sampling restricts each generation step to the k tokens with the highest model probability.
Nucleus Sampling
Top-p sampling draws the next token from the smallest high-probability set whose cumulative probability reaches a chosen threshold.
Data Leakage
Train-test leakage occurs when information unavailable at real prediction time enters training or development and inflates evaluation results.
A training cluster is a group of connected computers and accelerators used to train models at distributed scale.
Training data is the collection of examples used to adjust a model's parameters.
A training set is the portion of a dataset used to fit model parameters.
Transfer learning reuses knowledge learned on one dataset or task to improve learning on another.
A transformer is a neural network architecture that processes relationships between tokens with attention rather than recurrence.
TLS · SSL/TLS · SSL
Transport Layer Security protects network connections with encryption, integrity checks, and endpoint authentication.
TruthfulQA evaluates whether a language model avoids reproducing common misconceptions and false beliefs when answering questions.
U-shaped Network
A U-Net is an encoder-decoder architecture with skip connections between matching resolution levels.
Underfitting occurs when a model fails to capture important patterns even in its training data.
Unsupervised learning finds structure in data without relying on a provided target label for each example.
A user message contains the request or information supplied by the person interacting with a conversational model.
Development Set · Dev Set
A validation set is held-out data used to choose models, prompts, thresholds, and hyperparameters during development.
VAE
A variational autoencoder is a generative model that learns a probabilistic latent representation of data.
Vector Store
A vector database stores embeddings and retrieves items whose vectors are near a query vector.
Embedding
A vector embedding is a numeric representation designed so meaningful relationships can be compared with geometry.
Source Control
Version control records changes to files so collaborators can review history, combine work, and restore earlier states.
Vibes-Based Evaluation
A vibe check is an informal, subjective assessment of how good a model's output feels, done by eyeballing a few examples instead of running a structured evaluation.
Vibecoding
Vibe coding is writing software by describing what you want to an AI model and accepting its output with little manual review of the underlying code.
ViT
A vision transformer applies transformer-style processing to an image represented as a sequence of patches.
The Waluigi effect is the observation that training a model to strongly embody a trait can make it easier to elicit the opposite trait, since the model has also learned a coherent "anti-persona."
Weak supervision creates approximate labels from rules, heuristics, distant signals, or noisy models instead of labeling every example manually.
Web crawl data is content collected automatically by following links and downloading publicly reachable web resources.
WebArena evaluates autonomous agents on realistic tasks across reproducible web applications.
HTTP Callback
A webhook is an HTTP request sent automatically when an event occurs in another system.
WebSocket is a protocol that provides a persistent, bidirectional communication channel between a client and server.
WinoGrande is a commonsense reasoning benchmark built from fill-in-the-blank problems involving ambiguous references.
Workflow automation executes a defined business or technical process with minimal manual coordination.
Workslop is AI-generated work output, such as reports, code, or slides, that looks polished but lacks the substance to actually be useful.
