graphics processing unit communication bottleneck
A graphics processing unit communication bottleneck occurs when moving or synchronizing data between processors takes longer than the useful computation it supports.
AI Terminology
Search clear AI terminology definitions, acronyms, related concepts and the reviewed videos that explain them.
Showing 661–680 of 1136 terms
Clear filtersgrounded AI agent
A grounded AI agent bases its answers and actions on relevant, verifiable information rather than relying only on model memory or unsupported inference.
hard negative
A hard negative is a training example that looks relevant to an AI model but is actually the wrong match for the task.
hard red-list language model watermarking
Hard red-list language model watermarking forbids one keyed subset of tokens and allows sampling only from the complementary preferred set.
hardware-aware code optimization
Hardware-aware code optimization adapts algorithms, memory access, scheduling, and communication to the real characteristics of the target processor and system.
hardware-bound system
A hardware-bound system is software designed around a particular physical device, sensor, controller, or machine.
hardware-software co-design
Hardware-software co-design develops AI hardware and software together so each is optimized around the needs and constraints of the other.
headless application programming interface
A headless application programming interface exposes an application's data or operations independently of its graphical user interface so another program can use them directly.
held-out evaluation
Held-out evaluation checks an AI system or proposed change on examples that were not used to choose or tune that change.
heterogeneous robot team
A heterogeneous robot team combines different robot types so their distinct physical or sensing capabilities can contribute to one task.
high-stakes AI
High-stakes AI is AI used where errors can materially affect health, rights, safety, finances or other important outcomes.
historical engineering benchmark
A historical engineering benchmark tests AI on past technical problems using only the information that was available before each problem was originally solved.
hosted web application
A hosted web application runs on remote infrastructure and is accessed through a web browser over a network.
human control
Human control is the ability of people and institutions to direct, limit, supervise, or stop an AI system's actions.
human evaluation
Human evaluation asks people to judge AI outputs against criteria that may include correctness, preference, usefulness, aesthetics and intent.
human-directed AI research
Human-directed AI research keeps people responsible for hypotheses and priorities while AI systems execute bounded experimental work.
human-out-of-the-loop AI
Human-out-of-the-loop AI is a workflow in which an AI system executes defined actions without human approval at each operational step.
hybrid inference
Hybrid inference combines local AI execution with cloud models so each request can use the environment best suited to it.
hybrid retrieval
Hybrid retrieval combines different search or retrieval methods to find information that an AI workflow needs.
idempotency
Idempotency means that repeating the same operation with the same identity has the same intended effect as performing it once.