Glossary
Plain definitions of the terms you meet in AI news. Terms in our stories link here.
5
- 510(k)
- The FDA's most common route for medical devices, which clears a device shown to be substantially equivalent to one already on the market.
A
- AGI
- Artificial general intelligence, a loosely defined goal of AI that matches or exceeds people across most cognitive tasks.
- AI Act
- The European Union's law regulating AI systems according to their level of risk.
- AI agent
- An AI system that takes actions toward a goal, such as using tools, browsing, or running code, with limited step-by-step human direction.
- AI alignment
- The work of making AI systems pursue the goals and values their developers and users intend.
- Algorithmic impact assessment
- A structured review of an automated system's possible effects on people, done before or while it is used.
- Algorithmic management
- Using software, including AI, to assign, monitor, evaluate, or discipline workers.
- Annual recurring revenue
- A company's yearly revenue from ongoing subscriptions or contracts, usually projected from a recent month.
- API
- An application programming interface, a way for software to send requests to a service, such as an AI model, and receive results.
- ASIC
- An application-specific integrated circuit, a chip designed for one kind of task, such as running AI models.
B
- Benchmark
- A standard test used to measure and compare AI models on a set task.
- Benchmark contamination
- When test questions, or their answers, have leaked into a model's training data, inflating its score.
- Benchmark saturation
- When top models score so close to the maximum on a test that it no longer separates them.
C
- Capacity vs. consumption (MW vs. MWh)
- Megawatts measure how much power can be supplied or drawn at a moment; megawatt-hours measure energy used over time.
- Chatbot
- A program that holds a text or voice conversation with users, now usually powered by a large language model.
- Circular deal
- An arrangement in which a supplier invests in or lends to a customer that then spends the money on the supplier's products.
- Class action
- A lawsuit in which one or a few plaintiffs sue on behalf of a larger group with similar claims.
- Class deviation
- An authorized departure from standard federal purchasing rules that applies to a whole category of contracts.
- Clinical validation
- Evidence that a medical tool's output is accurate and meaningful for the patients and setting it is intended for.
- Compute (AI)
- The processing power used to train and run AI models, usually supplied by specialized chips in data centers.
- Compute deal
- An agreement to buy, rent, or supply computing capacity for AI, often over many years.
- Context window
- The maximum amount of text, measured in tokens, that a model can consider at once.
- Curtailment
- Cutting back how much electricity a customer draws, or how much a power plant produces, usually to protect the grid.
D
- Data center
- A building or campus housing servers, storage, and networking equipment, along with the power and cooling they need.
- De Novo
- An FDA pathway for new kinds of low- to moderate-risk medical devices that have no existing equivalent on the market.
- Discovery (litigation)
- The pretrial stage of a lawsuit in which each side must hand over relevant documents and information to the other.
- Disengagement
- A moment when an autonomous system stops controlling a vehicle or robot and a human takes over.
- Distillation
- Training a smaller model to imitate a larger one, so it is cheaper to run.
E
- Eval
- A structured test of what an AI model can do or how it behaves, including tests for risks.
- Executive order
- A written directive from the US president telling federal agencies how to carry out their work.
- Exposure (AI and jobs)
- A measure of how many of a job's tasks AI could in principle perform or speed up, not a forecast of job losses.
F
- Fair use
- A US legal doctrine that allows some uses of copyrighted work without permission, decided case by case.
- FASCSA
- The Federal Acquisition Supply Chain Security Act of 2018, a US law that lets officials bar risky technology from federal systems.
- Feasibility study
- An early, limited study of whether a plan can practically be carried out, before a larger commitment.
- Fine-tuning
- Additional training of an existing model on a smaller, focused dataset to adapt it to a task.
- FLOPs
- Floating-point operations, a count of the basic arithmetic steps a computer performs, used to measure AI training effort.
- Formal proof
- A mathematical proof written in a precise computer language so that software can check every step.
- Foundation model
- A large, general model trained on broad data that can be adapted to many different tasks.
- Frontier model
- A model at or near the most capable level available at a given time.
- Frontier safety framework
- A developer's published policy for testing models for dangerous capabilities and the safeguards it will apply at each level.
- Funding round (Series A, B, C...)
- A stage of venture investment in a private company; later letters usually mean a more mature company raising larger sums.
G
- General-purpose AI model
- Under the EU AI Act, a model that can competently perform a wide range of tasks and be built into many different systems.
- Generative AI
- AI that produces new content, such as text, images, audio, video, or code.
- GPU
- A graphics processing unit, a chip that performs many calculations in parallel and is widely used to train and run AI.
- Grid interconnection
- The process and approval needed to connect a power plant or a large electricity user to the grid.
H
- Hallucination
- When an AI model states something false or made up as if it were true.
- HBM
- High-bandwidth memory, stacked memory chips packaged next to an AI processor so data can move to it very quickly.
- High-risk AI system
- Under the EU AI Act, an AI system used in an area where errors could seriously affect health, safety, or rights, which must meet strict requirements.
- Humanoid robot
- A robot with a roughly human body shape, typically with two arms and two legs, designed to work in human spaces.
I
- Inference
- Running a trained model to produce an answer or output.
J
- Job displacement
- Workers losing their jobs or job tasks because the work is automated or reorganized.
L
- Large language model
- An AI system trained on very large amounts of text to predict and generate language.
- Large load
- A single customer, such as a data center, that draws an unusually large amount of power from the grid.
M
- Model Context Protocol
- An open standard for connecting AI applications to outside tools and data sources.
- Multimodal
- Able to take in or produce more than one kind of data, such as text, images, audio, and video.
O
- Open-source AI
- AI released under terms that allow anyone to use, study, change, and share it, usually including more than the weights.
- Open-weight model
- A model whose trained parameters are published so anyone can download and run it.
P
- Parameters (AI models)
- The numerical values inside a model that are adjusted during training and determine its behavior.
- Peer review
- Evaluation of a research paper by independent experts before a journal or conference publishes it.
- Pre-money and post-money valuation
- A company's value before new investment is added (pre-money) and after it (post-money).
- Preemption
- The legal principle that a higher level of government's law can override or block a lower level's law on the same subject.
- Preprint
- A research paper shared publicly before it has been peer reviewed.
- Procurement
- How governments and other organizations buy goods and services, including AI tools, under formal rules.
- Prompt injection
- An attack that hides instructions in content an AI system reads, to make it act against its user's intent.
- Prospective study
- A study that sets its plan in advance and then follows people or cases forward in time.
- Public benefit corporation
- A for-profit company whose directors are required to weigh a stated public benefit alongside shareholders' interests.
- PUE
- Power usage effectiveness, the ratio of a data center's total energy use to the energy used by its computing equipment.
R
- Randomized controlled trial
- A study that assigns participants by chance to different groups, so differences in results can be attributed to the intervention.
- Reasoning model
- A model trained to work through a problem in intermediate steps before giving a final answer.
- Red teaming
- Deliberately trying to make an AI system fail or misbehave in order to find weaknesses before others do.
- Retrieval-augmented generation
- A technique in which a model first fetches relevant documents and then uses them to write its answer.
- Retrospective study
- A study that looks back at data already collected, such as past medical records or scans.
- RLHF
- Reinforcement learning from human feedback, a training method that uses human ratings to shape model behavior.
- Robotaxi
- A taxi service in which a self-driving car carries passengers without a human driver in control.
- Rulemaking
- The formal process US agencies use to write binding regulations, usually including a public comment period.
- Run-rate
- A projection of yearly revenue made by extending a recent short period, such as a month or quarter, across a full year.
S
- Safety monitor
- A company employee who rides in a self-driving vehicle to watch it and intervene if needed.
- Shadow library
- An online collection that distributes copyrighted books or papers without permission from rights holders.
- Small modular reactor
- A nuclear reactor design that is smaller than a conventional plant and meant to be built largely from factory-made parts.
- Summary judgment
- A court ruling that decides all or part of a case without a trial, because the key facts are not genuinely in dispute.
- System card
- A document a developer publishes with a model describing its intended uses, test results, risks, and safeguards.
T
- Teleoperation
- Controlling or guiding a robot or vehicle from a distance by a human operator.
- Token (AI)
- A small chunk of text, such as a word or part of a word, that a language model reads or writes.
- TPU
- A Tensor Processing Unit, Google's custom chip designed for AI workloads.
- Training (AI models)
- The process of building a model by adjusting its parameters on large amounts of data.
- Training data
- The text, images, or other material used to teach an AI model.
U
- Uprate
- An upgrade that raises the maximum power output of an existing power plant, especially a nuclear plant.
V
- Valuation
- The price investors put on a whole company, usually implied by what they pay for a slice of it.
W
- WARN Act
- A US law requiring many large employers to give advance notice of plant closings and mass layoffs.