Learn / Glossary
61 terms
A B C D E F G H I L M N O P Q R S T U W
Browse the alphabet or use your browser’s Find command to look up a term.
Try a shorter term, a common abbreviation or another task.
Search all tasks
A
The share of predictions that are correct across the cases in a test.
Usage and source When a rare class matters, inspect that class separately. A high overall result can hide poor detection.
Related: Recall · Google’s classification metrics course
A system in which a model can choose steps and call tools as it works towards a goal.
Usage and source For a research assistant, ask which sources it can read, which actions it may take and where it needs a person’s approval.
Related: Tool or function call · Anthropic’s guide to building agents
Describes an AI system that can take steps towards a goal. People use the word for systems with different degrees of freedom.
Usage and source An assistant comparing train routes might choose which timetable to check next. Ask which choices the software makes, which tools it can use and who can approve a booking.
Agent · Anthropic on agents and workflows
Work to identify and reduce the risk that an AI system harms people or its surroundings.
Usage and source Ask what harm was tested, in what setting, and what happens when the system fails.
Related: Guardrail · NIST AI trustworthiness framework
An agreed way for software to request data or actions from another system.
Usage and source A model becoming available through an API lets a developer call it from code. Look for separate evidence of access in a consumer app.
Related: Software development kit · AWS guide to APIs
A proposed kind of AI that could perform well across a broad range of intellectual tasks. Researchers use different tests for how broad and how capable.
Usage and source An outstanding chess score tells you about chess. To assess a claim about AGI, ask which other tasks were tested, how difficult they were and what counted as success.
Evaluation · DeepMind’s proposed levels of AGI
Software that performs tasks such as interpreting language, recognising an image or making a prediction.
Usage and source Ask what task the particular system performs and how its output will be checked.
Related: Machine learning · Google’s introduction to machine learning
B
A shared task set and scoring method used to compare systems under specified conditions.
Usage and source Compare the dataset version, tool access and scoring rules before comparing two reported scores.
Related: Evaluation · Anthropic’s evaluation guide
A release stage in which selected or public users can try a feature before its mature release. Providers define their own stages.
Usage and source Check who can join and which support or stability commitments the publisher makes.
Related: General availability · GitHub release terminology
C
OpenAI’s chat product. You interact with an app that uses models and may also provide tools.
Usage and source If an announcement mentions a new GPT model for developers, check the app’s own release notes before assuming you can select it in ChatGPT.
GPT · OpenAI’s ChatGPT introduction
Mixing details from different people, situations or rules so an answer applies them to the wrong case.
Usage and example Check which person, date, product or exception the original statement actually covers.
Try the museum case · Misattribution
The amount of information a model can handle in one request, usually measured in tokens.
Usage and source Instructions, messages and attachments can consume the budget. Check the model’s input and output limits before submitting a long file.
Related: Token · Gemini token guide
D
How information about people is collected, used, shared and retained, and the risks those activities create for them.
Usage and source Before submitting a real document, check what the specific product does with the input, who may access it and for how long.
Related: Data retention · NIST Privacy Framework
How long a service keeps particular kinds of data and the conditions under which it deletes them.
Usage and source Check the terms for your product and account. Retention, model training and human access are separate questions.
Related: Data privacy · Anthropic retention guidance
A notice that users should plan to move away from a feature or model. A separate shutdown date may follow.
Usage and source Record the announcement date and the date requests are expected to stop working as separate events.
Related: Application programming interface · Gemini API deprecation guide
E
A list of numbers that represents an item so a system can compare it with others. Similar meanings can appear close together.
Usage and source Semantic search may retrieve “annual leave” for “holiday allowance”. A close match still needs checking for relevance.
Related: Retrieval-augmented generation · Google’s embeddings course
A test that compares a system’s outputs against stated criteria on a chosen set of tasks.
Usage and source For meeting summaries, require the correct decisions and owners. Keep examples the system has never seen for a fair check.
Related: Benchmark · Anthropic’s evaluation guide
F
Presenting an invented detail, quotation or source as real.
Usage and example A complete record can establish that a supposed notice never existed. If a search simply comes up empty, mark the claim unverified while you investigate.
Try the museum case · Hallucination
Showing several examples of the task inside the prompt before asking for a new result.
Usage and source Give two short notices and their desired summaries, then ask for a summary of a third. The examples stay in the input.
Related: Fine-tuning · Gemini prompting guide
Continuing training on selected examples to adapt a model for a task or behaviour.
Usage and source Training on examples of a house style changes the model. Putting examples into one prompt is a different option.
Related: Few-shot prompting · Google’s generative AI glossary
G
A publisher’s label for a feature’s regular release on the stated product or platform.
Usage and source A generally available API may still require an eligible account, region or paid plan. Read the scope attached to the label.
Related: Rollout · Google Cloud feature stages
AI that produces new material, such as text, pictures or sound, from patterns it learned.
Usage and source A convincing sentence can contain an invented detail. Check important claims against evidence.
Related: Hallucination · Google’s generative AI glossary
A name for a family of models. Generative means they can produce content. Pre-trained describes an initial learning stage. Transformer names the underlying design.
Usage and source In a release headline, ask whether GPT names a model or whether the story is about ChatGPT, the product a person opens.
ChatGPT · OpenAI’s early generative pre-training research
Connecting a model’s response to evidence, such as a document, search result or database record.
Usage and source Check whether the cited passage supports the specific statement and whether its date fits the question.
Related: Retrieval-augmented generation · Google Cloud grounding guide
A control intended to limit unwanted behaviour, such as checking an input or restricting an action.
Usage and source A filter can miss harmful material or flag harmless material. Its existence is one part of a safety assessment.
Related: AI safety · Gemini safety settings
H
A broad term for a plausible-sounding AI output that presents false information as fact. An invented source is one example.
Usage and source Open a cited source and find the passage that supports the claim. A real link alone provides little assurance. These labels describe what the evidence shows. A claim you have yet to verify may still turn out to be true.
Related: Fabrication · Unsupported certainty · Grounding · Google’s generative AI glossary
L
A model trained on large amounts of data to work with language. A typical text generator predicts successive tokens.
Usage and source The model can compose a fluent answer to a question even when it lacks evidence for the answer.
Related: Token · Google’s language-model course
The waiting time between a request and a response. A streaming service can measure the wait to the first token or to the final answer.
Usage and source “Faster” needs a measurement: compare the same task and decide which wait matters to the user.
Related: Inference · Claude latency guide
M
A way to build software by learning patterns from examples.
Usage and source A spam filter can learn from messages people have already marked as spam. Its performance on fresh mail matters.
Related: Training · Google’s introduction to machine learning
Crediting a statement or work to the wrong person or source.
Usage and example A quotation can be word-for-word accurate while the named speaker is wrong. Match the words to the speaker in the original record.
Try the museum case · Conflation
The learned part of an AI system that turns an input into a prediction or output.
Usage and source A release may change the underlying model while the name of the app stays the same.
Related: Product or app · Google’s introduction to machine learning
A document describing a model, its intended uses, available evaluations and limitations.
Usage and source Read the tested languages, tasks and conditions before applying a reported result to a different setting.
Related: Benchmark · Google DeepMind model cards
A shared way for an AI application to connect to outside tools and information.
Usage and source A museum assistant might use an MCP connection to search its catalogue. Reading a record and editing it are different permissions. Check what the connection allows.
Tool or function call · Official MCP introduction
Able to work with more than one kind of information, such as text, pictures or sound.
Usage and source Read the input and output lists separately. A model that reads a picture may produce only text.
Related: Model · Gemini multimodal research paper
N
A model made of connected layers of simple calculations. During training, it adjusts the strength of connections to learn patterns from examples.
Usage and source To recognise handwriting, a network can learn from images labelled with the correct digit. During training, backpropagation works backwards through the layers to calculate how changes in their weights would affect the error.
Related: Training · Google’s neural networks course
O
When a model learns features of its training examples that work poorly on new examples.
Usage and source Test on independently held-out data that resembles the situations where the model will be used.
Related: Test set · Google’s overfitting guide
P
A numerical value the model learns during training and uses to transform its inputs.
Usage and source A parameter count describes scale. To judge usefulness, compare performance on the task you need.
Related: Evaluation · Google’s machine learning glossary
Further training after the initial learning stage to shape how a model behaves on tasks people care about.
Usage and source A lab might show a model example answers, then use human preferences or other signals to improve later responses. Different labs use different recipes.
Learning from human feedback · OpenAI’s InstructGPT example
The initial stage in which a model learns patterns across many examples. For a language model, one training task is predicting the next piece of text.
Usage and source A model predicting the next part of a sentence gets a signal when its prediction differs from the training text. Many such updates shape its later output.
Training · OpenAI’s generative pre-training research
In classification, the share of items flagged as positive that really are positive.
Usage and source For spam, precision asks how often a flagged message actually is spam. Numerical precision is a different use of the word.
Related: Recall · Google’s classification metrics course
The service a person uses. It may combine a model with a chat screen, search, files or other tools.
Usage and source For a model announcement, check whether the change reaches the app you use or only a developer interface.
Related: Application programming interface · Google DeepMind’s Gemini overview
The instructions and material you give a model for a particular response.
Usage and source For a summary, supply the passage, name the reader and say which facts to retain.
Related: Few-shot prompting · Gemini prompting guide
Q
Representing model numbers with a smaller set of values to reduce their storage or computation requirements.
Usage and source Check output quality and speed on your hardware. The benefit depends on the method and implementation.
Related: Precision · Hugging Face quantisation guide
R
A ceiling on the number of requests or the volume of work allowed in a specified period.
Usage and source A request may fit the context window and still exceed the service’s per-minute allowance.
Related: Context window · Gemini rate limits
A model designed to spend computation working through a task before returning an answer.
Usage and source A higher thinking setting can alter the time and resources used. Verify the final result independently.
Related: Latency · Gemini thinking guide
The share of all genuine positive cases that a system successfully finds.
Usage and source A spam filter can flag almost no legitimate mail yet miss many spam messages. Check recall alongside precision.
Related: Precision · Google’s classification metrics course
A training method in which a learner takes actions and uses a reward signal to improve its choices over time.
Usage and source A game-playing program can receive a reward for winning. The designer chooses what earns a reward, so the scoring rule matters.
RLHF · OpenAI’s introduction to reinforcement learning
A way to train a model using a reward signal informed by people’s judgements about its outputs.
Usage and source In the InstructGPT study, people ranked candidate replies. Those rankings helped train a reward model. A preferred answer can still contain a factual error.
Try ranking two replies · OpenAI’s InstructGPT research
Looking up relevant material and giving it to a model as context for an answer.
Usage and source An outdated document can still be retrieved. Check the retrieval step separately from whether the answer uses the passage faithfully.
Related: Grounding · Lewis et al., Retrieval-Augmented Generation
If you saw “RFL” in a discussion of training, check what the author spelled out. Two terms you may have been looking for are RL and RLHF.
Usage and source RL uses a reward signal. RLHF draws on human feedback to shape that signal. Choose the term that matches the description in your source.
RL · RLHF
The process of delivering a change to users. During a staged rollout, different groups receive it at different times.
Usage and source Check which country, plan, platform or account group the announcement names.
Related: General availability · Google Play staged releases
S
A collection of tools, libraries or examples for building software on a platform.
Usage and source A Python SDK may handle the request format for an API. The service still determines access and usage limits.
Related: Application programming interface · AWS guide to SDKs
T
Examples reserved to measure performance after training and model selection.
Usage and source Repeatedly choosing a model because it scores best on this same set weakens it as a final test.
Related: Evaluation · Google’s dataset guide
A unit a model uses to represent input or output. In text, a token can be part of a word, a whole word or punctuation.
Usage and source A word count is a poor substitute for a token count across languages and models. Check the relevant tokenizer.
Related: Context window · Gemini token guide
A way to charge for model use based on the tokens processed or generated. Input and output can carry different rates.
Usage and source To estimate the cost of a task, look at both token totals and any separately billed tools.
Related: Token · Claude API pricing guide
A way for a model to request a defined operation from another system, such as looking up a document.
Usage and source The surrounding application determines which operations exist and whether to perform a request.
Related: Agent · Anthropic’s guide to building agents
Adjusting the model’s internal numbers using examples and a learning objective.
Usage and source Changing wording in a single chat changes its input. Training changes the parameters used on later inputs.
Related: Inference · Google’s introduction to machine learning
A neural-network design that uses attention to learn relationships between parts of its input.
Usage and source Attention can help a model connect a pronoun with a name earlier in a sentence.
Related: Large language model · Vaswani et al., Attention Is All You Need
U
Stating an outcome as settled when the available evidence leaves it open.
Usage and example The date an inspection ends does not by itself establish when a lift will reopen. State the limit or ask the responsible person.
Try the museum case · Hallucination
W
An arrangement of steps that passes work between models, tools or people.
Usage and source A fixed route that drafts, checks and forwards an answer can be a workflow even if no model chooses the next step.
Related: Agent · Anthropic’s guide to building agents
Put a term to work Read a rollout announcement · Inspect a four-lab comparison · Explore developer releases