Less jargon. More understanding. Get clear definitions, see how each idea works in practice, and go deeper when you want to.
This is a guide to concepts, not a list of controls available in every model. Our model guides explain provider capabilities, app controls and integration limits.
Search terms, abbreviations and definitions. Choose a topic to narrow the list.
Showing 60 of 60 terms
Technical concepts
Related search terms: agentic · agent
Software that uses a model and tools to work through actions toward a goal.
In practice
An authorized assistant looks up order details before drafting a support response.
An agent is more than a chat response. Its tool access and permissions determine what it can actually do; important actions may need human approval.
Technical concepts
Related search terms: application programming interface
A defined way for one piece of software to request features or data from another.
In practice
A website sends a structured request to a service instead of someone filling in its interface.
API stands for application programming interface. Each service defines its accepted inputs, responses and access requirements.
AI basics
Related search terms: AI
Software that performs tasks such as recognizing patterns, understanding language or making predictions.
In practice
A tool sorts customer questions by topic before a person replies.
AI is an umbrella term. Not every AI system generates content, and useful performance does not imply human understanding.
Video
Related search terms: portrait · landscape
The relationship between a picture’s width and height.
In practice
A 16:9 frame is wide, while a 9:16 frame is tall.
Aspect ratio describes shape, not pixel count. Two files can share a ratio while having very different resolutions.
Responsible use
Related search terms: fairness · AI bias
A systematic tendency in data, decisions or outputs that can produce uneven or unfair results.
In practice
Generated job descriptions repeatedly associate certain roles with one group of people.
Bias can arise from data, design choices and how a system is used. Reviewing varied examples is more informative than checking just one output.
Music & audio
Related search terms: 16-bit · 24-bit
For digital audio, the number of bits used to represent each sample’s amplitude.
In practice
A 24-bit recording uses more possible amplitude values than a 16-bit recording.
Bit depth affects quantization precision and available dynamic range. It is distinct from sample rate and compressed-file bitrate.
Responsible use
Related search terms: C2PA · Content Credentials
Information about where a digital asset came from and how it has been changed.
In practice
A content credential records creation and editing information attached to an image.
Signed provenance can make recorded history tamper-evident. It does not prove that the depicted event or every claim in the content is true.
Chat & writing
Related search terms: context length
The amount of information a model can consider in a single request, usually measured in tokens.
In practice
A long document, your question and conversation history all take up context.
Input and output limits vary by model. A large window is not permanent memory or a guarantee that every detail will be used correctly.
Images
Related search terms: inference steps · sampling steps
The number of refinement steps used by a diffusion generation process.
In practice
A model designed for a few steps may generate quickly without needing dozens of passes.
More steps can take longer without improving the result. Step counts should be interpreted within the same model and sampling method.
Images
Related search terms: diffusion model · latent
A family of generation methods that learns to turn noisy data into structured content.
In practice
An image generator gradually refines a noisy starting representation into a scene.
Some systems work in a compressed latent representation rather than directly on pixels. Not all image-generation models use identical methods.
Technical concepts
Related search terms: vector · semantic search
A numerical representation that helps software compare the meaning or features of information.
In practice
A search system can match “car repair” with documents about vehicle maintenance.
Embeddings are vectors. Similarity in that space can help retrieval, but a nearby match is not proof that two statements are equivalent or true.
Chat & writing
Related search terms: few shot · in-context learning
Including a few examples that demonstrate the output pattern you want.
In practice
Show two product descriptions in your brand voice, then ask for a third.
The examples consume context and influence the current response. This is not the same as fine-tuning the model.
Technical concepts
Related search terms: finetuning · tuning
Additional training that adapts an existing model using a more specific set of examples.
In practice
A developer trains a base model further on examples of a particular response format.
Fine-tuning changes learned parameters. It differs from including examples in a prompt or retrieving documents for a single request.
Video
Related search terms: FPS · frames per second
The number of video frames shown each second, expressed as frames per second.
In practice
A 24 FPS clip shows 24 frames during each second of playback.
Frame rate affects motion presentation, not clip length or resolution. Converting frame rates can duplicate, remove or synthesize frames.
Technical concepts
Related search terms: tool calling · tool use
A model producing a structured request for an application to use a tool or function.
In practice
The model requests a weather lookup with a city name instead of inventing the forecast.
The surrounding application executes permitted calls and returns results. A model-generated tool request is not itself authorization to perform an action.
Technical concepts
Related search terms: API gateway · upstream
An intermediary service that receives an application’s request and connects it to a backend service.
In practice
An app reaches a model through a gateway rather than calling the model developer directly.
A gateway can expose its own API, limits and supported parameters. Direct-provider documentation may not describe that gateway’s exact behavior.
AI basics
Related search terms: GenAI
AI that produces content, including text, images, video or audio, from an input.
In practice
Describe a quiet mountain scene and a model creates a new illustration.
Generation draws on learned patterns. A newly generated result is not automatically accurate, unique or cleared for every use.
Chat & writing
Connecting a model’s response to supplied documents, retrieved records or other reference material.
In practice
Ask for a summary based on your actual product documentation rather than general model knowledge.
Grounding can improve relevance and traceability. It does not guarantee correctness; check that cited evidence really supports each claim.
Images
Related search terms: CFG · classifier-free guidance
A setting in some generators that changes how strongly generation is steered toward the prompt.
In practice
Compare supported settings to balance prompt adherence with a more natural-looking result.
Higher is not always better: excessive guidance can introduce artifacts. The parameter’s meaning and useful range depend on the model.
Responsible use
Related search terms: confabulation
An AI-generated statement that appears plausible but is unsupported or incorrect.
In practice
An answer cites a research paper that does not actually exist.
Confident language is not evidence. Verify important claims against reliable sources, especially when decisions depend on them.
Responsible use
Related search terms: human in the loop · HITL
A person checking an AI result before it is relied on, published or used to take action.
In practice
Review a generated customer email for correct facts and an appropriate tone before sending it.
Review should match the consequences of error. Checking wording alone is insufficient when factual accuracy, safety or permissions matter.
Images
Related search terms: img2img · I2I
Creating a new image using an existing image as part of the input.
In practice
Use a sketch as a starting point for a more polished concept illustration.
How closely the output follows the original depends on the model and supported settings. It is not a guarantee of exact object preservation.
Video
Related search terms: I2V · image to video
Generating a video with a still image as a visual starting point or reference.
In practice
Animate a landscape photo with drifting clouds and a subtle camera move.
Reference handling varies. A model may use a first frame, reference image or other conditioning rather than preserving every pixel throughout.
AI basics
Using an already-trained model to produce a result from a new input.
In practice
When you ask for three headline ideas, producing the answer is inference.
Inference uses learned model parameters. Additional retrieval or tool calls may happen around that process without retraining the model.
Images
Related search terms: mask · masked editing
Generating or replacing content within a selected area of an image.
In practice
Select an unwanted object and ask for background that fits the surrounding scene.
A mask commonly identifies the edit area. Mask support and preservation outside that area depend on the implementation.
Music & audio
Related search terms: no vocals · backing track
Music without a sung lyrical part, built around instruments or other musical sounds.
In practice
A piano-and-strings backing track leaves room for your spoken narration.
Requesting an instrumental in prose does not guarantee a vocal-free result. Dedicated controls and model behavior vary; listen before using the track.
Video
Related search terms: first frame · last frame
A frame or point in time used to define an important visual state in a sequence.
In practice
Where supported, a starting image and ending image guide the transition between two scenes.
In editing, keyframes can set properties such as position. AI start/end-frame controls are a related but model-specific generation feature.
AI basics
Related search terms: LLM · language model
A model trained on large amounts of language data that can interpret and generate text.
In practice
An LLM can turn rough meeting notes into a readable summary.
Many language models predict successive tokens. Some also accept images or other inputs; those capabilities depend on the particular model.
Technical concepts
Related search terms: response time · time to first token
The delay between making a request and receiving a response or part of one.
In practice
A reply starts appearing quickly but takes longer to finish.
Time to first output and time to completion measure different things. Network conditions, queues, tools and model computation all contribute.
Music & audio
Related search terms: looping · seamless loop
A section of audio or music played repeatedly.
In practice
Repeat a short drum pattern underneath a longer presentation.
A seamless loop needs its ending and beginning to join cleanly. Generated audio may need trimming or a crossfade before it repeats smoothly.
Music & audio
Related search terms: custom lyrics · vocals
The words intended to be sung or spoken as part of a song.
In practice
A chorus contains the repeated words you want a vocalist to sing.
Lyrics and a style description serve different purposes. Check each music guide: the app may send the main prompt as lyrics rather than provide a separate input.
AI basics
Related search terms: ML
A way to build software that learns patterns from examples instead of relying only on hand-written rules.
In practice
A spam filter learns from messages labeled as spam or legitimate email.
Training adjusts a model using data. How well it works on new examples matters more than memorizing its training examples.
AI basics
The trained system that turns an input into a prediction or generated result.
In practice
You choose one model for drafting an email and another for creating an image.
A model is not the whole application. The surrounding product supplies its interface, tools, safeguards and connection to the model.
AI basics
Related search terms: modality · vision
Able to work with more than one kind of information, such as text and images.
In practice
Ask a question about a chart by providing both the chart image and your question.
Input and output support are separate. A model that understands an image does not necessarily generate images or accept video.
Images
A separate instruction used by some models to discourage unwanted visual features.
In practice
Where supported, a negative prompt might mention blur or unwanted text.
Not every model accepts a negative-prompt parameter. Putting negative wording in a regular prompt is not necessarily equivalent.
Images
Related search terms: generative expand
Generating new content beyond the original edges of an image.
In practice
Extend a square landscape into a wide banner by adding scenery at the sides.
Unlike simple resizing, outpainting invents additional visual content. New details should be reviewed for continuity with the original.
Responsible use
Related search terms: PII · personally identifiable information · privacy
Information that identifies a person or can be linked to them.
In practice
A customer’s name, email address and order history should not be copied into an unrelated prompt.
Remove unnecessary identifiers and check the applicable service’s privacy and retention terms before sharing sensitive material.
Chat & writing
Related search terms: input
The instructions and context you give an AI model to guide its response.
In practice
Write a friendly welcome email for new customers in fewer than 100 words.
Prompts can include examples, documents or media when supported. Be clear about the task, audience, constraints and desired output.
Chat & writing
Related search terms: prompt design
Writing, testing and refining instructions to get more useful responses from a model.
In practice
Replace “write an ad” with an audience, product benefit, tone and length limit.
Good prompts make requirements explicit. Prompt changes do not alter model weights or unlock unsupported tools and controls.
Responsible use
An attempt to make an AI system follow unwanted instructions hidden in its inputs or reference material.
In practice
A retrieved web page tells an assistant to ignore its task and disclose private information.
Treat external content as information, not authority. Applications need permission boundaries and tool safeguards beyond prompt wording alone.
Technical concepts
Related search terms: model provider · inference host
The organization or service that develops a model or makes it available to use.
In practice
A model developer and the gateway serving that model may be different companies.
Check whether a guide means the model creator, inference host or intermediary. A familiar model name alone does not identify its delivery route.
Chat & writing
Related search terms: reasoning model
Working through intermediate steps to solve a problem rather than giving only an immediate answer.
In practice
Compare delivery options against a budget, deadline and package size.
Reasoning-focused models may spend additional computation on a task. An explanation that sounds logical still needs its assumptions and result checked.
Video
Related search terms: pixels · 1080p · 4K
The pixel dimensions of an image or video frame.
In practice
A 1920 × 1080 video frame is 1,920 pixels wide and 1,080 pixels tall.
More pixels do not automatically mean sharper source detail. Compression, focus, generation quality and upscaling also affect appearance.
Technical concepts
Related search terms: RAG
Looking up relevant information and providing it to a model to help produce an answer.
In practice
Retrieve the relevant help-center article before answering a product question.
RAG adds context at request time rather than retraining the model. Retrieval quality and faithful use of the retrieved material still need evaluation.
Music & audio
Related search terms: kHz · Hz · sampling rate
How many samples of an audio signal are represented per second.
In practice
48 kHz audio contains 48,000 samples per second for each channel.
Sample rate is not bitrate or musical tempo. Converting to a higher rate does not restore detail missing from the original recording.
Images
Related search terms: random seed
A number used to initialize the random process behind a generation.
In practice
Reuse a supported seed while changing one prompt detail to make comparisons easier.
A seed alone cannot guarantee the same result. Model version, other parameters, hardware and implementation can all affect reproducibility.
Music & audio
Related search terms: stem separation · multitrack
Separate audio parts or groups, such as vocals, drums and bass, used to build a mix.
In practice
Lower the drum stem while keeping the vocal at its original volume.
Stems exported from a project differ from parts estimated by AI separation. Separation can leave artifacts, and a finished audio file does not imply stem access.
Technical concepts
Delivering a response in pieces as they become available rather than waiting for everything to finish.
In practice
An answer appears a few words at a time while the model continues generating.
Server-sent events are one way to transport incremental updates. Streaming can improve perceived responsiveness without reducing total generation time.
Chat & writing
Related search terms: system instructions
Instructions supplied by an application to establish how its assistant should behave.
In practice
A support assistant is instructed to answer concisely and avoid inventing refund policies.
System instructions are separate from an ordinary user message. They guide behavior but are not a substitute for enforced permissions or security checks.
Chat & writing
A generation setting that changes how much randomness is used when choosing output tokens.
In practice
A supported lower setting may give less varied wording across repeated drafts.
Temperature is not a truthfulness slider. Supported ranges differ, and a low value does not guarantee identical or factually correct answers.
Music & audio
Related search terms: BPM · beats per minute
The pace of music, commonly described in beats per minute.
In practice
A steady 90 BPM beat has 90 beats in one minute.
Tempo and mood are different. A BPM value in a generation prompt is guidance unless the model and integration enforce it.
Images
Related search terms: text to image · T2I
Creating an image from a written description.
In practice
A ceramic mug on a pale blue table, soft morning light, editorial product photograph.
Text guides composition and style, but exact lettering, counts and fine details can require revision. Supported settings vary by model.
Music & audio
Related search terms: music generation
Generating music from a written description of its sound, mood or style.
In practice
Gentle acoustic guitar, warm piano and a relaxed mood for a product walkthrough.
Lyrics, duration and instrumental controls are integration-specific. A provider’s full feature set is not necessarily available through the app.
Music & audio
Related search terms: TTS · speech synthesis
Converting written words into generated spoken audio.
In practice
Turn an approved tutorial script into a narrated voiceover.
Speech synthesis is different from music generation or speech transcription. Voice options, pronunciation controls and permissions depend on the service.
Video
Related search terms: T2V · text to video
Generating a moving clip from a written description of a scene or action.
In practice
A slow camera move across a sunlit workshop as dust floats through the light.
Describe the subject, action and camera movement. Audio, duration and other capabilities depend on the model and the application’s integration.
Chat & writing
Related search terms: tokens · tokenization
A small unit a model processes, such as a word fragment, punctuation mark or representation of other media.
In practice
A long word may use several tokens rather than just one.
Token counts depend on the tokenizer, language and media type. A word count is only a rough guide to context use or token-based pricing.
Chat & writing
Related search terms: nucleus sampling · top p
A setting that limits token choices to a group whose combined probability reaches a chosen threshold.
In practice
A lower supported top-p can narrow the set of words considered for the next token.
Also called nucleus sampling. It works differently from temperature; values and support are model-specific, not universal creative presets.
AI basics
The process of adjusting a model using examples so it learns patterns useful for a task.
In practice
A language model learns patterns from a large collection of text before people use it.
Training and using a model are different stages. Sending a prompt is not, by itself, a model-training operation.
Images
Related search terms: super-resolution · upscale
Increasing an image’s pixel dimensions, sometimes using AI to estimate additional detail.
In practice
Enlarge a small illustration for a larger web layout.
AI upscaling can invent textures or alter details; it cannot reliably recover information absent from the source image.
Chat & writing
Related search terms: zero shot
Asking a model to do a task without showing examples of the desired answer.
In practice
Classify this review as positive, neutral or negative.
Zero-shot describes the prompt, not whether the model encountered similar tasks during training. Clear instructions still matter.