The Tool Desk
Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
AI vocabulary can blur together: machine learning is part of AI, generative AI is one kind of system, and an LLM is not the same thing as every foundation model. This glossary gives 63 useful terms in plain language, grouped by how the concepts fit together. The selection is practical rather than canonical; terminology can vary across organizations and change as technology develops.
AI foundations
1. Artificial intelligence (AI)
A broad field focused on building computer systems that perform tasks associated with human capabilities, such as recognizing patterns, understanding language, or making decisions. A system that classifies photos is AI even if it does not create new images. Machine learning is one approach within AI.
2. Machine learning (ML)
A way to build AI systems by having them learn patterns from data rather than writing every decision as a fixed rule. A spam filter trained on labeled messages is a machine-learning system. ML is within the broader field of AI.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
3. Deep learning
A form of machine learning that uses neural networks with many layers to learn complex patterns. It is commonly used for speech, image, and language tasks. Not every machine-learning method is deep learning.
#1 Best Overall
4. Neural network
A machine-learning model made of interconnected computational units that transform input data to produce an output. A network may learn to identify objects in photos by adjusting its internal parameters during training. Deep learning uses neural networks with many layers.
5. Algorithm
A defined procedure for performing a computation or solving a problem. An algorithm might sort a list of numbers or specify how a model updates its parameters. An algorithm is a method; a model is the learned or specified system that applies a method to inputs.
6. Model
A system that has learned patterns from data, or that otherwise encodes rules for producing outputs from inputs. A model might predict whether a transaction is fraudulent. A model is not the same as the full AI application around it, which can include data sources, interfaces, and safeguards.
7. Dataset
A collection of data used to build, test, or operate a system. A dataset for recognizing plants might contain photographs and labels naming each plant. The data’s coverage and quality affect what a model can learn.
8. Feature
An input characteristic used by a model, such as a customer’s purchase history or the pixel values in an image. In a traditional prediction system, people may select features explicitly; other systems learn useful representations from raw inputs.
9. Label
The target answer associated with an example in supervised learning, such as “spam” for an email or “oak” for a tree image. Labels help a model learn which outputs correspond to which examples. A label is not the same as the model’s prediction.
10. Supervised learning
A machine-learning approach in which examples include target labels, and the model learns to predict those labels. Training on past house sales with known prices to estimate a new listing’s price is an example.
11. Unsupervised learning
A machine-learning approach that looks for patterns in data without target labels supplied for each example. It can group similar customer records, for instance. A discovered group is a pattern in the data, not automatically a meaningful category.
12. Reinforcement learning
A learning approach in which a system takes actions and receives rewards or penalties, using that feedback to improve its choices over time. A simulated robot might learn to navigate by receiving a reward for reaching its destination.
13. Classification
A task that assigns an input to one or more categories, such as deciding whether a message is spam. Classification predicts a class; it does not necessarily generate a new passage, image, or other content.
14. Prediction
An estimate of an unknown or future value based on available inputs. A model might predict next month’s energy use from earlier readings. Predictions can be uncertain and should not be treated as known facts.
Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchPC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 1115. Generative AI
AI that generates new content, such as text, images, audio, video, or code. A text generator can draft a product description from a prompt. This differs from a classifier that only selects a label, even though both may use machine learning.
Rank #2
Models, data, and operation
16. Foundation model
A model trained on broad data that can be adapted or used for a range of tasks. A foundation model may handle text, images, or multiple modalities. The term describes a broad model category, while an LLM specifically focuses on language.
17. Large language model (LLM)
A model trained to process and generate language, often by predicting likely next tokens in context. It can draft an email or answer a question, but fluent wording does not prove its claims are correct. An LLM is a type of model; foundation models can also cover other modalities.
18. Multimodal
Describes a system that can handle more than one kind of input or output, such as text and images. A multimodal assistant might answer a question about a picture. The word describes modalities supported, not whether the system is accurate.
19. Training
The process of adjusting a model using data so it learns patterns or relationships. During training, a language model may repeatedly update its parameters to improve its predictions. Training differs from inference, when the model is used to produce an output.
20. Inference
The process of running a trained model on inputs to produce an output. Asking a language model to summarize a paragraph is inference. It is distinct from training, which changes model parameters based on data.
21. Parameter
A value inside a model that is adjusted during training and influences its outputs. Parameters are part of the model’s learned internal structure; they are not the same as the prompt a person supplies at runtime.
22. Fine-tuning
Additional training that adapts an already trained model to a narrower task, style, or data domain. A company might fine-tune a model on examples of its preferred support responses. Fine-tuning changes model behavior through training; it is different from supplying reference documents in a prompt.
23. Pretraining
An initial, broad training stage that helps a model learn general patterns from a large collection of data. Later training or adaptation may specialize it. Pretraining is one stage in a model’s development, not the same as using the model for inference.
24. Token
A unit of text a language model processes. A token can be a whole word, part of a word, punctuation, or another text unit, so token counts do not map one-to-one to word counts. Token is a processing unit, not a synonym for word.
25. Tokenization
The process of breaking text into tokens before a model processes it. A long or uncommon word may be split into multiple pieces. Tokenization explains why the same-length passages can use different numbers of tokens.
26. Context window
The amount of text or other input a model can consider at one time, measured in tokens for language models. A longer context window can accommodate more material in a single interaction, but it is not automatically persistent memory across separate sessions.
Do these 3 things before closing this tab:
1Fix the driver behind crashes, sound loss and screen glitches2Clear out junk files and repair common Windows errors3Scan for outdated or missing drivers - takes under a minute27. Prompt
The input or instruction provided to a generative model to guide its output. “Explain photosynthesis to a 10-year-old in three sentences” is a prompt. A prompt supplies immediate context; it does not necessarily change the underlying model.
28. Prompt engineering
The practice of writing and organizing prompts to elicit more useful model responses. Specifying the audience, task, constraints, and desired format can reduce ambiguity. Better wording can improve an answer, but it cannot guarantee truth.
29. System prompt
An instruction supplied to shape how an AI application behaves across an interaction, often before a user’s message. It may define the assistant’s role or response constraints. It is part of the input context, not a substitute for model training or safety evaluation.
30. Temperature
A setting used by some generative models to control how varied their outputs are. Lower values generally favor more predictable choices, while higher values can produce more variety. The setting does not make a response more factually reliable.
Free tools Windows power users keep installed
One-click scans. No signup required.
31. Embedding
A numerical representation of data that can capture relationships among items. Texts with similar meaning may have embeddings that are close together in a vector space, helping a retrieval system find relevant passages even when they use different wording.
32. Vector
An ordered set of numbers that can represent an item or its features. An embedding is a kind of vector representation. Vector similarity can help locate related items, but similarity alone does not establish that two passages make identical claims.
33. Vector database
A database designed to store and search vector representations, often by similarity. A question-answering system might use it to locate passages whose embeddings resemble a question’s embedding. It is one possible component of retrieval, not a generative model.
34. Semantic search
Search that aims to find material related in meaning, rather than matching only exact words. A search for “ways to reduce a power bill” might find a passage about lowering electricity costs. Embeddings are one technique that can support semantic search.
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Clear out junk files and repair common Windows errorsFree Scan →35. Overfitting
A training problem in which a model learns details of its training examples too closely and performs poorly on new data. A model that memorizes training cases but fails on unfamiliar ones is overfit. Evaluation on separate data can help reveal this problem.
36. Generalization
A model’s ability to perform well on examples beyond those it was trained on. A handwriting recognizer that works on new writers’ samples has generalized beyond its training examples. Strong performance on training data alone does not establish generalization.
37. Evaluation
The process of measuring how well a model or AI system performs against specified tasks or criteria. Tests might assess accuracy on a set of labeled examples or the usefulness of generated answers. Results depend on what was tested and do not prove performance in every real-world setting.
Generative AI outputs and techniques
38. Text generation
The creation of written output by a generative model, such as drafting a summary or completing a sentence. A text generator predicts a sequence of tokens; the resulting prose may sound confident even when it needs fact-checking.
39. Image generation
The creation of an image from an input, often a text description or another image. A user might request an illustration of a red bicycle beside a tree. The description guides the output but does not guarantee every detail will appear correctly.
40. Diffusion model
A type of generative model that learns to create data by progressively transforming noise into a structured output. Many image generators use a diffusion process to produce pictures from prompts. It is one approach to generation, not a synonym for all generative AI.
41. Transformer
A neural-network architecture widely used in language models and other AI systems. It uses attention mechanisms to process relationships among parts of an input. “Transformer” names an architecture, while “LLM” names a language-focused model category.
42. Attention
A mechanism that helps a model weigh relationships among parts of its input when producing an output. In a sentence, attention can help a model relate a pronoun to an earlier noun. The term does not mean the model consciously pays attention.
Recommended Free Tools
43. Hallucination
An AI-generated claim or detail that is false, unsupported, or made up, despite being presented plausibly. A chatbot might invent a book citation. Hallucinations are a reason to verify consequential claims rather than relying on fluent presentation.
44. Grounding
Connecting an AI response to relevant information, such as supplied documents or retrieved sources. An answer grounded in a company policy document can be more relevant to that policy. Grounding can support an answer but does not guarantee that it is correct or that the source itself is accurate.
45. Retrieval-augmented generation (RAG)
A technique that combines information retrieval with generative AI: a system retrieves relevant material, adds it to the model’s prompt as context, and generates a response based on that augmented input. For example, a support assistant can retrieve passages from a help center before drafting an answer. RAG changes the information available at response time; it does not guarantee truth. Google Cloud’s RAG overview describes this workflow.
46. Fine-tuning versus RAG
These are different ways to adapt an AI application. Fine-tuning trains a model further; RAG retrieves material to provide context at response time. A team updating a frequently changing handbook may use retrieval so answers can draw on current documents, whereas fine-tuning is an adaptation of the model itself.
Windows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallCrashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minute47. Ground truth
The best available reference answer or measured value used to assess a prediction. In a test set, human-verified labels may serve as ground truth. It is a benchmark for comparison, not a claim that every label is infallible.
48. Synthetic data
Data generated artificially rather than collected directly from real-world events or people. Synthetic examples can help test a system or supplement training data. Their usefulness depends on how well they represent the situations the system will encounter.
49. Data augmentation
Creating modified versions of existing training examples to increase variety, such as slightly changing an image’s orientation. Augmentation can help a model handle variations, but it does not automatically correct gaps or bias in the original data.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Retrieval, tools, and AI applications
50. Retrieval
Finding relevant information from a collection, such as searching documents for passages related to a question. Retrieval supplies material; generation uses a model to compose new output. RAG connects these two functions.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
51. Knowledge base
An organized collection of information an application can consult, such as product manuals or internal policies. A knowledge base can be searched by people or connected to an AI retrieval system. Its usefulness depends on the accuracy and currency of its contents.
Best Value
52. AI agent
A system that can pursue a task by selecting actions, potentially using tools or interacting with other systems. An agent might search for information and then prepare a report. The label covers varied designs; it does not by itself indicate autonomy, reliability, or safety.
53. Tool calling
A model’s ability to request that an application run an available tool, such as a calculator or search function. The application executes the tool and may return its result to the model. A model requesting a tool is not the same as the model directly controlling every part of the application.
54. Function calling
A structured form of tool calling in which a model produces arguments for a defined function, such as a request to look up a delivery status. The surrounding software typically decides whether and how to run it. The function’s output may inform a later model response.
The Tool Desk
Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →55. API
An application programming interface: a defined way for software systems to request services or exchange data. An AI application might use an API to send a prompt to a model. An API is an integration mechanism, not the model itself.
56. Chatbot
A software application that interacts with users through conversation. A chatbot may rely on scripted rules, retrieval, a language model, or a combination. The interface alone does not tell you which underlying technology it uses.
57. AI workflow
A sequence of steps that combines AI and other software to complete a task. A workflow might accept a support request, retrieve a policy passage, draft a response, and route it for human review. It describes the process around a model, not just the model’s generation step.
Responsible and trustworthy AI
58. Bias
A systematic skew in data, model outputs, or decisions that can disadvantage some people or cases. A hiring model trained on unrepresentative historical records may reproduce those patterns. Detecting bias requires examining the system’s context and effects, not only its overall accuracy.
Free tools Windows power users keep installed
One-click scans. No signup required.
59. Fairness
A goal of treating people or groups justly in an AI system’s decisions and effects. Fairness can involve competing criteria, and what is appropriate depends on the use case and applicable rules. A single aggregate score cannot settle every fairness question.
60. Explainability
The extent to which people can understand why an AI system produced an output or how it works. A simple decision tree may offer a more direct explanation than a large neural network. An explanation should be useful and faithful to the system, not merely persuasive.
61. Transparency
Making relevant information about an AI system visible, such as its purpose, limitations, data practices, or role in a decision. Transparency helps people understand the system’s use; it is distinct from being able to explain every internal computation.
62. Privacy
Protecting information about people and controlling how it is collected, used, shared, or retained. An AI service that processes personal records raises privacy questions about the data and the surrounding application. Privacy is not assured simply because a model produces an anonymous-looking response.
63. AI risk management
The ongoing practice of identifying, assessing, and addressing risks associated with AI systems across their use and lifecycle. The risks can include unreliable outputs, privacy harms, security issues, or unfair effects. NIST’s The Language of Trustworthy AI: An In-Depth Glossary of Terms provides vocabulary for trustworthy AI and is intended to be used with the NIST AI Risk Management Framework or on its own.
How to distinguish commonly confused AI terms
- AI and machine learning: AI is the broader field; machine learning is one way to build AI systems.
- Generative and predictive AI: Generative systems create content; classification and prediction systems assign categories or estimate values. A system can combine both kinds of capability.
- LLM and foundation model: An LLM focuses on language. A foundation model is a broader category that can include models working across multiple modalities.
- Token and word: A word may map to one token or several; a token is the unit a model processes.
- Prompt context and persistent memory: A prompt provides information for an interaction, within the model’s context window. Persistent memory requires an application or system to retain and supply information across interactions.
- Retrieval and generation: Retrieval finds existing material; generation composes new output. RAG combines them.
- Grounding and truth: Grounding gives a response relevant source material, but the answer can still misread that material or make unsupported claims.
- Training and inference: Training adjusts a model using data; inference runs the model to produce an output.
References for evolving terminology
For broader generative-AI vocabulary, see Google Cloud’s Generative AI glossary. For responsible-AI language and risk concepts, consult the NIST glossary and consider its relationship to the NIST AI Risk Management Framework. Definitions describe useful distinctions, but specific products may use terms differently.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
