AI terminology is useful when it helps a team make a concrete decision. It becomes a distraction when a label stands in for the system’s actual behaviour.
These twenty-five terms cover models, retrieval, adaptation and operation. Definitions describe the engineering concept rather than a promise that a particular tool is reliable, secure or ready for your use case.
Models and learning
Artificial general intelligence
AGI describes a broadly capable hypothetical form of AI, but there is no single universally accepted operational definition. It is not a useful acceptance criterion for a bounded business application. Specify the tasks and tests you actually need.
Narrow AI
A term for AI designed or evaluated for a limited set of tasks. A system can be highly capable within that scope and fail outside it. A broad conversational interface does not remove task-specific limitations.
Neural network
A parameterised computational model made of connected operations, trained to approximate patterns or functions. Its structure and training objective influence what it can learn. Calling it a network does not imply human understanding.
Transformer
A neural-network architecture using attention mechanisms to process relationships between representations. Transformers underpin many language models, but architecture alone does not determine quality, context capacity or suitability for a business task.
Gradient descent
A family of optimisation methods that update parameters using gradients of a loss function. Reducing training loss is not the same as proving real-world correctness, security or generalisation.
Large language model
A model trained on large amounts of language and other data to predict or generate sequences. Applications can add retrieval, tools and validation. Those surrounding capabilities should not be confused with knowledge or authority inherent in the model.
Adaptation and input
Fine-tuning
Additional training that adapts a model’s parameters to examples or a task. It can improve particular behaviours, but it is not an inspectable store of current business facts. Evaluation needs held-out examples and regression checks.
LoRA and parameter-efficient fine-tuning
Methods that adapt a limited set of parameters rather than updating every base-model weight. The LoRA research reports results for specific experiments; there is no universal accuracy or cost percentage for every task.
Embedding
A numerical representation used to compare items such as text passages. Similarity can help retrieval, but nearby representations do not establish that two documents say the same thing or that a passage answers the question.
Context window
The token capacity available to a model request under its interface’s accounting rules. It is not durable application memory. Important state and completed external actions need an authoritative record outside the conversation.
Hallucination
Generated content that is unsupported, inaccurate or fabricated relative to the task and evidence. A fluent answer or real citation can still contain an unsupported claim. Evaluate the specific failure rather than treating the term as one uniform cause.
Prompt engineering
Designing and testing the instructions, examples and information supplied to a model. Prompts can guide behaviour, but they are not an enforceable access-control boundary. Permissions belong in software and connected systems.
Retrieval and agents
AI agent
Software using a model to help select steps towards a goal, often with tools. The surrounding application must validate requests and authorise actions. The degree of autonomy varies, and more autonomy is not automatically more useful.
Retrieval-augmented generation
Supplying externally retrieved information to generation at request time. Retrieval can use several search methods. It needs source ownership, access control, freshness and evaluation; it does not guarantee that an answer is supported.
Vector database
A store supporting search over vector representations. It can be part of retrieval, but it does not by itself provide a complete knowledge system, establish document authority or manage every user permission.
Chain of thought
A generated sequence of intermediate reasoning text. It can be useful in some tasks, but research on faithfulness shows that it should not be treated as a complete causal explanation of a model’s output.
Agent framework
A library or toolkit for assembling model calls, tools and workflows. It can provide useful building blocks. Choosing one does not automatically establish the application’s permissions, business rules, recovery or operating ownership.
Generative media
Diffusion model
A generative-model family commonly trained to reverse a noising process. Diffusion approaches are used in image and other generation systems. The technique does not establish originality, accuracy or rights to use a particular output.
Stable Diffusion
A family of diffusion-based models and related implementations. Versions and licences differ. Evaluate the actual model and its terms rather than assuming that one name implies identical capabilities or unrestricted use.
Text-to-image
Generating images from a text request, sometimes with additional reference inputs. Outputs may need factual, brand and rights review. A visual that looks plausible is not documentary evidence of a real event or product.
Multimodal AI
Systems that process or generate more than one type of information, such as text and images. Supported formats, context limits and quality vary. Evaluate each modality and the relationships the task requires the system to understand.
Safety and operation
Responsible AI
An umbrella term for practices addressing the effects, risks and accountability of AI. It needs specific requirements and evidence to be meaningful. It is not a technical feature that can simply be switched on.
AI alignment
Work concerned with how a system’s behaviour relates to intended objectives and human expectations. A model following ordinary instructions in a demonstration is not proof that it will behave as intended in every situation.
Reinforcement learning from human feedback
A family of training approaches using human feedback, often through a learned preference signal. It can shape behaviour, but it does not establish factual correctness or eliminate harmful outputs.
AI red-teaming
Structured adversarial testing intended to find failures and weaknesses. Scope, attack methods and tested versions matter. Finding no issue in a particular exercise is not proof that no vulnerability exists.
Use the terms to ask better questions
Ask what the system actually receives, stores, retrieves and changes. Identify where software checks a proposal and where a person decides. Then ask for task-specific evidence.
Our glossary includes related operational terms, while AI Software Engineering explains the application work around the model.
Written by Raj Dalal, Advisor at Ryzolv.
Back to Insights