The key components of modern AI, explained so that you can follow a project discussion and ask the questions that matter.
This is not a course in mathematics or model training. You will learn what central terms mean in day-to-day business, which role they play in a solution and why a model alone is not a finished application.
what an LLM processes and how it produces an answer
why an AI model is not the same as a finished application
which capabilities an AI platform brings together
how AI, machine learning, deep learning and generative AI relate
what reasoning means in AI models and when extra compute is justified
These questions connect a technical term to a concrete organisational decision.
Do we need a model, a finished application or a shared platform?
Where should the AI obtain current and internal information?
Which parts of the solution must remain under our control?
A large language model is an AI model that learns statistical patterns from very large collections of text and uses them to generate fitting next pieces of text.
An AI platform is a central software foundation through which employees and applications can use AI with approved data, knowledge and business systems in a controlled way.
An AI model processes inputs and generates outputs. An AI application combines one or more models with a specific task, interface and business logic. An AI platform provides shared technical and organisational building blocks for multiple applications.
Artificial intelligence includes systems that infer predictions, content, recommendations or decisions from inputs. Machine learning is a family of data-driven methods within AI; deep learning is machine learning based on multi-layer neural networks. Generative AI creates new content and today usually relies on deep learning. AGI, by contrast, is not a current product category but a hypothetical concept without one universally accepted definition.
Traditional software calculates a result defined by code and rules. Generative AI probabilistically creates text, images or other content from a model, so it can respond flexibly but not with complete predictability.
Tokens are small units into which a model breaks text and, in some systems, other content. The context window is the maximum token budget shared by instructions, conversation, sources, tool results and the response.
Multimodal AI can process or generate different data types such as text, images, audio and video within an application and relate information across them.
Prompt engineering is the systematic design and testing of instructions, context and examples so that a model performs a clearly defined task as reliably as possible.
Fine-tuning is additional training of a pretrained model on a curated dataset so that it performs particular tasks, formats, terminology or behavioural patterns more consistently.
Alignment is the overarching goal of bringing an AI system's behaviour closer to its intended tasks, human instructions and safety rules. In reinforcement learning from human feedback, or RLHF, people compare model responses. Their preferences are converted into a reward signal that is used to optimise the language model further.
Model choice balances task quality, language and modalities, latency, cost, context needs, data processing, adaptability and operational responsibility. A small language model (SLM) can be a better fit for narrowly bounded language tasks than a larger LLM.
Reasoning in current AI models means additional processing during inference: the model may internally break a task into smaller parts, compare intermediate results and adjust its approach before producing the final output. It is a technical capability for handling certain tasks, not human thought or consciousness. Greater reasoning effort can improve quality, but it usually increases latency, token usage and cost and must therefore be evaluated on your own cases.
We translate your starting point into understandable options and a realistic next step.
