AI concepts by group

What is artificial intelligence?

AI is a field of research and engineering that builds systems able to produce predictions, recommendations, decisions or content from inputs. It includes machine learning, perception systems, pattern recognition and generative models.

A spam filter, demand forecast and writing assistant can use AI in very different ways. With deterministic automation, the same inputs and conditions produce the same result; a generative model may vary. Ask what task the system performs, how it was evaluated and where it can fail.

In a rule-based system, people write conditions such as “if X happens, do Y.” In machine learning, training adjusts parameters from examples to recognize patterns. People still choose the data, objective and evaluation; rules and models can work together.

Predictive AI estimates an outcome, such as a music recommendation, arrival time or signal of possible fraud. Generative AI creates new text, images, audio or code. This distinction describes the main use: generative models also make predictions while generating.

What is MCP?

Model Context Protocol is an open standard for connecting AI applications to external systems. An MCP client can discover and use resources or tools exposed by servers, including files, databases and specialized services.

MCP is not a model and does not make an answer automatically correct. It expands what an assistant can access or do; every connection needs appropriate permissions, clear limits and care with shared data.

RAG, agents and tools

RAG retrieves passages from documents to support an answer. An agent can plan steps, consult tools and review results; a tool is an external capability such as search or file access.

Retrieval does not automatically validate a source. An agent can also take a wrong action if granted broad permissions or given malicious instructions in external content. Limit access and require confirmation for consequential actions.

How the concepts fit together

An assistant may use an LLM to interpret a question, RAG to retrieve documents and MCP to access a tool. An agent coordinates these steps. Training, inference, context and tokens describe different parts of the process.

When reading a release, identify what actually changed: the model, interface, integration or availability. This keeps unlike announcements from being compared as if they were the same.

Machine learning, training and inference

Machine learning is a part of AI in which models adjust parameters from examples. Training is that adjustment stage; inference uses an already trained model to produce an output for a new input.

An ordinary assistant conversation does not, by itself, mean the model's parameters changed. Chat history, product memory and training are different mechanisms; check the service policy to learn how your data may be used.

What is an LLM?

LLM means Large Language Model. It learns patterns from large datasets and generates a sequence of tokens in response to context. It can summarize, draft, classify and assist with code.

Fluency does not guarantee truth. A model may produce a convincing error, miss details in a document or reflect bias. Serious use calls for sources, tests and human review.

A model is a mathematical structure whose parameters are adjusted during training. Think of it as an engine, not a human brain. ChatGPT is the product people use, with its own interface and features; GPT names a family of models that may power it. An application can offer different models.

Tokens, context window and memory

Tokens are units used to process input and generate output. As a rough guide, 1,000 tokens correspond to about 750 words in common English; the ratio changes with language, text and model. The context window limits how much material can fit in processing and varies by model and product.

A larger window does not guarantee equally accurate recall of every detail. Persistent memory is a separate feature when available. For long tasks, split the work, ask for references to passages and check the original document.