A handy reference of key terms and concepts you'll encounter throughout the programme.
Entries are sorted alphabetically. If you come across a term that's missing, drop a note in Slack!
Watch this TED talk from Microsoft AI CEO explaining what AI is:
Watch: What is AI anyway? →An AI agent is a software that can decide what to do next in order to reach a goal, instead of just answering a single question and stopping. It can use tools (like search or email), remember some information, and take several steps in a row to complete a task with limited human guidance.
Not only used in AI context, but in software development. An API (Application Programming Interface) is a safe doorway that lets one program talk to another program in a controlled way. In AI, an API often lets an app send text to a model (like a chatbot) and receive the model's response without needing to host or build the model itself.
Very simple "if this, then that" kind of logic — a pre-programmed, baked-in solution. In AI, automation can mean things like automatically replying to customer emails, sorting documents, or generating reports based on data.
The practice of carefully choosing what information to give an AI model so it has just enough background to respond well. "Context bloat" happens when you stuff in too much information, making the model slower, more expensive, and sometimes more confused. "Context rot" refers to old, irrelevant, or conflicting information that quietly reduces answer quality over time.
The maximum amount of text (measured in tokens) that an AI model can pay attention to at once. If you go beyond this limit, older parts of the conversation or document must be shortened, dropped, or summarized for the model to keep working.
Short for evaluations — tests used to check how well an AI system performs on specific tasks, such as answering questions correctly or following safety rules. They often use prepared question sets and scoring methods to compare different models and to see if new versions are getting better or introducing new risks.
Stands for "Generative Pre-trained Transformer," a family of autoregressive models created to generate text, code, and other content from natural-language prompts. Autoregressive models use the previous words to predict the next word in a sequence.
Large Language Model — a kind of AI model trained on huge collections of text so it can recognize and produce language that feels natural to humans. It can answer questions, summarize documents, translate languages, and follow instructions by predicting what words should come next in a sequence.
Model Context Protocol — a way of connecting AI models to external tools and data sources in a consistent, structured manner. Different tools (like databases, search engines, or company systems) can be safely plugged into models, making it easier to build capable AI assistants without custom integrations each time. Created by the Anthropic team.
Setups where several AI agents work together, each handling part of a larger task. For example, one agent might gather information, another checks the quality, and a third produces the final answer — with information passed between them like a relay team.
Product Requirements Document. A written description of what you are building, who it is for, and what problems it solves. Used to align thinking before writing code.
The instruction or question you give to an AI model. Good prompts are specific, include context, and describe the desired output format.
The skill of writing and structuring the instructions given to an AI model so it produces useful, reliable results. It can involve choosing clear wording, giving examples, and specifying style or format, so the model better understands what the user wants.
Retrieval-Augmented Generation — a method where an AI model first looks up relevant information (retrieval) and then uses that information to generate an answer (generation). This helps the model stay closer to up-to-date or specialized sources instead of relying only on what it learned during training. Perplexity is a good example of how it's built on RAG principles.
The small pieces that AI models use to process text, which can be whole words or just parts of words depending on the language. According to MIT, a token is typically 4 characters in English, or about ¾ of a word. The model counts these pieces to measure how long an input or output is, which affects cost, speed, and whether the text fits into the context window.
A system for tracking changes to files over time. Git is the most common version control system. GitHub is a platform for hosting Git repositories and collaborating on code.
An informal term for adjusting prompts and examples until an AI system "feels" like it is responding in the desired tone, style, or personality, rather than focusing on precise technical rules. Someone experimenting with vibe coding might tweak wording, add sample conversations, or change descriptions of the assistant's role to steer how it behaves.
Join a hands-on programme and learn to build real AI-powered products — not just read about them.