Glossary
Every term used by the lessons you can open. These are applied automatically: the first time a lesson mentions an abbreviation it is spelled out, and later mentions carry the definition on hover. You should rarely need this page mid-lesson, and it grows as you open more of the library.
17 terms · 6 abbreviations
A
- agentic AI
- Systems in which models decide and act autonomously through tool calls, shifting the risk question from what the model says to what it is able to do.
- agentic
- AIArtificial Intelligence
- Any system built to perform tasks that would ordinarily require human intelligence. It describes the goal, not the technique, so a hand-written rules engine qualifies.
- foundations
- alignment
- The work of shaping a trained model's behaviour towards intended norms and refusals, usually through curated examples and preference-based reinforcement rather than architectural change.
- mlllm
- APIApplication Programming Interface
- A defined interface for programmatic access between systems. In AI it is usually where the model is exposed, and therefore where abuse, extraction and rate limits apply.
- foundations
E
- embedding
- A dense vector representation placing semantically similar items near one another so meaning can be compared numerically. Embeddings retain enough information to leak their source content.
- foundationsllmdata
F
- fine-tuning
- Continued training of a pre-trained model on a smaller, targeted dataset to specialise its behaviour, which also reintroduces risk when that dataset is not controlled.
- mlllm
G
- GenAIGenerative AI
- AI whose purpose is to synthesise new artefacts such as text, images, audio, video or code, rather than only to classify or score inputs that already exist.
- foundationsllm
- guardrail
- A control sitting around a model rather than inside it, screening inputs and outputs against policy. Guardrails constrain behaviour without changing what the model learned.
- defencellm
H
- hallucination
- Fluent output that is not grounded in fact or in supplied sources, a direct consequence of models predicting likely continuations rather than looking anything up.
- llmfoundations
I
- inference
- Runtime use of a trained model: computing a distribution over possible outputs and sampling from it, which is why identical prompts need not produce identical answers.
- foundationsml
M
- MCPModel Context Protocol
- An open protocol for connecting models to external tools, data sources and prompts through a standard client-server interface, so integrations need not be rebuilt per application.
- agenticllm
- MLMachine Learning
- The subset of AI in which behaviour is derived from data rather than written by hand, by adjusting parameters during training to fit examples.
- foundationsml
R
- RLHFReinforcement Learning from Human Feedback
- An alignment method that trains a reward model from human preference comparisons, then uses reinforcement learning to push the base model towards the responses people rated higher.
- mlllm
S
- system prompt
- The operator-supplied instructions placed ahead of user input to set a model's role, constraints and tone. It is a behavioural control, not a security boundary.
- llmdefence
T
- tokenisation
- Breaking text into words, sub-words or characters and mapping each to a numeric identifier, the first step before anything reaches the model itself.
- foundationsllm
- tool calling
- The mechanism by which a model requests execution of a defined function, letting it query databases, call APIs, run code, manage files or send messages.
- agenticllm
V
- vector database
- A store optimised for similarity search over embeddings, used to serve retrieval in RAG systems. Pinecone, Weaviate, Chroma and FAISS are common examples.
- llmdata