Advertisement
A
- Agent
- An AI system that plans and takes multiple actions (calling tools, browsing, writing files) to reach a goal, instead of answering a single prompt.
- Alignment
- Research and techniques that make AI systems behave according to human intentions and values.
- API
- Application Programming Interface — the way software sends requests to an AI model and receives results, usually billed per token.
- Attention
- The mechanism inside transformers that lets a model weigh which earlier words matter most when producing the next one.
B
- Benchmark
- A standard test set used to compare models, such as coding, maths or reasoning suites. Useful, but never the full picture.
C
- Chain of thought
- Having a model work through intermediate reasoning steps before giving a final answer, which often improves accuracy on hard problems.
- Chatbot
- A conversational interface to an AI model. Modern chatbots are powered by large language models.
- Context window
- The maximum amount of text (measured in tokens) a model can consider at once, including your prompt, files and its reply.
D
- Diffusion model
- A generative model that creates images, audio or video by gradually removing noise from random data.
- Distillation
- Training a smaller, cheaper model to imitate a larger one so it keeps much of the quality at lower cost.
E
- Embedding
- A list of numbers representing the meaning of text or images, used for semantic search, clustering and recommendations.
- Evals
- Systematic tests you write for your own AI feature to measure quality, regressions and failure modes.
F
- Few-shot prompting
- Including a handful of worked examples in a prompt so the model copies the pattern and format you want.
- Fine-tuning
- Further training a pre-trained model on your own examples to specialise its style or task performance.
- Foundation model
- A large model trained on broad data that can be adapted to many downstream tasks.
- Function calling
- A model returning a structured request to run a specific tool or function, letting apps connect AI to real systems.
G
- Generative AI
- AI that produces new content — text, images, audio, video or code — rather than only classifying existing data.
- GPU
- Graphics Processing Unit — the parallel chips used to train and run most modern AI models.
- Grounding
- Tying a model's answer to trusted sources such as documents or search results to reduce made-up information.
- Guardrails
- Rules, filters and checks placed around an AI system to keep its outputs safe, on-topic and compliant.
H
- Hallucination
- When a model states something false with confidence. Mitigated with grounding, citations and verification.
I
- Inference
- Running a trained model to get outputs. Inference cost and speed matter most for live products.
J
- Jailbreak
- An attempt to trick a model into ignoring its safety rules. Responsible builders test against these.
K
- Knowledge cutoff
- The date after which a model has no training data, so it may not know recent events without search.
L
- Large language model (LLM)
- A neural network trained on huge amounts of text to predict and generate language.
- Latency
- How long a model takes to respond. Important for chat, voice and real-time applications.
- LoRA
- Low-Rank Adaptation — an efficient fine-tuning method that trains small add-on weights instead of the whole model.
M
- MCP
- Model Context Protocol — an open standard for connecting AI assistants to tools and data sources in a consistent way.
- Mixture of experts (MoE)
- A model design where only some specialised sub-networks activate per token, improving efficiency at scale.
- Multimodal
- Able to understand or generate more than one type of data, such as text plus images, audio or video.
N
- Neural network
- Layers of connected mathematical units whose weights are learned from data; the basis of deep learning.
O
- Open weights
- Models whose trained parameters are published so anyone can download, run and adapt them.
- Overfitting
- When a model memorises its training examples and performs poorly on new data.
P
- Parameters
- The learned numeric weights inside a model. More parameters often means more capability and higher cost.
- Prompt
- The instructions and context you give a model. Clear goals, constraints and examples produce better results.
- Prompt injection
- Hidden or malicious instructions in content an AI reads that try to hijack its behaviour.
Q
- Quantization
- Storing model weights at lower precision so it runs faster and uses less memory, with small quality trade-offs.
R
- RAG
- Retrieval-Augmented Generation — fetching relevant documents first and giving them to the model so answers are grounded.
- Reasoning model
- A model trained to spend extra compute thinking through problems before answering, improving maths, code and logic.
- Reinforcement learning from human feedback (RLHF)
- Training a model using human preference ratings so its responses become more helpful and safe.
S
- Semantic search
- Search by meaning using embeddings, so a query finds relevant content even without matching keywords.
- System prompt
- Top-level instructions that set an assistant's role, rules and tone for an entire conversation.
T
- Temperature
- A setting controlling randomness. Low values give focused, repeatable output; high values give more variety.
- Text-to-speech (TTS)
- Converting written text into spoken audio with synthetic voices.
- Token
- A chunk of text (roughly ¾ of an English word) that models read and write. Pricing and limits are counted in tokens.
- Tool use
- A model's ability to call external tools — search, calculators, code execution, APIs — to complete tasks.
- Training data
- The examples a model learns from. Its quality, licensing and diversity shape model behaviour.
- Transformer
- The neural network architecture behind modern LLMs, built around attention.
V
- Vector database
- A database optimised to store embeddings and find the most similar items quickly; core to RAG.
- Vibe coding
- Building software mainly by describing what you want to an AI and iterating on its output.
Z
- Zero-shot
- Asking a model to do a task with no examples, relying on its general training.