Glossary
The words, in plain English
Every term the course uses, with the pub analogy it was born in. There are 36 entries, grouped by the serving that introduces them.
Serving 1
The Pub Before Opening
- Defining the problem
-
Decide the dish before the delivery.
Choosing exactly what you want the AI to do before gathering anything. A publican does not light the fire at random; the dish comes first.
Related: Data collection, Data cleaning
- Data collection
-
The morning delivery at the back door.
Gathering the raw information an AI learns from, from wherever it lives: ledgers, bookings, sensors, websites.
Related: Defining the problem, Data cleaning
- Data cleaning
-
Picking over the vegetables.
Fixing and removing bad raw data: duplicates, errors, missing values, anything that would ruin the stew. Most of the work in a real AI project sits here.
Related: Data collection, Feature engineering
- Feature engineering
-
Browning the lamb.
Turning raw data into the concentrated signals a model actually learns from, the way seared lamb builds the base of the stew. You never serve the raw crate; you serve the essence.
Related: Data cleaning, Model training
- Algorithm
-
Picking the recipe.
The method a machine follows to learn from data. Same ingredients, different recipes, different stews.
Related: Model training
- Model training
-
The pot on the fire.
The step where the machine studies the prepared data again and again, adjusting itself each pass until its answers stop improving. Shorter than everyone thinks, once the prep is done.
Related: Algorithm, Evaluation
- Evaluation
-
Taste and adjust.
Testing the trained model on examples it has never seen, the way a cook tastes before serving. Too bland means back to the pot.
Related: Model training, Deployment
- Deployment
-
Serving the bowl.
Putting the finished model to work in the real world, where real people use it and real consequences follow.
Related: Evaluation, Model drift
Serving 2
The Three Things Machines Learn
- Classification
-
Fresh carrot or spoiled carrot.
Teaching a machine to sort things into named groups from labelled examples. The most common job in machine learning.
Related: Regression, Clustering
- Regression
-
How many bowls on a rainy Friday.
Teaching a machine to predict a number from past patterns: sales, crowds, prices, temperatures.
Related: Classification
- Clustering
-
The regulars sort themselves into corners.
Letting a machine find natural groups in data nobody has labelled. Nobody tells it the groups; it finds them.
Related: Classification
Serving 3
The Chef With a Thousand Eyes
- Neural network
-
Twenty cooks, one stockpot.
A model built from many small, simple decision units stacked in layers. No single unit understands anything; together, they do.
Related: Computer vision, Large language model
- Computer vision
-
How a pub reads a photo.
Machines learning to recognise what is in an image, from the face that unlocks your phone to carrots on a sorting belt.
Related: Neural network
Serving 4
The Chef Who Invents New Dishes
- Generative AI
-
The chef who invents new dishes.
AI that creates new text, images, or sound, rather than sorting or predicting what already exists.
Related: Large language model, Classification
- Large language model
-
The chef who has read every cookbook.
A model trained on enormous amounts of text, whose core skill is predicting the next word well enough to write whole passages.
Related: Generative AI, Attention
- Attention
-
Keeping an eye on the right pot at the right moment.
The mechanism inside modern AI that weighs which parts of the input matter most for the word it is writing now.
Related: Large language model
- Prompting
-
How you order changes what you get.
Writing the instruction you give a generative model. Small changes in wording change the answer, which is why it is a skill.
Related: Retrieval-augmented generation
- Retrieval-augmented generation
-
Checking the pantry before you answer.
Giving a model the exact documents it needs at the moment you ask, so it answers from your facts instead of guessing. Your facts go on the counter before the fire starts.
Related: Prompting, Hallucination
- Foundation model
-
The master chef, trained over years.
A very large AI, trained at great cost on enormous amounts of text or images, that arrives already skilled. Most AI work today starts from one and adapts it.
Related: Pre-training, Fine-tuning
- Pre-training
-
The cooking school years.
The enormous first training run that gives a model its general skill. Expensive, slow, and usually someone else's job.
Related: Foundation model, Fine-tuning
- Fine-tuning
-
Coaching the chef to cook your style.
Extra training that teaches an existing model a specific style, format, or task, using far less data and money than building one from scratch.
Related: Pre-training
- Open weights
-
The stew plus the recipe card.
A model you can download, run yourself, and even adjust, while the training data and full method stay secret.
Related: Open source, Closed model
- Open source
-
The open kitchen.
A model released with weights, training code, and data details, so anyone can rebuild or deeply change it. You watched it cook.
Related: Open weights, Closed model
- Closed model
-
Ordering at the hatch, kitchen hidden.
A model you can only use through its owner's service. You send a request, receive the result, and never see the kitchen.
Related: Open weights, Open source
- Model drift
-
The regulars' tastes slowly change.
What happens when the world changes and the model does not: slowly, its answers stop fitting. The fix is monitoring and retraining, not blame.
Related: Deployment
Serving 5
The Pub That Thinks for Itself
- Reinforcement learning
-
A nod and a grumble, trial and error.
Training by consequences: the machine tries things, earns rewards or penalties, and keeps what paid off. No instructions, only feedback.
Related: AI agent
- AI agent
-
The publican who runs the whole service.
AI that takes a goal and works until the job is done, planning steps, using tools, and checking its own progress along the way, instead of answering one question.
Related: Reinforcement learning, Multi-agent system
- Multi-agent system
-
A team of specialists behind one bar.
Several agents with different jobs passing work to each other, the way the cook, the server, and the bookkeeper run one pub.
Related: AI agent
Serving 6
When AI Gets a Body
- Embodied AI
-
The day the recipe gets hands.
AI connected to a physical body, with sensors and movement, that must survive gravity, heat, and a slippery ladle.
Related: Robot sensors, Moravec's paradox
- Robot sensors
-
The five senses of a robot publican.
Cameras, microphones, and force and heat sensors that let a robot perceive the physical world, each standing in for a sense a publican never thinks about.
Related: Embodied AI
- Moravec's paradox
-
The crumble problem.
The finding that tasks humans find hard, like chess, are easy for AI, while tasks we find easy, like slicing soda bread, are brutally hard. The skill is in the body.
Related: Embodied AI, Sim-to-real
- Sim-to-real
-
Practising in a toy pub first.
Training a robot in a simulation, where failing costs nothing and time runs fast, before it ever touches the real ladle.
Related: Moravec's paradox, Embodied AI
Serving 7
The Pub You Can Trust
- Hallucination
-
Confidently serving the wrong bowl.
When an AI states something untrue with complete confidence, because it is predicting plausible words, not checking facts.
Related: Retrieval-augmented generation, Large language model
- Bias
-
A pub that only learned from one street.
Systematic unfairness in a model's answers, inherited from skewed training data. Bias in, bias out.
Related: Data collection
- AI ethics
-
The promise the publican keeps at 6 in the morning, when nobody is watching the back door.
The standard the people behind an AI system hold when nobody is checking: asking before taking data, telling the truth about what the system can do, and serving nothing they would not feed their own family.
Related: AI governance, Data collection, Bias
- AI governance
-
The licence, the inspection book, and the unannounced inspector.
The procedures and checks that make AI trust provable: testing before release, written records of what a model contains, risk levels that decide how hard the inspection is, and a named human who stays answerable for decisions that touch a life.
Related: AI ethics, Evaluation, Deployment
The AIFN Community
Learning AI is better with company.
Ask questions, share what worked, and keep pace with fellow learners.