Glossary
The words, in plain English
Every term the course uses, with the stand analogy it was born in. There are 36 entries, grouped by the serving that introduces them.
Serving 1
The Shuk Before Sunrise
- Defining the problem
-
Decide the pita before the crates arrive.
Choosing exactly what you want the AI to do before gathering anything. A stand keeper does not heat the oil at random; the sandwich comes first.
Related: Data collection, Data cleaning
- Data collection
-
The eggplant crates at the shuk gate.
Gathering the raw information an AI learns from, wherever it lives: sales notes, sensors, websites, clipboards.
Related: Defining the problem, Data cleaning
- Data cleaning
-
Salting the eggplant, then patting it dry.
Fixing and removing bad raw data: duplicates, errors, bruised slices, anything that would ruin the batch. Most of the work in a real AI project sits here.
Related: Data collection, Feature engineering
- Feature engineering
-
Frying the slices to deep gold.
Turning raw data into the concentrated signals a model actually learns from, the way fried eggplant becomes the flavour the pita is built on. You never serve the raw crate; you serve the essence.
Related: Data cleaning, Model training
- Algorithm
-
The recipe, written and followed.
The step-by-step procedure an AI uses to learn from examples. A recipe is not the dish, and an algorithm is not the model; it is the knowing-how.
Related: Model training
- Model training
-
Twenty years of morning rushes.
The learning run itself: showing the algorithm examples again and again until its inner settings settle. AI Sabich calls it remembering mornings.
Related: Algorithm, Evaluation
- Evaluation
-
The first pita of the day, tasted.
Testing the trained model on examples it has never seen, the way a stand keeper tastes before the shutters open. If it fails here, the customers never meet it.
Related: Model training, Deployment
- Deployment
-
Opening the shutters.
Putting the trained model to work in the real world, where real mornings are messier than the practice kitchen.
Related: Evaluation, Model drift
Serving 2
The Three Things Machines Learn
- Classification
-
Firm or tired, slice by slice.
Sorting things into named boxes: this eggplant is firm, that one is tired; this message is spam, that one is real.
Related: Regression, Clustering
- Regression
-
How many pitas tomorrow?
Predicting a number, not a label: tomorrow's pita count from a hundred past mornings, rain and holidays included.
Related: Classification
- Clustering
-
Regulars who never met.
Letting similar things find each other with no labels at all: the early birds, the lunch crowd, the extra-amba people.
Related: Classification
Serving 3
The Stand With a Thousand Eyes
- Neural network
-
Layers in one pita.
A model built from many small simple units stacked in layers, the way a sabich stacks hummus, eggplant, egg, salad and amba into one sandwich.
Related: Computer vision, Large language model
- Computer vision
-
Knowing the coins at a glance.
Machines learning to read images: faces, X-rays, licence plates, ripe fruit.
Related: Neural network
Serving 4
The Cook Who Invents New Dishes
- Generative AI
-
A new amba nobody wrote down.
AI that creates new text, images or sound, instead of only predicting numbers or labels.
Related: Large language model, Classification
- Large language model
-
The friend who read the whole shuk.
A generative model trained on enormous amounts of text, learning the patterns of language so well it can write and answer in it.
Related: Generative AI, Attention
- Attention
-
Ears on the voice that is ordering.
The trick inside modern language models: while reading, the model decides which earlier words matter most right now, the way a stand keeper's ear finds the one voice that is actually ordering in the din.
Related: Large language model
- Prompting
-
Ordering clearly, with the extras named.
Telling a generative AI what you want. Vague order, vague sandwich. The words you choose are the work.
Related: Retrieval-augmented generation
- Retrieval-augmented generation
-
The chalkboard behind the counter.
Answering with facts fetched fresh from your own documents at question time, not only from memory. The chalkboard keeps the cook honest about today's prices.
Related: Prompting, Hallucination
- Foundation model
-
The big batch every stand starts from.
One huge general model, trained once at great cost, that many teams adapt to many jobs instead of starting from zero.
Related: Pre-training, Fine-tuning
- Pre-training
-
The Friday cooking, done in advance.
The big first learning pass over a vast pile of examples. Iraqi families cooked the sabich parts the day before and served them cold on the Sabbath morning; the heavy work happened early. Same idea.
Related: Foundation model, Fine-tuning
- Fine-tuning
-
Savta's twist on the amba.
A small extra training pass that specialises a general model to your stand, your words, your customers, without re-cooking everything.
Related: Pre-training
- Open weights
-
Recipes shared at the shuk gate.
Models whose learned numbers are published for anyone to download, inspect and run themselves.
Related: Open source, Closed model
- Open source
-
The recipe with the method.
Code and recipe published so anyone can check it, change it, and share the improvement back.
Related: Open weights, Closed model
- Closed model
-
The locked recipe book.
A model you can only use through its owner's counter: you order, it serves, and the book stays shut.
Related: Open weights, Open source
- Model drift
-
Tastes move; the recipe must too.
A deployed model slowly going stale as the world changes under it: new customers, new prices, new mornings. Retraining is the repair.
Related: Deployment
Serving 5
The Stand That Thinks for Itself
- Reinforcement learning
-
Tuning amba by smiles.
Learning from rewards instead of examples: try, taste the reaction, keep what earned the smile. One batch at a time.
Related: AI agent
- AI agent
-
The helper who runs the whole stand.
AI that does not only answer but acts: takes the goal, uses tools, checks its own work, and comes back with the sandwich made.
Related: Reinforcement learning, Multi-agent system
- Multi-agent system
-
Fryer, salad, wrapper.
Several agents with divided jobs that hand work to each other, like the stations of a busy stand.
Related: AI agent
Serving 6
When AI Gets a Body
- Embodied AI
-
The cook with hands.
AI running in a physical body that senses and acts in the real world, not only on a screen.
Related: Robot sensors, Moravec's paradox
- Robot sensors
-
Eyes, ears, and oily fingertips.
The cameras, microphones and touch readings a robot uses to know the world; its only door to what is actually happening.
Related: Embodied AI
- Moravec's paradox
-
The full-pita problem.
The hard-won discovery that reasoning is easy for machines and effortless hand work is brutal: stuffing a pita full without tearing it defeats robots that can plan a whole warehouse.
Related: Embodied AI, Sim-to-real
- Sim-to-real
-
Practice kitchen, real shuk.
Training a robot in simulation, then facing physics: the wind, the crowd, the spilled tahini. Simulation is the practice kitchen; the shuk is the shuk.
Related: Moravec's paradox, Embodied AI
Serving 7
The Stand You Can Trust
- Hallucination
-
Change given, confidently wrong.
A generative AI stating something false with total calm, because fluent words are what it learned, not truth. Count your change.
Related: Retrieval-augmented generation, Large language model
- Bias
-
Only the morning customers counted.
Systematic unfairness an AI inherits from skewed examples: train only on the early birds and the stand will misread everyone who comes after nine.
Related: Data collection
- AI ethics
-
The honest scale.
The standard kept when nobody checks: consent, honesty, fairness. A stand keeper's promise, in silicon.
Related: AI governance, Data collection, Bias
- AI governance
-
The certificate on the wall.
The audits, records, risk levels and named humans that make AI answerable, like the certificates above a stand that any customer can point at.
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.