Room 01 · Key Terms
35 terms, in the order they make sense. Each one gives you the definition up front, then a toy to play with until it clicks. Start at the first stop, or jump to any word.
No term matches that. Try a shorter word, or .
Before tokens, prompts or agents, you need to know what the thing actually is. Everything after this builds on these seven.
Computer systems that do tasks we usually connect with human thinking: understanding language, recognizing images, making predictions.
ClarificationMost AI today isn't programmed with rules. It learns from examples, which is the very next term.
A way of building AI where the computer learns patterns from examples instead of following rules a person wrote. That learning step is called training.
ExampleNobody can write a rule for what a cat looks like. Show a system thousands of cat photos and it works out the pattern itself.
Builds on AI
Layers of connected nodes that pass numbers along. Each connection has a strength, and those strengths decide what the network does.
ClarificationLoosely inspired by brain cells, but it's math, not biology. Training just means adjusting the connection strengths until the answers come out right.
Builds on Machine Learning
The adjustable numbers inside a model: the connection strengths it tunes during training.
ExampleWhen people call a model "big," they mean it has a lot of parameters. More parameters can hold more patterns, but cost more to run.
Builds on Neural Network
The trained system itself: all those learned parameters, packaged up to do the work. Claude Opus, Sonnet and Haiku are three different models.
ClarificationThe model isn't the app. ChatGPT and the Claude app are places you chat; the model is the engine underneath. Opus, Sonnet and Haiku run from largest to smallest: bigger is usually smarter, smaller is faster and cheaper. One app can switch between models, and companies release new ones all the time.
Builds on Parameters
A model trained on enormous amounts of text to predict the next token. ChatGPT, Claude and Gemini are all built on LLMs.
Clarification"Large" means billions of parameters, not a big box. Under all the fluent writing, it's doing one thing over and over: guessing the next piece of text.
Builds on Model
AI that creates new content (text, images, music, video) instead of only sorting or recognizing what already exists.
CaveatAn LLM is one kind of generative AI; image and music generators are others. "Generated" means new, not true. It can make things that look real and aren't.
Builds on LLM
Now follow your words through an LLM: chopped into pieces, turned into numbers, held in memory, and answered one piece at a time.
The pieces a model chops text into before processing. Usually word-chunks, not whole words.
Example"unbelievable" might split into "un," "believ," "able." The model never sees letters the way you do, which is exactly why it sometimes miscounts them.
Builds on LLM
Sequences of N consecutive words or tokens, used to study or predict language patterns.
Example"the cat sat" is a 3-gram (a trigram). Early language tech guessed the next word by counting which n-grams showed up most: a simpler ancestor of today's models.
Builds on Tokens
A way of turning words into numbers, coordinates in space, so a machine can measure how related two meanings are.
See it in motionWords as coordinates sounds abstract until you watch it. You place a word on a meaning map yourself in the Learning Lab. Open the Learning Lab
Builds on Tokens
The maximum amount of text a model can hold in mind at once: your prompt and its response combined.
CaveatGo past it and the earliest material slips out of memory. A long conversation can make a model "forget" what you told it at the very start.
Builds on Tokens
The moment a trained model actually runs: it takes your input and produces an output.
ClarificationTraining is when the model learns; inference is when it performs. Every time you hit enter, you trigger inference. You're not teaching it anything new.
Builds on Machine Learning
A setting that controls how adventurous a model's word picks are. Low is safe and predictable; high is surprising and sometimes weird.
ExampleKeep it low for facts and formatting, turn it up for brainstorming. Too high and it stops making sense.
Builds on LLM
A model learns in two big phases, and it can learn the wrong lesson.
The first, massive training phase where a model learns language patterns from enormous amounts of text.
ClarificationThis is the "pre" in pre-trained. It builds broad ability; fine-tuning comes after to specialize. Pre-training is expensive and rare; fine-tuning is cheap and common.
Builds on Machine Learning
Further training an already-trained model on a narrower dataset to specialize it.
ExampleTake a general model, train it on thousands of legal contracts, and you get one fluent in legalese, without building a whole new model from scratch.
Builds on Pre-Training
When a model learns its training data too well, noise included, and stumbles on anything new.
ExampleLike a student who memorizes the exact practice test, then bombs the real one because the wording changed. Memorization isn't understanding.
Builds on Pre-Training
A trained model can handle things nobody trained it for. These are the abilities that make prompting work, which is the next stop.
A model's ability to handle a task it was never explicitly trained on, using general knowledge.
ClarificationDon't confuse this with zero-shot prompting. Prompting is what you do (no examples in your request); zero-shot learning is what the model can do (succeed at an unseen task). One's a writing move, one's a capability.
Builds on Pre-Training
A model's ability to pick up a new task from just a handful of examples.
ClarificationSame split as above: few-shot prompting is you supplying the examples; few-shot learning is the model being sharp enough to generalize from them. The examples are the bridge between the two.
Builds on Zero-Shot Learning
A model that works with more than one kind of input or output: text, images, audio, video.
ExampleTake a photo of a confusing graph and ask what it means. The model reads the image and answers in words.
Builds on Generative AI
You know how it works. Now, how you actually talk to it: prompting, and seven prompt types worth knowing by name.
The act of writing input to pull a useful output from an AI. The core skill of this whole class.
ClarificationA prompt isn't just a question. It's an instruction, a frame, and a set of constraints at once. Better prompting beats a better model more often than people expect.
Builds on LLM
A simple, direct instruction with no examples given.
Example"Translate this into French." You're trusting the model to already know how. Fast, but you're not steering style or format; you get its default.
Builds on Zero-Shot Learning
Provide 2–4 examples for the AI to mimic in style or format.
ExampleShow it three movie reviews rewritten as haiku, then hand it a fourth movie. It matches the pattern. Here, the examples are the instruction.
Builds on Few-Shot Learning
A direct command built on strong action verbs and specific constraints.
ExampleThe power lives in the constraints. "Write about dogs" drifts everywhere; "List five facts about border collies in under 100 words" lands exactly where you aimed.
Assign the AI an expert identity before the task.
Example"You are a marine biologist" primes different vocabulary, depth and assumptions than no role at all. Same question, sharper answer.
Give background (audience, purpose, platform) before you ask.
ExampleThe model can't read your mind. "Write a post" and "Write a LinkedIn post for hiring managers" produce two completely different things from one small difference.
A sequence of prompts where each step builds on the last output.
ExampleFirst prompt brainstorms ten ideas; second narrows to three; third drafts one. You think in stages instead of demanding everything in a single shot.
Behind-the-scenes instructions that configure the AI's behavior before any user input.
ClarificationIt's the difference between telling someone what to do once and setting the rules they follow all day. "Always respond in formal English" shapes every later answer without being repeated.
AI outputs are drafts, not answers. These are the three ways a fluent answer goes wrong.
When a model produces confident, fluent information that is simply false.
CaveatThe danger is the confidence. It won't sound unsure. Invented citations and made-up "facts" read exactly like real ones. Always verify.
Builds on LLM
Systematic skew in a model's output, inherited from skewed training data.
ClarificationIt isn't the model "having opinions." It's a mirror of patterns in what it read. If the data leaned one way, the output leans too, often invisibly.
Builds on Pre-Training
When a model tells you what you want to hear instead of what's true or useful.
CaveatModels are partly trained on human ratings, and people tend to rate agreement highly. So if you lead the question, it often follows. Ask for weaknesses, not reassurance.
Builds on Prompting
So far it answers and stops. These keep going: they take steps, use tools, and build things.
An AI system that doesn't just answer but takes actions on its own to complete a goal, often across several steps.
ClarificationA regular model responds and stops. An agent keeps going: it can search, use tools, check its own work, and decide the next move without you prompting each step. You hand it the goal, not the instructions.
Builds on Chain
An AI agent framed as a virtual "employee" that owns an ongoing role or workflow rather than a single task.
ClarificationSame engine as an agent, different packaging. "Agent" describes what it does (acts toward a goal); "digital worker" describes how it's sold: as a teammate that handles a job like an inbox or a scheduling queue, day after day.
Builds on Agent
Building software by describing what you want in plain language and letting AI write the code, often without reading every line of it.
CaveatAI researcher Andrej Karpathy coined the term in 2025. It's fast, and parts of this site were built that way. But code can look right and still break, so someone still has to test it and know enough to spot what's wrong.
Builds on Agent
The words you'll hear in the news. They only make sense once you know everything before them.
The work of making AI systems do what people actually intend, and stay helpful, honest and safe while doing it.
ClarificationAI does what you said, not what you meant. Give it the goal "get clicks" and it will find clicks however it can. Alignment is closing that gap.
Builds on Agent
A hypothetical AI that could learn and do nearly any thinking task a person can, not just the ones it was built for.
CaveatIt doesn't exist yet. Experts disagree about what would count, and about whether or when it will arrive. Be skeptical of anyone who says it's definitely here, or definitely impossible.
Demo answers and percentages on this page are written by hand to show how the real thing behaves. They aren't generated live.