A Beginner's Field Guide

AI 101
From Zero

Understand the core ideas, products, and vocabulary of AI through everyday analogies — written for anyone who wants to make sense of this revolution.

Contents

8 chapters · 65 terms

00 What is AI? Foundations 01 Core AI Concepts 9 terms 02 Products & Tools 10 terms 03 AI Training Terms 10 terms 04 Developer Tools 12 terms 05 Infrastructure 14 terms 06 Version Control / Git 8 terms 07 Evaluation (Eval) 6 terms

Chapter 00Foundations

What is AI?

Before learning any jargon, let's get clear on what AI actually is — and how it differs fundamentally from the kind of programs we've known all along.

The simplest explanation

AI (artificial intelligence) is technology that lets computers imitate the way humans think and learn. The earliest AI just followed rules: "If it rains, take an umbrella." Modern AI, by contrast, can pick up patterns on its own from huge amounts of data — just like a person. You don't have to spell out every step; give it enough examples and it figures out the rules itself.

Two ways of thinking

A · Traditional

Traditional Programs

You tell the computer exactly what to do at every step. "If A, then do B; if C, then do D." Every rule has to be written by a human in advance, so anything you didn't anticipate breaks the program.

AnalogyLike a recipe — follow each step exactly; miss one and dinner is ruined.

B · Artificial Intelligence

Artificial Intelligence

You feed the computer lots of examples and let it work out the patterns itself. No rules required — the AI learns. When it meets a new situation, it makes a reasonable guess based on what it has seen before.

AnalogyIt's like a chef who, after watching a thousand dishes being made, learns how to cook on their own.

A short history

AI isn't new — but it has leapt forward in recent years

1950s

The idea of AI is born; Turing asks whether machines can think.

1990s

Computers start beating world champions at chess.

2012

Deep learning delivers a breakthrough in image recognition.

2022

ChatGPT arrives and AI enters everyday life.

2026

AI can write, code, compose, and paint.

Chapter 01Core Concepts

Core AI Concepts

The most fundamental ideas behind how AI works. These nine terms underpin every chapter that follows — everything else is built on top of them.

9 terms

01.01

LLM

The "brain" behind AI assistants like ChatGPT, Claude, and Gemini. It learns the patterns of language by reading nearly all of the text on the internet, then uses what it has learned to understand and generate natural language.

AnalogySomeone who has read billions of books since childhood — they know a little about every topic.

01.02

Token

The smallest unit of text that AI processes. Roughly every 3–4 English letters equals 1 token; every 1–2 Chinese characters equals 1 token. Both the cost and the speed of AI depend on token count.

AnalogyWhen AI reads, it takes one small "bite" at a time — each bite is a token.

01.03

Context Window

The maximum amount of text the AI can "see" at once, measured in tokens. Anything beyond the limit gets "forgotten" — the oldest content drops out first. Bigger window, longer memory.

AnalogyThe AI's short-term memory — only so many documents fit on the desk; the older ones get pushed off to make room.

01.04

Prompt

The question or instruction you type into the AI. The clearer and more specific you are, the better the answer. "Prompt Engineering" is the craft of writing better instructions.

Analogy"Write me an article" vs. "Write me a 500-word article introducing AI to a 30-year-old office worker" — completely different results.

01.05

System Prompt

A behind-the-scenes instruction set up before the conversation starts that quietly tells the AI "who you are, what you should do, and what you must not do." Users never see it, but the AI follows it the whole time.

AnalogyThe internal briefing a manager gives staff before a meeting — customers never hear it, but staff stick to it.

01.06

Temperature

The dial that controls how "creative" the AI's answers are, on a scale from 0 to 2. Low temperature = safe and predictable. High temperature = wilder, more imaginative, but more likely to go off the rails.

AnalogyLow temperature is a careful accountant; high temperature is an artist who's had three espressos.

01.07

RAG

Before answering, the AI looks up relevant information in an external database, then uses what it found to compose a reply. This solves the problem of an AI's knowledge having a "cut-off date."

AnalogyGoing to the library to look things up before the exam, instead of relying on what you've memorized.

01.08

Hallucination

When AI "makes things up." It doesn't know the answer but doesn't want to admit that, so it confidently produces something that sounds plausible — but is entirely fabricated. Always double-check anything important.

AnalogyAsking a colleague a question; they aren't sure but don't want to look unsure, so they invent an answer on the spot.

01.09

Embedding

Turning words into number coordinates. "Cat" and "dog" end up close together; "cat" and "airplane" end up far apart. Embeddings are how AI knows which words mean similar things.

AnalogyPutting every word on a giant map — words with similar meanings live in the same neighborhood.

Compare

Prompt vs System Prompt vs Instruction

Term	User sees it?	When set	Purpose
Prompt	Yes	During each conversation	Your questions and instructions
System Prompt	No	Before the conversation starts	Sets the AI's role and rules
Instruction	—	Anytime	Umbrella term for both of the above

Chapter 02Products & Tools

Products & Tools

A look at how AI gets packaged into different products and tools — and the small but important differences between them.

Evolution

Bot → Assistant → Agent

Stage 01

Bot

Follows fixed rules; not very flexible.

Stage 02

Assistant

Reactive — answers one question at a time.

Stage 03

Agent

Plans on its own and takes initiative across multi-step tasks.

10 terms

02.01

Agent

An AI that can complete multi-step tasks on its own. It doesn't just answer questions — it makes a plan, calls tools, and takes actions, like searching for information, writing a report, and saving the file.

AnalogyYou say "book me a flight," and it compares prices, picks a seat, pays, and emails you the confirmation — all without you lifting a finger.

02.02

Copilot

Instead of replacing you, the AI sits beside you and helps. You're in charge; the AI suggests and autocompletes. You write code, it finishes the line; you write an article, it polishes the prose.

AnalogyA Copilot is like a co-pilot — you fly the plane. An Agent is autopilot — you give the destination and it flies there itself.

02.03

Plugin

An expansion pack that adds several capabilities at once. Install a "design plugin," for example, and you get a whole bundle of design-related tools. A plugin can contain multiple Connectors and Skills.

AnalogyEquipping an employee with a whole "department skill kit" — one install, a full set of related abilities.

02.04

Connector

A bridge that lets the AI read and write to one specific external service — cloud storage, project management tools, messaging platforms, and so on. Once connected, the AI can act on that service directly.

AnalogyHooking up a "phone line" to an external system — you can only call once the wire is connected.

02.05

Tool

A single capability the AI can call on — "search the web," "read a file," "run code." A Tool is the smallest unit; many Tools combine to form more complex abilities.

AnalogyAn individual implement in an employee's hand — a hammer, a ruler, a phone — each does one thing.

02.06

Skill

A pre-written set of instructions that teaches the AI the full process for a particular type of task. "Publish a social media post," for example, is a Skill containing steps for writing copy, choosing images, and posting.

AnalogyA standard operating procedure (SOP) an employee learns — every time this kind of work comes up, they follow the steps.

02.07

Workflow

A chain of tasks linked together to run automatically. For example: "Customer inquiry comes in → AI analyzes it → reply is generated → email is sent" — fully automatic, no human in the middle.

AnalogyA factory assembly line: raw materials in one end, finished products out the other, everything between handled automatically.

02.08

n8n

An open-source automation workflow tool you can self-host — free, flexible, and approachable. A drag-and-drop canvas lets you design flows that connect to hundreds of services and AI models.

AnalogyA tool for designing your own "automated assembly line": email arrives → data is sorted → spreadsheet is updated → notification is sent — all automatic.

02.09

Multimodal

An AI that doesn't just understand text but also images, audio, and video. A "modality" is a type of information; a multimodal AI can handle several types at once.

AnalogyThe old AI could only read books; the new one can also look at pictures, listen to sound, and watch videos.

02.10

MCP

An open protocol that lets AI models connect to external tools and services safely. Think of it as a USB standard — every tool plugs into the same kind of socket.

AnalogyThe "USB port" for AI — any tool that follows the MCP standard can plug straight in.

Telling them apart

Plugin / Connector / Tool / Skill — what's the difference?

Term	Analogy	What it contains	Example
Plugin	A whole skill kit	Connector + Skill + Tool	A "design Plugin" with design-tool connections and skills
Connector	A phone line	One external service	Connecting to cloud storage
Tool	A single instrument	One capability	The "search the web" function
Skill	An SOP	A flow of multiple Tools	The full "publish a social post" process

Chapter 03Training

AI Training Terms

How does an AI actually "learn"? Here's the training pipeline, from reading the entire internet to being graded by human teachers.

Training pipeline

The four stages of AI training

Stage 01

Pre-training

Reads almost every book — extremely expensive.

Stage 02

Fine-tuning

Specialized training for a particular task.

Stage 03

RLHF

Humans grade the answers to make the AI more useful.

Stage 04

Inference

Deployment — this is when you actually chat with the AI.

10 terms

03.01

Pre-training

The AI's "K–12 schooling." It's trained on enormous amounts of general data (essentially the whole internet) so the model picks up the basic patterns of language. Eye-wateringly expensive — tens or hundreds of millions of dollars.

AnalogySpending more than a decade in school, from kindergarten through university, building a foundation.

03.02

Fine-tuning

Taking a pre-trained model and training it further on specific data so it gets better at a particular domain or style. Cheaper than pre-training, but still resource-intensive.

AnalogyHiring a generalist chef and training them only in sushi — they go from all-rounder to sushi specialist.

03.03

RLHF

Real human reviewers rate the AI's answers — "this one's good," "that one's bad" — and the model learns to improve from those scores. Most modern AI uses this method to become more helpful and safer.

AnalogyA student hands in essays, the teacher grades them, and the student writes better next time based on the feedback.

03.04

Parameters

The number of "neural connections" inside the AI's brain. More parameters generally means a smarter model — but also a heavier and more expensive one. Top-tier models reportedly have over a trillion parameters.

AnalogyThe connections between neurons in a human brain — the more connections, the more powerful the processing.

03.05

Weights

The actual numerical values of those parameters. Training is the process of nudging these numbers until the AI's answers get better and better. Weights are the form in which all of an AI's learned knowledge is stored.

AnalogyEvery problem a student solves earns "experience points" — Weights are all that accumulated experience.

03.06

Dataset

The AI's "textbook." Before training, you need huge amounts of labeled data — say, millions of images tagged "cat" or "dog," or billions of lines of text. The quality of the data directly shapes the quality of the AI.

AnalogyThe teaching materials a school gives its students — the richer and more accurate they are, the better the students learn.

03.07

Benchmark

A "standardized test" for AI, so different models can be compared on the same questions. Common ones include MMLU (breadth of knowledge) and HumanEval (coding ability).

AnalogyA public exam — everyone takes the same test, so the scores are actually comparable.

03.08

Overfitting

The AI has memorized the training data too well and falls apart when it sees something new. It looks perfect on the practice questions but fails the moment you rephrase them.

AnalogyA student who rote-memorizes past exam papers — they ace the original questions, but reword them and they're stuck.

03.09

Quantization

"Slimming down" an AI model. You sacrifice a little precision in exchange for much faster inference — and the model can run on a regular computer or even a phone. It makes large models lightweight and practical.

AnalogyCompressing a 4K video to 1080p — slightly lower quality, much smaller file, plays just fine on a phone.

03.10

Distillation

A large model (the teacher) trains a smaller one (the student). After distillation, the small model can produce results close to the big one with far less compute, dramatically reducing cost.

AnalogyA senior professor distills a lifetime of knowledge for a young student, bringing the student close to the professor's level fast.

Three ways to improve an AI

Prompt Engineering vs RAG vs Fine-tuning

Method	Analogy	Cost	Best for
Prompt Engineering	Teaching staff how to ask better questions	Low · close to free	Quick wins on quality
RAG	Giving staff a reference library to consult anytime	Moderate	When the knowledge keeps changing
Fine-tuning	Putting staff through a full retraining program	High · expensive	When you need a specific style or deep expertise

Chapter 04Developer Tools

Developer Tool Terms

The tools and concepts developers use every day. Even if you don't write code, knowing these will make working with engineering teams much smoother.

12 terms

04.01

API

The bridge that lets two programs "talk" to each other. You don't need to step into the kitchen — you just order from the menu, and the kitchen (the AI) sends out the dish. Developers use APIs to tap into AI capabilities.

AnalogyThe order window at a restaurant — you say what you want, the chef makes it, and you don't need to know what's happening behind the counter.

04.02

SDK

A more complete toolkit than just an API — it bundles the API itself plus documentation, sample code, and helpers. Developers can get started faster and skip a lot of repetitive boilerplate.

AnalogyIf an API is the order window, an SDK is the whole package: window, menu, payment terminal — service in a single bundle.

04.03

Webhook

When something finishes, the system actively notifies your program. You don't have to keep asking "is it done yet?" — when it's done, a signal gets sent automatically. Common in automation pipelines.

AnalogyThe notification when your food delivery arrives — it pings you when it's there; you don't have to keep refreshing the app.

04.04

Hook

When a particular event happens, a piece of code automatically fires. For example, every time the AI finishes a reply, the conversation log is saved to a file. The system "hooks into" the action for you.

AnalogyA motion sensor at the door — someone walks in, the lights come on, no one has to flip a switch.

04.05

Endpoint

A specific "service address" inside an API. Each capability has its own endpoint — for instance, /chat is the entry point for chat, /image for image generation.

AnalogyDifferent shop addresses inside a mall — to visit a particular shop, you go to its address.

04.06

Rate Limit

A cap on how many times you can call an API per minute or per day, to prevent overload. Go over and the API temporarily refuses you.

AnalogyA restaurant rule that no one person can order more than 10 dishes per hour — so one customer can't hog every chef.

04.07

Streaming

Instead of waiting for the AI to think up the entire answer and dumping it all at once, the response appears word by word, like typing. Those characters tumbling onto your screen mid-chat? That's streaming. It cuts down on the feeling of waiting.

AnalogyThe chef serves each dish as it's ready — you don't have to wait until everything is plated before you start eating.

04.08

Async

After sending a request, your program doesn't sit and wait — it carries on with other work, and the AI nudges you when the answer is ready. The system can handle many things at once, much more efficient.

AnalogyPlace your order and then go do your shopping — they call you when the food is ready, instead of you standing at the counter.

04.09

IDE

A developer's "workshop" — code editor, debugger, and runner all in one. The most popular IDEs support virtually every language, and many now include built-in AI assistance.

AnalogyA chef's kitchen — stove, knives, ingredients all in one place, every step covered.

04.10

Package Manager

A helper that installs, updates, and manages third-party tools for you. Instead of hunting around websites, one command does it. Common ones include npm and pip.

AnalogyAn app store — search, tap install, done. No website hopping required.

04.11

API Key

Your "pass" for using an API service — a unique string of characters. Keep it secret; if it leaks, anyone can rack up charges on your account. Usually stored in a dedicated environment variable file.

AnalogyYour credit card number — hand it to someone and they can spend your money.

04.12

Sandbox

An isolated test environment where you can break things without affecting the real system. Developers can try new features safely, knowing real users won't be impacted if something goes wrong.

AnalogyKids playing in a sandbox — no matter how messy it gets, the rest of the world is fine.

Chapter 05Infrastructure

Infrastructure

The "foundations" beneath AI and software — servers, cloud, containers, deployment. Invisible, but holding everything up.

14 terms

05.01

Terminal

That black screen where you type commands. It's the window for "talking directly" to your computer — keyboard only, no mouse needed. Far faster than a graphical interface for many tasks.

AnalogyTelling the chef what you want directly, instead of tapping through a screen menu — quicker and more flexible.

05.02

Shell

The Terminal is the "window"; the Shell is the "translator" inside it. The Shell takes the commands you type and translates them into something the operating system understands. Common shells include bash and zsh.

AnalogyYou speak Cantonese; the translator turns it into a language the operating system understands.

05.03

SSH

Remotely controlling another computer over the network — securely and encrypted. Sitting in Hong Kong, you can SSH into a server in the US and operate it as if it were right in front of you.

AnalogyOperating a computer in the US through an encrypted phone line — every word is scrambled to outsiders.

05.04

VPN

An "encrypted tunnel" built across a public network. Your traffic appears to come from somewhere else, which protects your privacy and lets you bypass regional restrictions.

AnalogyA hidden private tunnel underneath a public road — outsiders can't see what car you're driving or where you're going.

05.05

Port

A server's "room number." A single computer can run many services at once, and Ports are how you tell them apart. 80 = website, 443 = secure website, 22 = SSH, 3000 = local development.

AnalogyOne building with many rooms — Ports are the room numbers.

05.06

Server

A computer that runs 24/7 to provide a service. Your website, your API, your database all live on a server. It can be a physical machine or a virtual one in the cloud.

AnalogyA restaurant kitchen that never closes — there's always someone there ready to take an order.

05.07

Cloud

Servers managed by someone else — you pay for what you use and skip buying hardware. The big three providers are AWS, GCP, and Azure.

AnalogyRenting warehouse space instead of building your own — take as much as you need, return it when you're done.

05.08

Container

Packaging an application together with everything it needs to run, so it works the same anywhere. Solves the classic "but it works on my machine" problem.

AnalogyA shipping container — pack everything inside, and any ship in any port can move it.

05.09

Docker

The most popular Container tool, used to package, distribute, and run containers. A staple in every developer's toolbox — it makes deployment standardized and simple.

AnalogyThe factory that builds shipping containers — same standard, same dimensions, fits anywhere.

05.10

Environment

All the settings an application needs to actually run. There are usually three: Dev (developer's local testing) → Staging (pre-launch dress rehearsal) → Production (the real thing, live to users).

AnalogyDev is the chef tasting in the back; Staging is inviting friends and family for a tasting; Production is opening the restaurant to the public.

05.11

Deploy

Moving your finished code onto a server so users can actually use it. CI/CD makes deployment automatic — submit code, and the system tests, packages, and deploys it without manual steps.

AnalogyTurning a draft recipe into the official printed version that the kitchen actually cooks from.

05.12

Cron Job

A task scheduled to run automatically at set times. For example, back up the database every night at 2 a.m., or clean up old files every hour — no human required.

AnalogyA pre-set alarm clock that does something on cue, no manual triggering.

05.13

Log

The system's automatically recorded "diary," noting every action and error. When something breaks, you read the logs to figure out why — like reviewing security camera footage after an incident.

AnalogyThe black box on a plane — it records everything during the flight so investigators can piece together what happened.

05.14

Checkpoint

AI training can run for days; the system saves progress at intervals. If something crashes, you can resume from the last Checkpoint instead of starting over.

AnalogyThe save point in a video game — you can save mid-level, and if you die, you respawn at the save point.

Chapter 06Version Control

Version Control

A "time machine" for code — every change is recorded, you can rewind anytime, and multiple people can collaborate without stepping on each other.

8 terms

06.01

Repository

A "folder" that holds all the code for a project — but with the full history of every change. Like a document's version history, except a hundred times more powerful.

AnalogyA super folder with version control — every edit is logged, and you can roll back any time.

06.02

Branch

Like a parallel universe. You leave the main code untouched, open a new branch to try a feature, and either merge it back when it works, or just delete the branch — no harm done.

AnalogyA road that forks off the main route — you take the new road to see where it goes, while everyone else keeps cruising on the main road.

06.03

Commit

A "save," with a note attached (like "Fix login bug"). Strung together, commits form the complete history of changes.

AnalogyHitting "save" and writing a memo — later you can see exactly what you changed and when.

06.04

Push / Pull

Push = "upload" your local changes to the cloud Repo. Pull = "download" the latest cloud code to your machine. The bread and butter of team collaboration.

AnalogyPush is uploading to a shared cloud drive; Pull is downloading the latest version your teammates uploaded.

06.05

Pull Request

You've made some changes — now you "request" that they be merged in, and a teammate reviews them. They merge it only if they're happy. It's the standard way teams keep code quality high.

AnalogySubmitting a report to your boss for sign-off — it only goes out once they've stamped it approved.

06.06

Merge

Combining the code from two branches. If two people changed the same lines at the same time, the system can't decide which to keep — that's a "Merge Conflict," and it has to be resolved by hand.

AnalogyTwo streams flowing into one river — if they're heading in different directions, they collide.

06.07

GitHub

A platform for hosting Repos online so teams can collaborate. The world's largest gathering place for open-source projects. Similar platforms include GitLab and Bitbucket.

AnalogyThe cloud drive for code — everyone parks their code here and works on it together.

06.08

README

The "instruction manual" of the Repo — the first thing anyone sees when they open a project. It explains what the project does, how to install it, and how to use it.

AnalogyThe leaflet inside a product's box — what it is, how to use it.

Chapter 07Evaluation

Eval — Evaluation

How do you know if an AI has actually improved? How do you put numbers on its performance? Without Eval, you have no idea whether your change was a step forward or a step back.

Why Eval matters

"It feels better" isn't enough

You tweak a Prompt and the AI seems to answer better — but "seems" isn't proof. You need data showing it actually got better, not just on a couple of cases while everything else regressed. Eval is the discipline of getting that proof.

6 terms

07.01

Eval

Giving the AI a "test" to measure how good and how accurate its answers are. It's a core part of AI development — every time you change a model or a Prompt, you run Evals to check the impact.

AnalogyEvery time you change your teaching method, you give the students a test to see whether their grades actually went up.

07.02

Human Eval

Real people score the AI's answers, judging "good" or "bad." The most accurate method, but also the slowest and most expensive. RLHF training depends on huge amounts of Human Eval.

AnalogyA teacher hand-grading essays — most accurate, but the teacher's time is limited and costly.

07.03

LLM-as-Judge

Letting another AI (usually a stronger model) score the answers. Faster and cheaper than Human Eval, but it carries its own biases and isn't fully objective.

AnalogyAn automatic essay-grading system — quick and cheap, but not as careful as a teacher.

07.04

Binary Eval

The simplest form of grading: right or wrong, no partial credit. Fast and easy to automate — perfect for checking whether the AI did some specific thing or not.

AnalogyPass/Fail grading — it only matters whether you cleared the bar, not by how much.

07.05

Accuracy

The proportion of questions the AI got right. The most intuitive evaluation metric. If the AI answers 100 questions and gets 85 correct, accuracy is 85%.

AnalogyA test score — the share of questions answered correctly.

07.06

Benchmark

An industry-standard exam, used to rank different AI models on the same set of questions. It's what gives "this model is better than that one" some hard data behind it.

AnalogyA public exam — everyone takes the same paper, so scores can actually be compared.

The Eval loop

How Eval fits into AI development

Step 01

Tweak

Adjust the Prompt or the model parameters.

Step 02

Run Eval

Have the AI take the standard test set.

Step 03

Check scores

Compare before-and-after results.

Step 04

Decide

Keep the change, or keep iterating.

Step 05

Loop

Keep optimizing until you hit the bar.

Three kinds of "checking"

Test vs Eval vs Benchmark

Term	Question it asks	Angle	Example
Test	"Does it run? Is anything broken?"	Engineering	Does the button click? Are there bugs?
Eval	"How well does it run? Is the quality good?"	AI research	Are the answers accurate? Any hallucinations?
Benchmark	"How does it stack up against others?"	Comparison and ranking	What did it score on MMLU?

Appendix AGlossary

Quick Glossary

Every term at a glance, instantly searchable. 65+ entries, organized by topic.

Search · 65+ terms

AI

Artificial intelligence — letting computers imitate human thinking

LLM

Large language model — the core of mainstream chat AI

Token

The smallest unit of text the AI processes

Context Window

The maximum content the AI can hold in mind at once

Prompt

The instruction or question you give the AI

System Prompt

A hidden background instruction the user never sees

Temperature

The dial that controls how creative the AI is

RAG

Letting the AI look things up before answering

Hallucination

The AI confidently inventing information

Embedding

Turning words into number coordinates

Agent

An AI that can complete multi-step tasks on its own

Copilot

An AI that assists you while you stay in charge

Plugin

A bundle of capabilities that extends the AI

Connector

A bridge that links the AI to an external service

Tool

A single capability the AI can call

Skill

A pre-defined instruction set for an AI workflow

Workflow

Multiple tasks chained to run automatically

MCP

A standard protocol for AI to connect to tools

n8n

An open-source automation workflow tool

Multimodal

An AI that handles text, images, and audio together

Pre-training

The initial training of an AI on huge amounts of data

Fine-tuning

Further training a model for a specific task

RLHF

Improving the AI by learning from human ratings

Parameters

The number of neural connections in the AI's brain

Weights

How the AI's learned knowledge is stored

Dataset

The data used to train the AI

Benchmark

A standardized test for measuring AI performance

Overfitting

The AI memorizes training data and fails on new questions

Quantization

Compressing an AI model so it's lighter to run

Distillation

A large model teaching a smaller one

API

The bridge that lets programs talk to each other

SDK

A developer toolkit that's more complete than an API

Webhook

A mechanism that pings you when something is done

Hook

Code that fires automatically when an event happens

Endpoint

The service address of a specific API capability

Rate Limit

A cap on how many API calls you can make in a window

Streaming

The AI outputs word by word instead of waiting

Async

Don't wait for the response — keep doing other things

IDE

A developer's workshop

API Key

The password for using an API — keep it secret

Sandbox

An isolated test environment

Terminal

The window where you type commands to the computer

Shell

The program inside the Terminal that translates commands

SSH

Encrypted remote control of another computer

VPN

An encrypted tunnel built across a public network

Port

The room number for different services on a server

Server

A computer that provides a service 24/7

Cloud

Servers managed by someone else

Container

An isolated package containing an app and its environment

Docker

The most popular container tool

Deploy

Moving code to a server so users can use it

CI/CD

Automated testing and deployment pipelines

Cron Job

A task that runs automatically on a schedule

Log

The system's automatically recorded run-time diary

Checkpoint

A save point during the training process

Git

A version-control tool for code

Repository

A repo holding code and its full history

Branch

A parallel-universe branch of the code

Commit

A single saved record of code changes

Pull Request

A review request to merge a branch

Merge

Combining the code from two branches

GitHub

The largest code-hosting and collaboration platform

README

The instruction file inside a repo

Eval

A test for the quality of an AI's answers

Human Eval

Real people scoring the AI's answers

Binary Eval

An evaluation with only right or wrong outcomes

Accuracy

The share of questions the AI gets right

Prompt Engineering

The craft of writing better AI instructions

Inference

The AI in use — answering questions

Latency

The waiting time before the AI responds