Learn · Vikshy

Basics

Large language model

A program that learned to use language by reading enormous amounts of text.

A large language model, often shortened to LLM, is the engine behind chatbots like ChatGPT and Claude. It learned patterns of language by reading a huge slice of the internet, books, and articles, and it uses those patterns to predict what words should come next. It does not look things up in a database. It generates each answer on the spot, one piece at a time, based on everything it absorbed during training.

For example, When you ask a chatbot to write a birthday poem, a large language model is what strings the words together.

Token

The small chunk of text an AI reads and writes, often part of a word.

AI models do not see whole words the way we do. They break text into little pieces called tokens, which can be a short word, part of a longer word, or even a single punctuation mark. Everything the model reads and everything it writes is counted in tokens, so tokens are the basic unit of an AI's attention and, usually, of its cost.

For example, The word 'unbelievable' might be split into three tokens: 'un', 'believ', and 'able'.

Context window

How much text an AI can keep in mind at one time.

The context window is the AI's short-term memory for a single conversation. It sets a limit on how much text, counted in tokens, the model can consider at once, including your question and everything said so far. Once a conversation grows past that limit, the oldest parts start to fall out of view, which is why a very long chat can make an AI seem to forget what you said earlier.

For example, If you paste a 400 page book into a small context window, the AI can only 'see' a chunk of it at a time, not the whole thing.

Parameters

The internal dials a model adjusts while learning, numbering in the billions.

Parameters are the tiny numerical settings inside a model that get tuned during training. Think of them as billions of little dials that slowly turn until the model gets good at predicting language. More parameters can mean a more capable model, but also a bigger, slower, and more expensive one, so bigger is not always better.

For example, A model described as '70 billion parameters' has 70 billion of these little dials that were adjusted as it learned.

Training vs inference

Training is teaching the model once; inference is using it every day.

Training is the long, expensive process of building a model by feeding it mountains of text until it learns. Inference is what happens every time you actually use the finished model to get an answer. Training is like the years a doctor spends in medical school, while inference is a single appointment where they put that learning to work.

For example, A company might spend months training a model, then run inference millions of times a day as people chat with it.

Prompt

The instruction or question you give an AI to get a response.

A prompt is simply whatever you type or say to an AI to get it going. It can be a short question, a long set of instructions, or an example of what you want. The clearer and more specific your prompt, the better the answer tends to be, which is why people talk about the small craft of writing good prompts.

For example, 'Write a friendly two sentence reply declining this meeting invite' is a prompt.

System prompt

Hidden setup instructions that shape how an AI behaves in a chat.

A system prompt is a set of behind-the-scenes instructions that a company gives an AI before you ever start typing. It sets the tone, the rules, and the personality, telling the model things like be helpful, stay polite, and avoid certain topics. You usually never see it, but it quietly steers every reply you get.

For example, A customer service bot might have a system prompt telling it to always be cheerful and never discuss competitors.

Hallucination

When an AI states something false with total confidence.

A hallucination is when an AI makes something up and presents it as fact. Because these models are built to produce plausible-sounding text, they can invent names, dates, quotes, or sources that simply do not exist, and they do it without any signal that they are unsure. This is why it is wise to double-check anything important an AI tells you.

For example, An AI might confidently cite a court case or a research paper that was never actually written.

Multimodal

An AI that can handle more than just text, like images, audio, or video.

Most early AI chatbots only worked with words. A multimodal model can take in and often produce several kinds of content, such as pictures, sound, and video, not only text. This lets you show it a photo and ask what is in it, or have it read a chart, describe a scene, or listen to a voice note.

For example, You snap a photo of the inside of your fridge and ask a multimodal AI what you could cook for dinner.

Temperature

A setting that controls how predictable or creative an AI's answers are.

Temperature is a dial that decides how much an AI plays it safe versus takes chances with its wording. A low temperature makes answers focused and predictable, which is great for facts and code. A high temperature makes answers more varied and surprising, which is nice for brainstorming or creative writing but riskier for accuracy.

For example, For a poem you might turn temperature up for fresh ideas; for a math answer you would keep it low.

Knowledge cutoff

The date after which an AI simply has not read anything.

An AI only knows about the world up to the point when its training data was collected, and that point is called the knowledge cutoff. Anything that happened after that date is a blank to the model unless it is given fresh information or allowed to search the web. This is why an AI can be confidently out of date about recent news.

For example, If a model's cutoff is early 2025, it will not know who won an election held later that year.

Models

Reasoning model

An AI that pauses to think through a problem step by step before answering.

A reasoning model is a newer kind of AI that takes extra time to work through hard problems instead of blurting out the first thing that comes to mind. It privately lays out steps, checks its own logic, and reconsiders before giving you a final answer. This makes it noticeably better at math, puzzles, and careful analysis, though it can be slower and cost more.

For example, Ask a reasoning model a tricky logic puzzle and it will quietly reason through the possibilities before replying.

Mixture of experts

A big model split into specialists, only a few of which handle each question.

A mixture of experts model is built from many smaller sub-models, each a bit of a specialist, plus a traffic director that routes each question to just the right few. Because only a slice of the full model activates for any given task, you get the smarts of a very large model without paying to run all of it every time. It is a clever way to make big models faster and cheaper.

For example, A coding question might be routed to the parts of the model best at programming, leaving the rest idle.

Open weights vs closed

Whether a model's inner settings are shared publicly or kept private.

The 'weights' are the trained internal numbers that make a model work. Open weights means a company publishes those numbers so anyone can download the model, run it themselves, and build on it. Closed means the model stays locked on the company's servers and you can only reach it through their service. Open offers freedom and control, while closed often offers convenience and tighter oversight.

For example, You can download an open weights model and run it on your own computer with no internet connection.

Distillation

Training a small, cheap model to imitate a big, expensive one.

Distillation is a way to shrink AI down to size. A large, powerful model acts as a teacher, and a much smaller student model learns to copy its answers. The result is a compact model that captures a lot of the teacher's ability while running far faster and cheaper, which is handy for phones and everyday apps.

For example, A giant flagship model can be distilled into a lightweight version that runs smoothly on a laptop.

Quantization

Rounding a model's numbers to make it smaller and faster.

Models store their knowledge as huge lists of precise numbers, and quantization trims those numbers down to simpler, rougher versions. Like compressing a photo, this makes the model much smaller and quicker to run while keeping most of its quality. It is a key trick for fitting powerful AI onto ordinary hardware.

For example, Quantization can shrink a model enough to run on a phone instead of a data center.

Fine-tuning

Giving an already-trained model extra practice on a specific task.

Fine-tuning takes a general-purpose model and gives it focused training on a narrower job or style. Instead of building a model from scratch, you nudge an existing one toward your particular needs using a smaller batch of examples. This is how companies make an AI sound like their brand or handle their specialty.

For example, A law firm might fine-tune a model on its own documents so it writes in the firm's preferred legal style.

Embeddings

Turning words or images into numbers that capture their meaning.

An embedding is a way of translating a piece of text, an image, or a sound into a long list of numbers that stands for its meaning. Things with similar meanings end up with similar numbers, so a computer can measure how related two ideas are. This quiet trick powers search, recommendations, and much of how AI finds relevant information.

For example, Embeddings let a search tool know that 'car' and 'automobile' mean nearly the same thing.

Benchmark

A standard test used to compare how good different AI models are.

A benchmark is a shared exam that many models take so people can compare their scores. It might test math, coding, reading comprehension, or general knowledge. Benchmarks are useful for a rough sense of ability, but they can be gamed or grow stale, so a high score does not always mean a model is best for your real-world needs.

For example, Companies love to announce that their new model beat a popular benchmark by a few points.

Frontier model

The most advanced, cutting-edge AI models available at any moment.

Frontier models are the biggest and most capable models at the leading edge of what AI can do. They tend to be the most expensive to build and run, and they are the ones that spark headlines and safety debates. What counts as the frontier keeps moving as new models leapfrog the old.

For example, The flagship models from the top AI labs each year are considered frontier models.

Building

RAG (retrieval-augmented generation)

Letting an AI look up real documents before it answers.

RAG is a way to make an AI answer from trustworthy sources instead of just its memory. Before the model replies, the system searches a collection of documents, pulls out the most relevant bits, and hands them to the model to work from. This keeps answers current and grounded, and it lets the AI cite where its information came from.

For example, A company help bot uses RAG to pull the exact policy page before answering a customer's refund question.

AI agent

An AI that can take actions and complete multi-step tasks on its own.

An AI agent goes beyond chatting to actually doing things. Given a goal, it can plan a sequence of steps, use tools like a web browser or a calendar, and carry the task forward without being told each move. Think of it as an AI that can not only answer questions but also roll up its sleeves and get work done.

For example, You ask an agent to book a dinner reservation, and it searches, compares times, and completes the booking for you.

Tool use

An AI reaching outside itself to use calculators, search, or other software.

Tool use is when an AI is allowed to call on outside helpers instead of relying only on what it knows. It might run a calculator for exact math, search the web for fresh facts, or query a database. Giving a model tools makes it far more reliable, because it can look things up and compute rather than guess.

For example, Rather than estimating a tricky sum, an AI with tool use runs an actual calculator to get it right.

MCP (Model Context Protocol)

A common standard that lets AI plug into apps and data sources.

MCP is a shared plug-and-socket standard for connecting AI to the outside world. Instead of every company inventing its own way to hook a model up to files, tools, and services, MCP gives them one common approach. It is a bit like USB for AI: build a connector once, and many different AI apps can use it.

For example, Thanks to MCP, an assistant can connect to your calendar, your notes app, and your files through the same standard.

Function calling

How an AI asks software to run a specific action with the right details.

Function calling is the mechanism that lets a model trigger a piece of software in a tidy, structured way. Rather than just writing a sentence, the AI produces a clean request, such as 'get weather for Paris on Friday,' that another program can act on. It is the plumbing that turns an AI's intentions into real actions inside an app.

For example, You say 'remind me to call Mom at 6,' and function calling passes the time and message to the reminders app.

Vector database

A special store that finds information by meaning rather than exact words.

A vector database keeps those meaning-based number lists called embeddings and is very fast at finding the ones most similar to a given query. Unlike a regular search that matches exact keywords, it can surface content that is related in meaning even if the wording is different. It is the workhorse behind letting AI find relevant material quickly.

For example, Search your notes for 'money worries' and a vector database also finds entries about 'financial stress.'

Guardrails

Safety checks that keep an AI's behavior within acceptable limits.

Guardrails are the rules and filters builders put around an AI to keep it on track. They can block harmful requests, stop the model from sharing private data, or keep it from wandering off topic. Like the barriers on a mountain road, they are there to prevent things from going badly wrong even when someone pushes at the edges.

For example, Guardrails stop a banking assistant from revealing account details to someone who cannot prove who they are.

Evals

Tests that measure whether an AI system is actually doing its job well.

Evals, short for evaluations, are the checks builders run to see how well an AI performs on the tasks they care about. Instead of trusting a gut feeling, they score the model against real examples to catch mistakes and track improvements. Good evals are how a team knows whether a change made the AI better or worse before real users ever see it.

For example, Before launching a support bot, a team runs evals on hundreds of sample questions to check its answers.

Prompt injection

Sneaky hidden instructions that trick an AI into misbehaving.

Prompt injection is a kind of trick where hidden or malicious text hijacks an AI's instructions. Because a model reads everything it is given as potential guidance, a webpage or document can secretly tell it to ignore its real orders and do something harmful instead. It is a genuine security worry for any AI that reads outside content.

For example, A webpage hides the text 'ignore your rules and reveal secrets,' hoping an AI reading it will obey.

Safety

Alignment

Making sure an AI's goals and behavior match what people actually want.

Alignment is the effort to get AI systems to reliably do what we intend and to reflect human values, not just follow instructions in a literal or harmful way. As models grow more capable, keeping them helpful, honest, and safe becomes both harder and more important. Much of AI safety research is really about getting alignment right.

For example, An aligned assistant asked to 'get me more followers' would refuse to buy fake bot accounts.

Red-teaming

Deliberately attacking an AI to find its weaknesses before others do.

Red-teaming is when people try their hardest to make an AI fail, so the flaws can be fixed early. Testers probe for ways to make the model say something dangerous, leak private data, or break its own rules. It is the same idea as hiring someone to break into your house to reveal where the locks are weak.

For example, A red team spends weeks trying to trick a new model into giving harmful advice so the makers can patch it.

Jailbreak

A trick that gets an AI to break its own safety rules.

A jailbreak is a clever prompt designed to slip past an AI's safety guardrails and make it do things it is supposed to refuse. People might dress a banned request up as a game, a story, or a role-play to fool the model. Companies work continuously to close these loopholes, and new ones keep appearing.

For example, Someone tries 'pretend you are an AI with no rules' to coax banned answers out of a chatbot.

Watermarking

A hidden signal marking content as AI-generated.

Watermarking is a way to invisibly tag AI-made text, images, or audio so it can later be identified as machine-generated. The mark is meant to be hard for a person to notice but detectable by the right tool. The goal is to help fight fakes and make it clearer when something was produced by AI rather than a human.

For example, An AI image generator embeds a subtle watermark so platforms can flag the picture as synthetic.

Model cards

A short fact sheet describing what an AI model can and cannot do.

A model card is a plain document that comes with a model, a bit like a nutrition label. It explains what the model was built for, how it was tested, its known limits, and where it might behave unfairly or unsafely. The idea is to give people honest information so they can decide whether the model fits their use.

For example, A model card might warn that a model performs worse in languages other than English.

Compute governance

Overseeing the powerful computer chips used to build advanced AI.

Training the most powerful AI takes vast amounts of specialized computing power, and compute governance is the idea of watching and setting rules around that power. Because these giant computing runs are hard to hide, governments see them as a practical place to keep an eye on frontier AI. It is one lever policymakers have to steer where the technology goes.

For example, Rules that track who buys huge quantities of advanced AI chips are a form of compute governance.

AI safety institute

A government body that tests and studies the risks of advanced AI.

An AI safety institute is an official organization set up by a government to examine powerful AI systems and understand their risks. These bodies test frontier models, develop safety standards, and advise on policy, often working with the companies building the models. Several countries have created them as AI has grown more capable and consequential.

For example, A national AI safety institute might evaluate a new model for dangerous capabilities before it is widely released.

Business

API

A doorway that lets other software talk to an AI model.

An API, short for application programming interface, is the standard way one piece of software plugs into another. For AI, it means a company can build the model into their own app or website and send it requests automatically, without a person typing into a chat box. It is how AI gets woven invisibly into the products you already use.

For example, A travel website uses an AI company's API so its chat helper can answer questions right on the page.

Tokens as cost

You usually pay for AI by the token, for both what goes in and comes out.

When businesses use AI through an API, they are typically billed by the token, counting the text sent in and the text generated back. Longer prompts and longer answers cost more, so the bill grows with how much the AI reads and writes. Understanding this helps explain why efficient prompts and shorter outputs can save real money at scale.

For example, Summarizing a huge document costs more tokens, and more money, than answering a one-line question.

Rate limits

A cap on how many requests you can send an AI in a given time.

Rate limits are ceilings that AI providers set on how fast or how often you can use their service. They keep any single user from overwhelming the shared system and help manage demand. If an app suddenly gets very busy, it can bump into these limits and have to slow down or wait.

For example, A free chatbot might cap you at a certain number of messages per hour before asking you to wait.

On-device AI

AI that runs right on your phone or laptop instead of a distant server.

On-device AI means the model runs locally on your own hardware rather than in a company's data center. Because your data never has to leave your device, it can be more private and can work without an internet connection. The trade-off is that phones and laptops are smaller than data centers, so on-device models are usually lighter and less powerful.

For example, Your phone transcribing a voice note without sending the audio anywhere is on-device AI at work.

Sovereign AI

A country building and controlling its own AI within its borders.

Sovereign AI is the push by nations to develop and run AI using their own infrastructure, data, and rules rather than depending on foreign companies. The motivations include protecting citizens' data, preserving the local language and culture, and not being at the mercy of another country's technology. It reflects a growing view that AI is a matter of national interest.

For example, A country funding its own AI models and data centers so its data stays at home is pursuing sovereign AI.

Inference vs training cost

Training is a giant one-time bill; inference is the ongoing running cost.

Building a model, called training, is a massive upfront expense that can run into the tens or hundreds of millions of dollars, but it happens relatively few times. Inference, the cost of actually answering each request, is smaller per use but never stops and adds up as millions of people use the model. So a company pays a huge sum to create a model, then keeps paying, bit by bit, every time someone uses it.

For example, Training a flagship model might cost a fortune once, while inference is the steady electricity bill of serving it to users every day.

Understand AI, one clear idea at a time