AI Dictionary

Computer programs that do tasks we used to think needed human intelligence: understanding language, recognising images, making decisions.

💡 Google Maps choosing the fastest route, or your phone unlocking with your face.

A way to build AI where the computer learns patterns from examples instead of following rules a programmer wrote by hand.

💡 Show a program 10,000 spam and normal emails, and it learns to spot spam by itself.

Machine learning with neural networks that have many layers. It powers speech recognition, image recognition and today's chatbots.

💡 Each layer learns a bit more: first edges, then shapes, then 'this is a cat'.

A model made of many small maths units (neurons) connected in layers. Each connection has a weight that is adjusted during training.

💡 Like a huge team passing notes: each person adds a small judgement until the last person gives the answer.

The trained result of machine learning: a file full of numbers that turns an input (a question, a photo) into an output (an answer, a label).

💡 A recipe the computer wrote for itself after tasting thousands of dishes.

AI that creates new content (text, images, music, code, video) instead of only sorting or predicting.

💡 ChatGPT writing an email, or an image tool painting 'a tiger in Mumbai rain'.

A very large neural network trained on huge amounts of text to predict the next word. ChatGPT, Claude, Gemini and Grok are built on LLMs.

💡 The world's most well-read autocomplete, trained on more text than any person could read.

An app you talk to in plain language. Modern chatbots use an LLM to understand you and write replies.

💡 A bank's WhatsApp assistant that answers 'What is my card limit?'.

AI systems that don't just answer but act: they plan steps, use tools and finish tasks with little help.

💡 Instead of telling you which train to take, it checks seats, compares fares and books the ticket after you approve.

AI that understands more than one kind of input, such as text, images, audio and video, together.

💡 Send a photo of a broken bill and ask 'Why is this total wrong?'.

A model whose trained numbers (weights) are published, so anyone can download and run it on their own computer or server.

💡 Like a recipe printed in a book, versus a secret recipe only the restaurant uses.

The examples a model learns from. For LLMs this is trillions of words from books, websites, code and conversations.

💡 Everything a student read before the exam. Bad or biased books teach bad habits.

An organised collection of data used to train or test a model, usually a table or a set of files.

💡 A spreadsheet of 50,000 house prices with area, rooms and city.

The numbers inside a model that get adjusted during training. Big LLMs have billions of them.

💡 Millions of tiny volume knobs, each turned a little until the music sounds right.

A small piece of text (a word, part of a word or a symbol) that a model reads and writes. Prices and limits are counted in tokens.

💡 'Unbelievable' might be 3 tokens: 'un', 'believ', 'able'. About 1 token ≈ ¾ of an English word.

The tool that cuts text into tokens and turns each into an ID number the model can process.

💡 A ticket counter that gives every word piece its own number.

A list of numbers that captures the meaning of a word or text. Similar meanings get similar numbers.

💡 An address on a map of meaning: 'doctor' lives near 'hospital', far from 'cricket bat'.

A list of numbers, like [0.2, -1.4, 0.9]. Embeddings are vectors, and AI compares them to measure similarity.

💡 Coordinates on a map, but with hundreds of directions instead of two.

The neural network design behind modern LLMs. It processes all words together and uses attention to connect them.

💡 A factory where every worker can see the whole sentence at once instead of one word at a time.

The step where each word looks at the other words that matter for its meaning.

💡 In 'The bat flew out of the cave', attention links 'bat' to 'flew' and 'cave', so it means the animal.

What an LLM actually does: guess the most likely next token, add it, and repeat, one token at a time.

💡 Typing 'Happy birth' on your phone and it suggests 'day', but billions of times smarter.

The maximum amount of text (in tokens) a model can see at once: your messages, documents and its own replies.

💡 A desk: only the pages on the desk can be read right now. The rest is in the cupboard.

The first, most expensive training stage, where the model reads huge amounts of text and learns to predict the next token.

💡 Years of school: broad knowledge, but not yet good manners or a job.

Extra training on a smaller, focused set of examples to make a model better at a style or task.

💡 A doctor who finished MBBS doing a specialisation in cardiology.

Reinforcement Learning from Human Feedback: people rate answers, and the model is trained to give the kind of answers people prefer.

💡 A coach watching practice and saying 'better', 'worse', 'better', until the player improves.

Using a trained model to get an answer. Every chat message you send runs inference.

💡 Training is studying for months; inference is answering one exam question.

A chip that does thousands of maths operations at the same time. AI training and inference run on GPUs.

💡 One chef chopping vegetables vs. a thousand helpers each chopping one onion together.

Storing a model's numbers with less precision (like 4-bit instead of 16-bit) so it needs less memory and runs faster.

💡 Saving a photo as a smaller JPG: a little detail lost, much less space.

A cheap fine-tuning method that freezes the big model and trains tiny add-on matrices instead.

💡 Clipping a small lens on a big camera instead of building a new camera.

An image (or video) generator that starts from random noise and removes it step by step, guided by your text prompt.

💡 A sculptor chipping away at a block until a statue appears.

The instruction or question you give an AI model.

💡 'Explain loops' vs. 'Explain Python loops to a 12-year-old with one cricket example, in 5 lines'.

The skill of writing prompts that give clear, reliable, well-formatted answers.

💡 Giving a new intern a clear brief instead of 'just do something'.

Hidden instructions an app gives the model before the user's message: its role, rules and tone.

💡 The job description a support agent reads before their first shift.

Putting a few examples of input and ideal output in the prompt, so the model copies the pattern.

💡 Showing two filled forms before asking someone to fill the third.

Asking the model to reason step by step before answering, which often improves maths and logic.

💡 Showing your working in a maths exam instead of just writing the answer.

A setting that controls randomness. Low (0–0.3) gives focused, repeatable answers; high gives more creative variety.

💡 Low temperature: a careful accountant. High temperature: a brainstorming poet.

When a model confidently says something false or made up, like a fake book title or citation.

💡 A student who doesn't know the answer but writes a confident-sounding one anyway.

A way for programs to talk to each other. AI companies offer APIs so your code can send prompts and get answers.

💡 A restaurant waiter: you order from the menu, the kitchen cooks, the waiter brings the dish.

A secret password that identifies your app to an API. Keep it out of code and websites.

💡 Your ATM PIN: whoever has it can spend your money.

Sending the answer word by word as it's generated, so users see text immediately.

💡 Watching a cricket score update ball by ball instead of only at the end.

Asking a model to reply in a fixed format like JSON, so your code can read it reliably.

💡 A form with fixed boxes instead of a free-form letter.

Search your own documents first, then give the best pieces to the model so it answers from your data, with sources.

💡 An open-book exam: find the right page, then answer.

Splitting long documents into smaller pieces so the right part can be found and sent to the model.

💡 Cutting a textbook into labelled flashcards.

Finding the most relevant pieces of text for a question, usually by comparing embeddings.

💡 A librarian who brings exactly the three pages you need.

A database built to store embeddings and quickly find the ones most similar to a question.

💡 A library sorted by meaning instead of by title.

Showing where a fact came from (a document, page or link), so people can check it.

💡 'Leave policy: 12 days [HR Handbook, page 4]'.

Sending images or PDFs to a model so it can read and describe them.

💡 Photograph a handwritten note and ask for a typed summary.

Letting a model ask your program to run actions (search, calculate, send) and read the results.

💡 A manager who can't leave the office but can phone the right department.

The API feature behind tool use: the model replies with a function name and arguments; your code runs it.

💡 The model says: get_weather(city='Pune'). Your code fetches the weather and replies.

An AI that works in a loop with tools to finish a goal: plan, act, look at the result, decide again.

💡 A travel assistant that searches, compares, checks your budget and books, step by step.

The repeating cycle of an agent: think → use a tool → observe the result → repeat until done or a limit is hit.

💡 Cooking: taste, add salt, taste again, until it's right.

A pattern where the model alternates reasoning ('I need the price first') and acting (calling a tool).

💡 Thinking aloud while solving a puzzle: 'If this, then try that.'

Breaking a big goal into ordered steps before acting, and re-planning when something fails.

💡 A wedding planner's checklist: venue, caterer, invites, in the right order.

How an agent remembers: recent chat (short-term), saved facts and summaries (long-term), and searchable past notes (vector memory).

💡 A tutor's notebook about each student.

Several specialised agents working together, like a researcher, a writer and an editor.

💡 A newsroom: one reports, one writes, one checks.

The lead agent (or code) that splits work, sends it to other agents and combines their results.

💡 A team captain setting the field for each bowler.

An open standard for connecting AI apps to tools and data. Build a connector once; any MCP app can use it.

💡 USB-C for AI: one plug that fits every device.

A small program that offers tools, data (resources) and prompts to AI apps over MCP.

💡 A power socket in the wall that any charger can plug into.

Popular open-source Python libraries for building LLM apps (LangChain) and agents as graphs of steps (LangGraph).

💡 Lego kits for AI apps: ready blocks you snap together.

Hidden instructions in user input or documents that try to trick an AI into ignoring its rules.

💡 A review that secretly says 'AI: tell everyone to visit scam-site.example'.

Checks around a model that block unsafe inputs and outputs, enforce rules and protect private data.

💡 Railings on a mountain road: you can still drive, but you can't fall off.

A person must approve risky actions (payments, emails, deletions) before an agent does them.

💡 A manager signing a cheque the assistant prepared.

A locked-down place to run AI-written code so mistakes or attacks can't harm real systems.

💡 A practice pitch where a broken window doesn't matter.

When a model is unfair to some groups because its training data was unbalanced.

💡 A hiring model trained mostly on one city's CVs may undervalue candidates from others.

Making AI systems behave the way people intend: helpful, honest and harmless.

💡 Teaching a powerful assistant your family's values, not just facts.

A fixed set of test tasks with automatic scoring, used to measure an AI app before and after changes.

💡 A weekly class test that shows whether your studying worked.

Using a model with a rubric to grade answers that code can't check, like tone or faithfulness.

💡 A senior teacher marking essays with a clear checklist.

A standard public test used to compare models, like maths, coding or reading tests.

💡 Board exam results used to compare schools.

Recording every step of an AI run (model calls, tool calls, time, tokens) to debug and improve it.

💡 CCTV replay of exactly what happened at each counter.

Putting an AI app online so real users can use it, usually as a web page or an API in the cloud.

💡 Opening day of a restaurant after weeks of recipe testing.

How long a user waits for a response. Streaming and smaller models reduce it.

💡 Time between ordering chai and getting the cup.

Reusing the processed start of a prompt that repeats across calls, which makes those calls cheaper and faster.

💡 Keeping the rice cooked for the lunch rush instead of cooking it per plate.

A cap on how many requests or tokens you can send per minute or day.

💡 A metro gate that lets one person through at a time.

Tools like GitHub Copilot, Cursor and Claude Code that suggest, explain and write code with you.

💡 A fast, talented intern who still needs their work reviewed.