100 AI terms explained like a normal person
AI gets much easier once the language stops sounding mysterious.
This guide explains 100 useful terms you will hear while using ChatGPT, Claude, image generators, AI agents, coding tools, search systems, and machine-learning products. You do not need a technical background. Every term includes a simple explanation, a real-world example, and a quick way to remember it.
Made by Life of Arjav
Practical AI, business systems, tools, and workflows without confusing textbook language.
How to use this guide
- Read one section at a time.
- Pick five terms you did not know.
- Try explaining them to someone else without checking the definition.
- Save this page and return whenever an AI conversation starts sounding too technical.
You do not need to memorise all 100 terms. The goal is to recognise them, understand the basic idea, and know why they matter.
How AI Models Learn
Parameter
What it means: A value the model learns during training.
Real-world example: A spam filter learns thousands of internal values that help it recognise suspicious emails.
Remember it like this: The model’s learned settings.
Hyperparameter
What it means: A training setting chosen by people rather than learned by the model.
Real-world example: A team chooses the learning rate and batch size before training begins.
Remember it like this: Parameters are learned. Hyperparameters are chosen.
Activation function
What it means: A rule inside a neural network that decides how strongly a signal should move forward.
Real-world example: An image model detects an edge and decides whether that signal matters enough for the next layer.
Remember it like this: A gate that decides what information gets through.
Gradient descent
What it means: A method that improves a model by adjusting it in the direction that reduces errors.
Real-world example: A house-price model changes its internal values after each wrong prediction.
Remember it like this: The model walks downhill toward fewer mistakes.
Learning rate
What it means: The size of each adjustment a model makes while learning.
Real-world example: A rate that is too high may overshoot the answer, while one that is too low makes training slow.
Remember it like this: The size of each learning step.
Epoch
What it means: One complete pass through the full training dataset.
Real-world example: Training on 100,000 reviews five times means five epochs.
Remember it like this: The model has seen every example once.
Batch size
What it means: The number of examples processed before the model updates itself.
Real-world example: An image model may process 64 photos together before updating.
Remember it like this: One small stack of training examples.
Overfitting
What it means: When a model memorises training data too closely and performs poorly on new data.
Real-world example: A spam filter works perfectly on old emails but fails on new scam wording.
Remember it like this: It memorised the practice sheet instead of learning the lesson.
Underfitting
What it means: When a model is too simple or insufficiently trained to learn important patterns.
Real-world example: A sales model misses obvious seasonal patterns because it has not learned enough.
Remember it like this: It has not learned the lesson properly yet.
Regularization
What it means: Techniques that reduce overfitting and help a model generalise.
Real-world example: A team limits model complexity so it works better on new customer data.
Remember it like this: Learn the pattern without memorising the page.
AI Image Generation
Diffusion model
What it means: A generative model that creates images by starting with noise and gradually cleaning it up.
Real-world example: An image generator turns something like television static into a realistic portrait.
Remember it like this: Create by removing noise step by step.
Latent space
What it means: A compressed mathematical space where an AI represents important features and ideas.
Real-world example: A model can move an image toward ‘older’, ‘brighter’, or ‘more cinematic’ inside latent space.
Remember it like this: The AI’s map of visual ideas.
Denoising
What it means: Removing noise from an image representation until a clear image appears.
Real-world example: A diffusion model repeatedly cleans a noisy image until the requested scene becomes visible.
Remember it like this: The picture slowly comes into focus.
Seed
What it means: A starting number that influences random choices during generation.
Real-world example: The same prompt and seed can often recreate a very similar image.
Remember it like this: The starting roll of the dice.
Guidance scale
What it means: A setting controlling how closely an image follows the text prompt.
Real-world example: Higher guidance may follow ‘red jacket at night’ more literally, though too much can look unnatural.
Remember it like this: The prompt-obedience dial.
Negative prompt
What it means: A list of things you want the model to avoid.
Real-world example: You add ‘blurry, extra fingers, watermark’ to reduce unwanted results.
Remember it like this: Tell the model what should not appear.
Inpainting
What it means: Replacing or editing a selected part of an image while keeping the rest.
Real-world example: Select a mug and ask the model to replace it with a laptop.
Remember it like this: Fix one area without rebuilding everything.
Outpainting
What it means: Extending an image beyond its original borders.
Real-world example: A square portrait is expanded into a wide banner with a matching background.
Remember it like this: Continue the picture outside the frame.
ControlNet
What it means: A method that gives an image model structural guidance such as a pose, edge map, or depth map.
Real-world example: Provide a pose and generate a person standing in that exact position.
Remember it like this: A blueprint for the image model.
Image-to-image
What it means: Creating a new image from an existing image plus instructions.
Real-world example: Upload a room and ask the model to redesign it in a Japanese minimal style.
Remember it like this: The original image becomes the starting point.
AI Search and Knowledge
Semantic search
What it means: Search based on meaning rather than exact words.
Real-world example: ‘How do I get my money back?’ can find a page called ‘Refund policy’.
Remember it like this: Search for the idea, not just the phrase.
Cosine similarity
What it means: A measurement of how similar two vectors are in direction.
Real-world example: A system sees that ‘cancel my subscription’ is close to ‘end my membership’.
Remember it like this: Measures how similarly two ideas point.
Dense retrieval
What it means: Search that uses embeddings to find semantically similar content.
Real-world example: A support bot retrieves the right article even when the customer uses different wording.
Remember it like this: Meaning-based search.
Sparse retrieval
What it means: Search that relies mainly on exact words and their importance.
Real-world example: A rare product code is easier to find when exact matching matters.
Remember it like this: Strong when the words themselves matter.
Hybrid search
What it means: A combination of semantic and keyword search.
Real-world example: A legal tool searches by meaning while still prioritising an exact case number.
Remember it like this: Meaning plus matching.
Cross-encoder
What it means: A model that reads a query and result together to judge relevance.
Real-world example: It reranks 20 retrieved passages and selects the five most useful.
Remember it like this: The final judge for search results.
Knowledge graph
What it means: A network of facts and the relationships between them.
Real-world example: It stores that a person works at a company and that the company owns a product.
Remember it like this: A map of connected facts.
Ontology
What it means: A formal definition of the types of things in a domain and how they relate.
Real-world example: A healthcare ontology defines patients, doctors, conditions, and treatments.
Remember it like this: The rulebook for organising knowledge.
Entity extraction
What it means: Finding names, organisations, places, dates, amounts, or products inside text.
Real-world example: An invoice reader extracts the supplier, due date, invoice number, and total.
Remember it like this: Pull the important nouns from messy text.
OCR
What it means: Optical character recognition, which converts text in images into machine-readable text.
Real-world example: A receipt photo becomes editable words and numbers.
Remember it like this: It lets computers read text from pictures.
AI Agents and Reasoning
Planner
What it means: The part of an agent that breaks a goal into smaller steps.
Real-world example: For competitor research, it plans discovery, data collection, comparison, and reporting.
Remember it like this: Decides what should happen next.
Executor
What it means: The part of an agent that performs planned actions.
Real-world example: It searches websites, calls APIs, or enters data into a CRM.
Remember it like this: The planner plans. The executor works.
Critic
What it means: A component that reviews output and points out weaknesses.
Real-world example: It checks whether an email is too long, vague, or off-tone.
Remember it like this: The internal reviewer.
Reflection
What it means: When an AI reviews a result and uses the lesson to improve the next attempt.
Real-world example: A coding agent studies a failed test and changes its approach.
Remember it like this: Pause and learn from the last attempt.
ReAct
What it means: A method where an AI alternates between reasoning and using tools.
Real-world example: It decides what information it needs, searches, reads the result, then continues.
Remember it like this: Reason, act, observe, continue.
Chain-of-thought
What it means: The sequence of reasoning steps a model may use to solve a problem.
Real-world example: A model breaks a maths problem into smaller calculations before answering.
Remember it like this: The path of reasoning behind the answer.
Self-consistency
What it means: Trying a problem several ways and selecting the most consistent result.
Real-world example: A model solves a logic question five times and chooses the answer reached most often.
Remember it like this: Several attempts vote.
Tree of thoughts
What it means: Exploring several possible reasoning paths before choosing one.
Real-world example: A strategy model considers pricing, product, and distribution paths separately.
Remember it like this: Think through branches, not one straight line.
Planning horizon
What it means: How far ahead an agent considers future steps and consequences.
Real-world example: A short-horizon agent plans one move; a long-horizon agent plans an entire project.
Remember it like this: How many moves ahead it can see.
Computer use
What it means: AI operating software through a screen, keyboard, mouse, or browser.
Real-world example: An agent opens a CRM, updates a record, and sends a follow-up.
Remember it like this: The AI operates the computer.
AI Safety, Privacy, and Risk
Data contamination
What it means: When information that should remain separate enters training or evaluation data.
Real-world example: A benchmark question appears in training data and makes the model look better than it is.
Remember it like this: The test leaked into the study material.
Benchmark leakage
What it means: When benchmark questions or answers become available to the model before testing.
Real-world example: A coding model saw parts of its test online during training.
Remember it like this: The model may have seen the exam.
Adversarial example
What it means: An input designed to confuse a model while appearing normal to people.
Real-world example: A tiny change to a road-sign image causes a wrong classification.
Remember it like this: A normal-looking input with a hidden trick.
Data poisoning
What it means: Adding harmful or misleading examples to training data.
Real-world example: Fake reviews are inserted so a model learns to favour one product.
Remember it like this: Bad data teaches the wrong lesson.
Backdoor attack
What it means: A hidden behaviour that activates only when a specific trigger appears.
Real-world example: A model behaves normally until a certain symbol produces an attacker-chosen result.
Remember it like this: A secret switch inside the model.
Model inversion
What it means: An attack that attempts to reconstruct sensitive information from model outputs.
Real-world example: Repeated queries are used to approximate features of training faces.
Remember it like this: Work backwards to guess the original data.
Membership inference
What it means: An attack that tries to determine whether a record was used for training.
Real-world example: Someone tests whether one person’s medical record was in the dataset.
Remember it like this: Was this person in the training data?
Differential privacy
What it means: Adding controlled randomness so individual records are harder to identify.
Real-world example: A company studies customer trends without exposing one person’s behaviour.
Remember it like this: Learn about the crowd without exposing one person.
Federated learning
What it means: Training across many devices without collecting all raw data centrally.
Real-world example: Phones improve a keyboard model locally and send only model updates.
Remember it like this: The learning travels while raw data stays home.
Model card
What it means: A document explaining a model’s uses, testing, limits, and risks.
Real-world example: A company reads a hiring model’s card before deploying it.
Remember it like this: The model’s product label and instruction sheet.
Running AI in Real Products
Throughput
What it means: The amount of work an AI system can process in a given time.
Real-world example: A support model handling 500 requests per minute has more throughput than one handling 50.
Remember it like this: How much work gets through the pipe.
Rate limit
What it means: A cap on how many requests can be sent within a period.
Real-world example: An API allows 100 requests per minute and blocks the rest.
Remember it like this: A speed limit for API usage.
Context compression
What it means: Reducing a long conversation or document while keeping important information.
Real-world example: A chat summarises its first 100 messages so it can continue efficiently.
Remember it like this: Pack a large suitcase into a smaller one.
Prompt routing
What it means: Sending requests to different models or workflows based on the task.
Real-world example: Simple questions go to a cheap model and legal analysis goes to a stronger one.
Remember it like this: Send the request to the right lane.
Model fallback
What it means: A backup model used when the preferred model is unavailable or fails.
Real-world example: If the main model times out, the app automatically switches to another.
Remember it like this: The spare tyre of an AI product.
Batch inference
What it means: Processing many AI requests together.
Real-world example: A company summarises 10,000 customer reviews overnight.
Remember it like this: Handle a pile of tasks together.
Cold start
What it means: The delay while an AI service or model loads for its first request.
Real-world example: A rarely used model takes several seconds because it is not already in memory.
Remember it like this: The engine needs to warm up.
Token budget
What it means: The maximum number of tokens allowed for input, output, or reasoning.
Real-world example: A workflow limits each response to control cost and length.
Remember it like this: The word-piece spending limit.
Response schema
What it means: A fixed structure for the model’s answer.
Real-world example: A lead workflow requires name, company, title, and confidence fields.
Remember it like this: The form the AI must fill in.
Session state
What it means: Information remembered during an active interaction.
Real-world example: A shopping assistant remembers your size and budget while you browse.
Remember it like this: Temporary memory for the current session.
Different Ways Machines Learn
Supervised learning
What it means: Learning from examples that include the correct answer.
Real-world example: A model sees emails labelled spam or not spam.
Remember it like this: Learning with an answer key.
Unsupervised learning
What it means: Finding patterns in data without supplied labels.
Real-world example: A system groups customers by buying behaviour without predefined segments.
Remember it like this: Find structure without an answer key.
Semi-supervised learning
What it means: Learning from a small labelled dataset and a much larger unlabelled one.
Real-world example: A company combines 1,000 labelled tickets with 100,000 unlabelled tickets.
Remember it like this: A few labels guide a larger dataset.
Reinforcement learning
What it means: Learning through actions, rewards, and penalties.
Real-world example: A game-playing agent earns rewards for winning moves.
Remember it like this: Learn by trying and seeing what pays off.
Transfer learning
What it means: Reusing knowledge from one task as a starting point for another.
Real-world example: A general image model is adapted to identify damaged factory parts.
Remember it like this: Reuse old knowledge for a new job.
Classification
What it means: Predicting which category an input belongs to.
Real-world example: A support message is labelled complaint, refund, sales, or technical issue.
Remember it like this: Choose a label.
Regression
What it means: Predicting a numerical value.
Real-world example: A model estimates the selling price of a home.
Remember it like this: Predict how much, how many, or how long.
Clustering
What it means: Grouping similar items without predefined labels.
Real-world example: An AI groups leads by company size, industry, and behaviour.
Remember it like this: Find natural groups.
Feature engineering
What it means: Creating or selecting useful inputs for a model.
Real-world example: A churn model uses days since login and number of complaints.
Remember it like this: Turn raw data into useful clues.
Data labeling
What it means: Adding correct tags or answers to training examples.
Real-world example: People draw boxes around cars so a model learns object detection.
Remember it like this: Create the answer key.
Computer Vision
Computer vision
What it means: The field of teaching computers to understand images and video.
Real-world example: A warehouse camera identifies damaged packages.
Remember it like this: Give machines the ability to interpret what they see.
Object detection
What it means: Finding objects and locating each one with a box.
Real-world example: A driving system detects cars, cyclists, lights, and pedestrians.
Remember it like this: What is here, and where is it?
Image segmentation
What it means: Labelling individual pixels to show the exact shape of regions or objects.
Real-world example: A medical model marks the precise boundary of a tumour.
Remember it like this: Detection draws a box. Segmentation marks the exact area.
Pose estimation
What it means: Identifying the positions of key body joints.
Real-world example: A fitness app tracks knees, hips, and shoulders during a squat.
Remember it like this: Turn the body into a moving skeleton.
Optical flow
What it means: Estimating motion between video frames.
Real-world example: A sports system tracks player direction and speed.
Remember it like this: Measure motion inside video.
Depth estimation
What it means: Predicting how far objects are from the camera.
Real-world example: A robot uses it to avoid furniture.
Remember it like this: Give a flat image a sense of distance.
Visual grounding
What it means: Connecting words to the exact region of an image they describe.
Real-world example: Ask where the red bag is and the model points to it.
Remember it like this: Link language to a place in the image.
Image captioning
What it means: Generating a written description of an image.
Real-world example: An accessibility tool says, ‘A child is flying a kite on a beach.’
Remember it like this: Turn a picture into a sentence.
Face recognition
What it means: Comparing facial features to identify or verify someone.
Real-world example: A phone checks whether the face matches its authorised user.
Remember it like this: Do these faces belong to the same person?
Super-resolution
What it means: Creating a sharper, higher-resolution version of an image.
Real-world example: An old product photo is enlarged and reconstructed with more detail.
Remember it like this: Make a small blurry image clearer.
Measuring AI Performance
Precision
What it means: Of everything marked positive, how much was actually positive.
Real-world example: If 100 payments are flagged and 80 are fraud, precision is 80 percent.
Remember it like this: When it says yes, how often is it right?
Recall
What it means: Of all real positive cases, how many the model found.
Real-world example: If 100 fraud cases exist and 90 are caught, recall is 90 percent.
Remember it like this: How many real cases did we find?
F1 score
What it means: A score that balances precision and recall.
Real-world example: A fraud team uses it when false alarms and missed fraud both matter.
Remember it like this: Balance accuracy and coverage.
Confusion matrix
What it means: A table showing correct and incorrect predictions.
Real-world example: A spam filter table shows spam caught, spam missed, and normal mail wrongly flagged.
Remember it like this: The full prediction scoreboard.
False positive
What it means: A positive prediction that is actually wrong.
Real-world example: A legitimate payment is flagged as fraud.
Remember it like this: The alarm rings when nothing is wrong.
False negative
What it means: A missed case that was actually positive.
Real-world example: A screening model fails to detect a real condition.
Remember it like this: The alarm stays silent when something is wrong.
Calibration
What it means: How well confidence scores match actual accuracy.
Real-world example: Predictions labelled 80 percent confident should be correct about 80 percent of the time.
Remember it like this: Confidence should match reality.
Decision threshold
What it means: The cut-off that converts a probability into a final decision.
Real-world example: A fraud system flags payments above 70 percent risk.
Remember it like this: The line between yes and no.
Holdout set
What it means: Data kept separate until final testing.
Real-world example: A team trains on 80 percent and saves 20 percent for evaluation.
Remember it like this: The unopened exam paper.
Cross-validation
What it means: Testing several times with different training and validation splits.
Real-world example: A model is evaluated across five different splits.
Remember it like this: Take several practice exams, not one.
How Language Models Are Built
Tokenizer
What it means: The component that breaks text into tokens.
Real-world example: ‘Unbelievable’ may be split into smaller pieces before processing.
Remember it like this: Cut language into model-sized pieces.
Vocabulary
What it means: The complete set of tokens a model can recognise and generate.
Real-world example: It may include words, punctuation, numbers, and code fragments.
Remember it like this: The model’s box of language pieces.
Positional encoding
What it means: Information that helps a model understand token order.
Real-world example: It distinguishes ‘dog bites man’ from ‘man bites dog’.
Remember it like this: Tell the model where each word sits.
Encoder-only model
What it means: An architecture mainly designed to understand or classify input.
Real-world example: It can analyse whether a review is positive or negative.
Remember it like this: Built mainly for reading and understanding.
Decoder-only model
What it means: An architecture that generates text one token at a time.
Real-world example: Many chatbots and coding assistants use this design.
Remember it like this: Built mainly for generating and continuing text.
Encoder-decoder model
What it means: An architecture with one part that reads and another that generates.
Real-world example: A translator reads English with the encoder and writes Hindi with the decoder.
Remember it like this: One side reads. The other writes.
Autoregressive model
What it means: A model that generates each token using previous tokens.
Real-world example: It predicts one word piece, then uses it to predict the next.
Remember it like this: Write one step at a time.
Masked language model
What it means: A model trained to guess missing tokens.
Real-world example: Given ‘The capital of France is [MASK]’, it predicts Paris.
Remember it like this: Learn by filling in blanks.
Speculative decoding
What it means: A speed method where a smaller model drafts tokens and a larger model checks them.
Real-world example: A fast draft model proposes words that a stronger model accepts or rejects.
Remember it like this: A fast assistant drafts. A stronger editor approves.
Test-time compute
What it means: Extra computation used while answering to improve the result.
Real-world example: A reasoning model explores several solutions before returning one.
Remember it like this: Think longer at the moment of use.
Quick self-test
Pick any term from this guide and answer three questions:
- What problem does it solve?
- Where would I see it in a real AI product?
- How would I explain it to a friend in one sentence?
When you can answer all three without copying the definition, you understand the term well enough to use it.
Keep learning with Life of Arjav
I share practical AI terms, useful tools, Claude and ChatGPT workflows, business systems, automation ideas, and simple ways to keep up with AI.
Follow @lifeofarjav on Instagram
Share this guide with someone who keeps hearing AI words but has no idea what they mean.