The AI Dictionary by Mundhra Labs
Every AI word, explained simply.
Plain words, a picture for every term, and as much depth as you want. 221 words so far, from ChatGPT to transformers.
Type a word, a question or a whole sentence. We find the right words.
Start here
New to AI words? Short paths that teach them in the right order, about ten minutes each.
Understand ChatGPTThe words that explain what happens when you chat with an AI.10 words →Use AI tools wellGet better answers and avoid the common traps.9 words →How machines learn, from zeroWhat training really means, step by step.10 words →AI that acts: agentsFrom chatbots that talk to agents that do things.9 words →How AI makes pictures, voice and videoThe ideas behind image, video and voice tools.9 words →Is AI safe? The key wordsThe words people use when they argue about AI risk.9 words →
Tools
Two other ways in, when you would rather not search.
The basics
The words everything else is built on.
- Artificial intelligenceAIArtificial intelligence is the effort to make computers do things that would need thinking if a person did them, like seeing, talking, planning and learning.
- Machine learningMLMachine learning is a way of building software where the computer works out its own rules from examples, instead of a person writing every rule.
- Deep learningDLDeep learning is machine learning with neural networks stacked in many layers, where each layer builds on the last to spot more complex patterns.
- Neural networkA neural network is a web of simple maths units, loosely inspired by brain cells, that learns by strengthening or weakening the links between them.
- AlgorithmAn algorithm is a precise, step-by-step recipe for solving a problem, written clearly enough that a computer can follow it without guessing.
- AI modelAn AI model is the trained part of an AI system: a file of learned numbers that turns an input, like a photo or a question, into an output.
- TrainingTraining is how an AI model learns: it makes a guess, checks how wrong it was, and nudges its numbers to do a little better, millions of times over.
- InferenceInference is using a trained AI model: feeding it something new, like your question or a photo, and letting it work out an answer.
- ParametersParameters are the adjustable numbers inside an AI model, set during training, that hold everything it has learned. Bigger models have more of them.
- Training dataTraining data is the collection of examples an AI model learns from. Whatever patterns, gaps and mistakes it contains, the model tends to pick up.
- Generative AIGenerative AI is AI that creates new things, like text, images, music or code, by learning the patterns in huge numbers of examples and producing fresh ones.
- Artificial general intelligenceAGIAGI, or artificial general intelligence, is the idea of an AI that can learn and do almost any thinking task a person can, not just one kind.
- Narrow AINarrow AI is AI built to do one kind of task, such as playing chess or recognising faces, often extremely well, but nothing outside it.
- SuperintelligenceASISuperintelligence is a hypothetical AI that would be far smarter than the best humans at nearly everything, from science to strategy to persuasion.
- Symbolic AIGOFAISymbolic AI is the older approach to AI in which people write knowledge as explicit symbols and rules, and the computer reasons with them step by step.
- Expert systemAn expert system is a program that captures a human specialist's know-how as hundreds of if-then rules and uses them to give advice in one narrow field.
- Turing testThe Turing test asks whether a machine can chat so much like a person that a judge, typing to both, cannot reliably tell which is which.
- AI winterAn AI winter is a stretch of years when big promises about AI went unmet, so money and interest dried up and research slowed.
- Data scienceData science is the craft of turning messy data into answers: asking a good question, cleaning the numbers, finding the pattern and explaining what to do.
- Big dataBig data means collections of data so large, fast or varied that one ordinary computer cannot store or analyse them, so the work is shared across many machines.
- Predictive AIPredictive AI learns patterns from past data to estimate something unknown, like tomorrow's sales or whether a payment is fraud, and gives a number or a likely answer.
- RoboticsRobotics is the engineering of machines that sense the world, decide what to do and then move, from factory arms to robot vacuum cleaners.
- Physical AIPhysical AI is AI with a body: models that take in what cameras and sensors see and turn it into movements of a robot arm, a humanoid or a car.
How machines learn
What actually happens when we say a model is trained.
- Supervised learningSupervised learning teaches a computer with examples that already have the right answer attached, until it can answer new ones by itself.
- Unsupervised learningUnsupervised learning gives a computer data with no answers attached and lets it find the hidden groups and patterns by itself.
- Reinforcement learningRLReinforcement learning trains an AI by trial and error: it tries actions, gets rewarded or penalised, and slowly learns what works best.
- Self-supervised learningSSLSelf-supervised learning lets AI learn from raw data with no human labels, by hiding part of the data and practising guessing it back.
- Labelled dataLabelled data is a set of examples that each come with the right answer attached, like photos tagged "cat" or emails marked "spam".
- FeatureA feature is one measurable detail a model looks at to make a prediction, like a home's size, age or distance from the metro.
- ClassificationClassification is when an AI sorts each input into one of a set of categories, like spam or not spam, cat or dog, fraud or fine.
- RegressionRegression is when an AI predicts a number, like a price, a temperature or a delivery time, by finding the pattern linking inputs to that number.
- ClusteringClustering is when an AI sorts items into groups of similar things on its own, without being told what the groups should be.
- Loss functionA loss function is the scorecard of training: one number that says how wrong a model's guesses are, so training can push it down.
- Gradient descentSGDGradient descent is how most AI models learn: check which way reduces the error, take a small step that way, and repeat.
- BackpropagationBackpropagation is how a neural network learns from a mistake: it traces the error backwards and works out how much each connection was to blame.
- OverfittingOverfitting is when an AI memorises its practice examples, quirks and all, so it looks perfect in training but fails on anything new.
- EpochAn epoch is one complete pass through all the training data. Training usually runs several epochs, so the model sees each example more than once.
- Learning rateThe learning rate sets how big a step a model takes each time it corrects itself during training: too small is slow, too big overshoots.
- Activation functionAn activation function is the small rule each artificial neuron applies to its total before passing it on, adding the bend that lets networks learn complex patterns.
- Artificial neuronAn artificial neuron is the tiny building block of a neural network: it weighs a few inputs, adds them up, and passes on a signal.
- Test setA test set is data kept locked away during training, so we can fairly check how well a model works on examples it has never seen.
- Transfer learningTransfer learning reuses what a model learned on one big task as a head start on a new task, so it needs far less data and time.
- Synthetic dataSynthetic data is made-up data that is generated to look like real data, used to train or test AI when real data is scarce, private or risky.
- Recurrent neural networkRNNA recurrent neural network reads a sequence one piece at a time, like words in a sentence, carrying a running memory of what came before.
- Decision treeA decision tree predicts by asking a short chain of yes or no questions, like a flowchart the computer learned from examples.
- Random forestA random forest is a crowd of decision trees, each trained on a different random slice of the data, that vote on the answer.
- Gradient boostingGradient boosting builds a strong model from many small decision trees, added one at a time, each fixing the mistakes the others still make.
- HyperparameterHyperparameters are the settings people choose before training starts, like how big each learning step is, while the model learns everything else itself.
- RegularisationRegularisation is any trick that stops a model from memorising its training examples, so it learns the general pattern and does well on new data.
- DropoutDropout randomly switches off some parts of a neural network at every training step, so no part can lean too hard on another, and the network memorises less.
- Batch sizeBatch size is how many training examples a model looks at before it adjusts itself once.
- Precision and recallPrecision asks how many of the model's alarms were right. Recall asks how many of the real cases it caught.
- Confusion matrixA confusion matrix is a table that shows, for every real category, what the model guessed, so you can see exactly which things it mixes up.
- Recommendation systemA recommendation system is the software that picks what to show you next, a video, song, product or post, by learning from what you and people like you chose before.
- Anomaly detectionAnomaly detection is AI that learns what normal looks like, then flags anything that does not fit, like a strange card payment at 3 a.m.
- Federated learningFederated learning trains one shared AI model across many phones or organisations, sending the model to the data instead of collecting the data in one place.
Inside language models
How ChatGPT-style models read, think and write.
- Large language modelLLMA large language model, or LLM, is an AI that has read a huge amount of text and writes by predicting a likely next word, again and again.
- TokenA token is a small chunk of text, often a word or part of a word, that an AI model reads and writes. It is the unit AI counts in.
- TokenizationTokenization is how an AI cuts text into small pieces called tokens. Common words become one piece, rare words get split into several.
- Context windowA context window is how much text an AI can keep in view at once. Anything that falls outside it, the AI simply cannot see.
- EmbeddingAn embedding turns a word, sentence or image into a list of numbers, so that things with similar meaning end up close together.
- VectorA vector is a list of numbers that can be drawn as an arrow. AI turns words and pictures into vectors, then compares their directions.
- TransformerA transformer is the design behind ChatGPT, Claude and Gemini. It reads a whole passage at once and works out which words matter to each other.
- AttentionAttention is how an AI decides which other words to focus on when it reads each word, so it can understand that word in context.
- Next-token predictionNext-token prediction is how chatbots write: they guess the most fitting next piece of text, add it, and repeat, one piece at a time.
- Pre-trainingPre-training is the first, biggest stage of building an AI model: it reads a vast amount of text and learns by guessing each next word.
- Fine-tuningFine-tuning takes an AI model that is already trained and teaches it a specific job or style with a much smaller set of examples.
- RLHFRLHF trains a chatbot using people's preferences: humans pick the better of two answers, and the model learns to give answers people prefer.
- Instruction tuningInstruction tuning teaches a raw language model to follow requests, by training it on many examples of an instruction paired with a good answer.
- TemperatureTemperature is a setting that controls how adventurous an AI is when picking its next word: low for safe and predictable, high for varied and surprising.
- Foundation modelA foundation model is one large AI model trained on broad data that many different apps and tools are then built on top of.
- Multimodal AIMultimodal AI can work with more than one kind of input or output at once, such as text, photos, voice and video, in a single conversation.
- Reasoning modelA reasoning model is an AI that works through a problem step by step, privately, before it gives its answer. It is slower, but better at hard problems.
- Chain of thoughtChain of thought means an AI writes out its working step by step before giving the answer, which helps it get multi-step problems right.
- Mixture of expertsMixture of experts is a way to build a huge AI model where only a few of its many parts switch on for each word, saving effort.
- DistillationDistillation trains a small AI model to imitate a big one, keeping much of the big model's skill in a model that is cheaper and faster to run.
- QuantizationQuantization shrinks an AI model by storing each of its billions of numbers less precisely, so it needs less memory and runs faster, with a small loss in quality.
- Natural language processingNLPNatural language processing, or NLP, is the part of AI that helps computers read, understand and produce everyday human language, written or spoken.
- Small language modelSLMA small language model is a compact version of the AI behind chatbots, small enough to run on a phone or laptop instead of a data centre.
- LoRALoRA is a cheap way to customise a big AI model: it freezes the model and trains a tiny add-on instead, often under 1% of the size.
- KV cacheA KV cache is a chatbot's saved notes on every word it has already read, so it does not redo that work each time it writes the next word.
- Positional encodingPositional encoding stamps each word with its place in the sentence, so an AI that reads all words at once still knows their order.
- Encoder-decoderAn encoder-decoder is an AI built in two halves: the encoder reads the whole input and understands it, the decoder writes the output from that understanding.
- Sentiment analysisSentiment analysis is AI that reads a piece of text and judges the feeling in it: positive, negative, neutral or mixed.
- Machine translationMachine translation is a computer turning text or speech from one language into another, like Google Translate turning Hindi into English.
- SummarisationSummarisation is AI shortening a long text, like an article, report or meeting, into a few lines that keep the main points.
- LogitsLogits are the raw scores an AI model gives each possible answer, before they are turned into percentages that add up to 100.
- Reward modelA reward model is an AI judge: it learns from people's choices which answers they prefer, then scores new answers so a chatbot can learn from them.
- Direct preference optimisationDPODPO teaches a chatbot straight from pairs of answers, one preferred and one not, nudging it towards the kind of answer people chose.
- Constitutional AIConstitutional AI trains a chatbot using a written list of principles: the AI checks its own answers against the list and learns from the corrections.
- Speculative decodingSpeculative decoding speeds up AI writing: a small, fast model guesses the next few words, and the big model checks them all in one go.
- Model cardA model card is a short fact sheet that comes with an AI model, saying what it is for, how it was tested, and where it falls short.
Using AI
The words you meet when you actually use AI tools.
- PromptA prompt is whatever you type or give to an AI to tell it what you want. The AI's answer is built from it.
- Prompt engineeringPrompt engineering is the skill of wording your request to an AI so it reliably gives you what you actually need.
- System promptA system prompt is a set of standing instructions an AI reads before your message, setting its role, rules and tone for the whole chat.
- Few-shot promptingFew-shot prompting means showing an AI a few worked examples inside your prompt, so it copies the pattern for your new case.
- ChatbotA chatbot is a computer program you talk to in ordinary language, by typing or speaking, and it replies like a person would.
- HallucinationAn AI hallucination is when a chatbot states something false, like a made-up fact or source, in the same confident tone it uses for the truth.
- Retrieval-augmented generationRAGRAG lets an AI look up the relevant pages in a set of documents first, then write its answer from what it found.
- Vector databaseA vector database stores things as lists of numbers that capture their meaning, so it can quickly find the items most similar to a question.
- AI agentAn AI agent is an AI you give a goal, not just a question. It plans steps, uses tools, checks results and keeps going until the job is done.
- Tool useTool use lets an AI model ask for outside help, like a search, a calculator or a booking system, and then use the result in its answer.
- Model Context ProtocolMCPMCP is a shared standard that lets any AI app plug into any tool or data source the same way, like one charging port for everything.
- APIAn API is a set doorway through which one program asks another for something, using an agreed format, and gets a reply back.
- AI assistantAn AI assistant is an AI you ask, by voice or text, to help with everyday things, and it can act for you, not just answer.
- AI memoryAI memory lets a chatbot carry useful facts about you, like your diet or your job, from one conversation into the next.
- Context engineeringContext engineering is deciding exactly what information an AI gets to read for a task: the right instructions, files, history and tool results, and nothing that distracts.
- Vibe codingVibe coding is making software by describing what you want to an AI, trying what it builds, and asking for changes, without reading the code yourself.
- Computer useComputer use means an AI operating a computer the way you do: looking at the screen, moving the mouse, clicking and typing.
- Deep researchDeep research is an AI mode that spends several minutes searching and reading many sources on your question, then writes a report with citations.
- Knowledge cutoffA knowledge cutoff is the date an AI's training reading stopped. It knows nothing that happened after it, unless it looks things up.
- Semantic searchSemantic search finds results that mean the same as your question, even when they share none of its words.
- ChunkingChunking means cutting long documents into small, self-contained pieces so an AI can search them and read only the part that matters.
- Structured outputStructured output makes an AI fill in a fixed form, with exactly the fields and format you asked for, instead of writing free text.
- Multi-agent systemA multi-agent system splits one big job between several AI agents, usually a lead agent that hands out parts and combines the results.
- Human in the loopHITLHuman in the loop means a person checks, approves or can stop what an AI does, instead of letting it act completely on its own.
- AI workflow automationAI workflow automation is a set chain of steps between your apps that runs by itself, with AI handling the parts that need reading or writing.
- AI search engineAn AI search engine searches the web for you, reads the top pages and writes a short answer with links, instead of giving you a list to open.
- Voice modeVoice mode lets you talk to an AI chatbot out loud and hear it answer, back and forth, almost like a phone call.
- AI browserAn AI browser is a web browser with an AI assistant built in that can read the pages you open and click through websites for you.
- Prompt cachingPrompt caching lets an AI service reuse the work it already did on the unchanged start of a prompt, making repeat requests faster and cheaper.
- Agent skillsAgent skills are packaged how-to guides, with instructions and sometimes code, that an AI agent opens only when a task calls for them.
Images, voice and video
How AI sees, hears, draws and speaks.
- Computer visionCVComputer vision is the part of AI that teaches computers to make sense of photos and video: what is in them, where, and what is happening.
- Convolutional neural networkCNNA convolutional neural network (CNN) looks at a picture through small sliding windows, finding edges first, then shapes, then whole objects.
- Image recognitionImage recognition is AI that looks at a picture and tells you what is in it, and often exactly where.
- Diffusion modelA diffusion model makes images by starting from pure random noise and removing a little of it, step by step, until a picture appears that matches your words.
- GANA GAN is two AI models trained as rivals: one makes fakes, the other tries to catch them, and both keep getting better.
- Text-to-imageText-to-image AI turns a sentence you type into a brand new picture that did not exist before.
- Text-to-videoText-to-video AI turns a sentence you type into a short, brand new video clip, frame by frame, with everything moving consistently.
- Speech recognitionSpeech recognition is AI that listens to spoken words and writes them down as text.
- Text-to-speechTTSText-to-speech is AI that reads written words aloud in a voice that sounds human.
- DeepfakeA deepfake is a video, photo or voice clip made or changed with AI so that a real person seems to do or say something they never did.
- OCROCR, optical character recognition, turns a photo or scan of words into real text you can copy, search and edit.
- Latent spaceLatent space is the compressed inner map a model builds, where each point stands for the essence of a picture, sound or sentence.
- Facial recognitionFacial recognition is AI that identifies or checks a person by turning their face into a pattern of numbers and comparing it with saved ones.
- Vision-language modelVLMA vision-language model is AI that can look at an image and talk about it: describe it, read it, or answer your questions about it.
- Image-to-imageImage-to-image AI takes a picture you give it, plus a few words, and makes a new picture that keeps the layout of yours.
- InpaintingInpainting is AI that fills in a part of a picture you mark, so it matches everything around it: removing a stranger, or adding something new.
- UpscalingAI upscaling makes a small or blurry picture bigger and sharper by guessing the fine detail that should be there.
- Image segmentationImage segmentation is AI that colours in every pixel of a picture by what it belongs to, so each object gets its exact outline.
- Music generationAI music generation turns a few typed words, like "a monsoon love song with tabla", into a new piece of music, often with singing.
- 3D generation3D generation is AI that makes a 3D object or scene, one you can turn and view from any side, from words, a photo or a short video.
- CLIPCLIP is an AI model from OpenAI that learned which captions go with which pictures, so it can measure how well any words match any image.
- ControlNetControlNet is an add-on for image generators that makes them follow a guide you give, like a stick-figure pose or an outline, not just your words.
- AI avatar and lip syncAn AI avatar is a video of a face, real or invented, that AI makes speak any words, with lips moving in sync with the voice.
- Object trackingObject tracking is AI that follows the same object through a video, frame after frame, and remembers which one is which.
Safety and society
Where AI goes wrong, and how people try to keep it right.
- AI biasAI bias is when an AI system treats some groups of people unfairly, usually because it learned from data that was already skewed.
- AlignmentAlignment is the work of making an AI pursue what people actually intend, not just the literal goal or score it was given.
- AI safetyAI safety is the work of making sure AI systems do not cause harm, whether by mistake, through misuse, or as they grow more capable.
- JailbreakA jailbreak is a trick that gets a chatbot to break its own safety rules, usually by disguising a request it would normally refuse.
- Prompt injectionPrompt injection is when text an AI reads, like an email or web page, contains hidden instructions that hijack what the AI does.
- GuardrailsGuardrails are the checks and rules wrapped around an AI model that stop it from receiving or giving harmful, off-topic or unsafe content.
- ExplainabilityExplainability is about answering "why did the AI decide that?", by showing which inputs mattered or what is happening inside the model.
- Red teamingRed teaming means deliberately attacking an AI system before release, trying to make it misbehave, so its flaws can be found and fixed.
- AI regulationAI regulation is the laws and rules governments set for how AI can be built and used, from outright bans to labelling AI content.
- AI watermarkingAI watermarking hides an invisible signal in AI-made text, images, audio or video, so software can later tell it was made by AI.
- AI and copyrightAI and copyright is the legal fight over two questions: may AI learn from people's work without permission, and who owns what AI makes?
- AI and privacyAI and privacy is about what happens to personal information when AI is trained on it or when you share it with an AI tool.
- SycophancySycophancy is when an AI tells you what you want to hear, agreeing or praising you, instead of what is true or useful.
- AI ethicsAI ethics asks which uses of AI are fair and right, who could be harmed, and who answers for it when things go wrong.
- AI and jobsAI can now do some tasks inside many jobs. Whether that mostly changes jobs or removes them is one of the big open questions of our time.
- Existential riskExistential risk is the worry that future AI, far more capable than today's, could cause a catastrophe that ends or permanently cripples humanity.
- MisinformationMisinformation is false information shared as if it were true. AI makes it cheaper and faster to produce, and harder to tell from the real thing.
- AI detectorAn AI detector is a tool that guesses whether text, an image or audio was made by AI. It gives a probability, not proof, and it can be wrong.
- Data poisoningData poisoning is slipping bad examples into the data an AI learns from, so the finished model behaves wrongly, often only in hidden, specific situations.
- Adversarial exampleAn adversarial example is an input changed very slightly, on purpose, so that an AI model gets it badly wrong while a person sees no difference.
- AI governanceAI governance is the whole system for steering AI: laws, standards, company policies and international agreements, plus the people and bodies that apply them.
- Responsible AIResponsible AI is the practice of building and using AI with checks for fairness, safety, privacy and accountability at every step, not just good intentions.
- Model collapseModel collapse is what happens when AI models are trained, generation after generation, on content made by earlier AI: rare details vanish and output drifts from reality.
Chips and scale
The hardware and numbers behind modern AI.
- GPUA GPU is a chip with thousands of small workers that all calculate at once. AI needs exactly that kind of teamwork.
- TPUA TPU is a chip Google designed to do one thing: the grid multiplications at the heart of AI, as fast and cheaply as possible.
- ComputeIn AI, compute means calculating power: how many sums your chips can do, and for how long. More compute usually buys a more capable model.
- Scaling lawsScaling laws are a steady pattern: give an AI model more compute, data and size, and its error falls by a predictable amount.
- BenchmarkA benchmark is a fixed test, the same questions and the same marking for every AI model, so their scores can be compared fairly.
- Data centerA data center is a huge building full of computers. AI data centers pack in tens of thousands of AI chips and the power and cooling they need.
- On-device AIOn-device AI runs the model on your own phone or laptop, so your request never has to leave it, even without internet.
- NPUAn NPU is a small AI engine built into your phone or laptop chip. It runs AI features quickly while using very little battery.
- FLOPsFLOPA FLOP is one basic sum with decimal numbers. Counting FLOPs measures how much work an AI job takes and how fast a chip works.
- Open-weight modelAn open-weight model is an AI model whose trained numbers anyone can download, run and adapt, unlike a model you can only use through an app.
- CUDACUDA is Nvidia's software that lets programmers hand big piles of maths to a GPU, so thousands of its cores work on them at once.
- HBMHBM is a stack of memory chips placed right beside an AI chip, so the chip can be fed numbers fast enough to stay busy.
- AI acceleratorASICAn AI accelerator is a chip built to do the maths of AI much faster, and with less power, than an ordinary processor.
- Inference costInference cost is what it costs to run an AI model each time it answers, usually counted in tokens, the small chunks of text it reads and writes.
- LatencyLatency is the waiting time between asking an AI something and seeing its answer begin, and then finish.
- Tokens per secondTokens per second is how fast an AI model writes: how many small chunks of text it produces each second.
- Training clusterA training cluster is thousands of AI chips wired together so tightly that they can train one model as if they were a single giant computer.
- AI energy useAI energy use is the electricity, and the water, that AI chips and their data centers consume to train models and answer people.
- Chip export controlsChip export controls are government rules that limit which countries and companies may buy the most powerful AI chips and the machines that make them.
- Cloud AICloud AI means renting AI computing power or ready-made AI models over the internet, instead of buying and running the chips yourself.
Names you hear
The products and companies, and what they actually are.
- ChatGPTChatGPT is an app: a chatbot made by OpenAI. The thinking happens in a model called GPT that runs inside it.
- GPTGPT-4GPT is a family of large language models made by OpenAI. ChatGPT is the app you chat with; GPT is the engine inside it.
- ClaudeClaude is a family of AI models made by the company Anthropic, and also the name of the chat app you use them in.
- GeminiGemini is Google's family of AI models, and also the name of Google's chatbot app, which used to be called Bard.
- LlamaLlama is a family of large language models from Meta that anyone can download and run on their own computers.
- Microsoft CopilotCopilot is Microsoft's name for the AI assistant built into its products, from Windows and Word to Outlook and the Edge browser.
- MidjourneyMidjourney is a service that turns a written description into pictures, made by a small independent company of the same name.
- DALL-EDALL·EDALL-E was OpenAI's family of models that made pictures from written descriptions. It was retired in 2026, replaced by GPT Image.
- Stable DiffusionSDXLStable Diffusion is a family of image-making AI models from Stability AI that anyone can download and run on their own computer.
- SoraSora was OpenAI's AI that turned a written description into a short video clip. OpenAI shut the Sora app and service down in 2026.
- DeepSeekDeepSeek is a Chinese AI company. The name also covers its free chatbot app and its family of open-weight language models.
- GrokGrok is an AI chatbot and model family first made by xAI, Elon Musk's AI company. It lives inside the social network X and in its own app.
- PerplexityPerplexity is an AI search app. It looks up web pages for your question, then writes one answer with numbered links to its sources.
- Mistral AIMistral AI is a French AI company. It builds large language models, releasing many as open weights, plus its own chat app.
- OpenAIOpenAI is an American AI company. It builds the GPT family of models and the ChatGPT app that runs on them.
- AnthropicAnthropic is an American AI company. It makes the Claude family of AI models and the Claude app.
- Google DeepMindGoogle DeepMind is Google's main AI research lab. It builds the Gemini models and science systems such as AlphaGo and AlphaFold.
- NvidiaNVIDIANvidia is an American chip company. It makes GPUs, processors widely used to train and run AI models, plus the software to program them.
- Hugging FaceHugging Face is a website and company where people share AI models and datasets, a bit like an app store or GitHub for AI.
- AlphaGoAlphaGo is a program from DeepMind that learned to play the board game Go and, in 2016, beat one of the world's best players.
- MicrosoftMicrosoft is an American technology company. In AI it makes the Copilot assistant, rents out computing to AI makers, partners with OpenAI and builds its own models.
- Meta AIMeta AI is the AI assistant built into WhatsApp, Instagram, Facebook and Messenger, made by Meta, the company that owns them.
- Apple IntelligenceApple Intelligence is not an app. It is Apple's name for the AI features built into the iPhone, iPad and Mac software.
- Amazon (Alexa+ and Bedrock)Amazon's AI reaches people in two main ways: Alexa+, the assistant in Echo speakers, and Bedrock, a cloud service that rents AI models to businesses.
- QwenQwen is a family of large language models made by Alibaba, the Chinese technology group. Many of them can be downloaded and run by anyone.
- CursorCursor is a code editor app with AI built in. You describe what you want in plain words, and AI writes or changes the code for you.
- Sarvam AISarvam AI is an Indian AI company in Bengaluru. It builds AI models that understand and speak Indian languages, for chat, voice and documents.
- NotebookLMNotebookLM is a Google app where you upload your own notes, PDFs or videos, and an AI answers questions using those sources.
- ElevenLabsElevenLabs is an AI audio company. Its models turn written text into natural-sounding speech, copy voices and dub videos into other languages.
- RunwayRunway is an AI company and web app for making and editing video. You describe a shot, or start from a photo, and it generates the clip.
- SunoSuno is an AI app that makes complete songs, with lyrics, music and singing voices, from a short written description.
- AlphaFoldAlphaFold is an AI system from Google DeepMind that predicts the 3D shape of a protein from the list of building blocks it is made of.
#
A
- Activation function
- Adversarial example
- Agent skills
- AI accelerator
- AI agent
- AI and copyright
- AI and jobs
- AI and privacy
- AI assistant
- AI avatar and lip sync
- AI bias
- AI browser
- AI detector
- AI energy use
- AI ethics
- AI governance
- AI memory
- AI model
- AI regulation
- AI safety
- AI search engine
- AI watermarking
- AI winter
- AI workflow automation
- Algorithm
- Alignment
- AlphaFold
- AlphaGo
- Amazon (Alexa+ and Bedrock)
- Anomaly detection
- Anthropic
- API
- Apple Intelligence
- Artificial general intelligence
- Artificial intelligence
- Artificial neuron
- Attention