OpenAI
ChatGPT is the product; GPT names the underlying model family.
ChatGPT · GPT · Sora
A broad, polished consumer and developer ecosystem across text, code, images, voice and video.
General assistance, coding, creative work and multimodal tasks.
Toolkit library
AI words and companies, explained without the jargon.
Look up the language behind AI, from LLMs and AGI to robotics and deepfakes. Then compare 30 leading model builders and AI ecosystems, what they make and where each is strongest.
Understand the language
Plain-English AI glossary
Start with the short definition, then open any term for a fuller explanation and a real-world example.
27 terms
FoundationsArtificial intelligence (AI)Computer systems designed to do tasks that normally need human-like intelligence.
AI is the umbrella term for software that can recognise patterns, understand language, make predictions, recommend actions or create content. It does not think or feel like a person. Most AI is built for a particular kind of task and learns patterns from examples rather than following only hand-written rules.
Example
Face unlock recognising you, Netflix suggesting a programme and ChatGPT drafting a story are all different uses of AI.
Related: Machine learning · Generative AI · AGI
How AI worksMachine learning (ML)A way of building AI by letting a computer learn patterns from data.
Instead of programming every rule, developers give a machine-learning system many examples. The system adjusts itself until it can make useful predictions about new examples. What it learns depends heavily on the quality and variety of its training data.
Example
Show a system thousands of labelled cat and dog photos and it can learn to classify a new photo.
Related: Training data · Model · Neural network
FoundationsGenerative AIAI that creates new text, images, audio, video or code from instructions.
Generative AI learns patterns in existing material and uses those patterns to produce a new output. It is not simply copying one item from its training data, but it can reproduce biases, mistakes or recognisable elements, so its work still needs human review.
Example
Ask for a poem about space, a picture of a robot garden or a first draft of computer code.
Related: LLM · Prompt · Multimodal AI
How AI worksLarge language model (LLM)An AI model trained on enormous amounts of language to predict and generate text.
An LLM breaks language into tokens and repeatedly predicts what should come next. This lets it answer questions, summarise, translate and write in many styles. It can sound confident without understanding truth in the human sense, which is why facts and sources must be checked.
Example
GPT, Claude, Gemini, Llama and Mistral are families of large language models.
Related: Token · Transformer · Hallucination
FoundationsArtificial general intelligence (AGI)A proposed future AI able to learn and perform most intellectual tasks at human level or beyond.
AGI has no universally agreed test or definition. Today’s systems can be remarkably capable across many tasks, but they still make basic errors and need human direction. Companies may use the term differently, so claims that AGI has arrived should be treated carefully.
Example
An AGI might learn medicine, design a bridge and plan a lesson without needing a separate specialist system for each task.
Related: AI · LLM · Agent
How AI worksAlgorithmA step-by-step set of instructions for solving a problem.
Algorithms are not automatically AI. A recipe and long division are algorithms too. In computing, algorithms tell software how to sort information, calculate a result or update a model while it learns.
Example
A route-planning algorithm compares possible roads to find a fast journey.
Related: Model · Machine learning
How AI worksModelThe learned pattern system that turns an input into a prediction or output.
Training creates a model by adjusting a huge collection of numerical settings called parameters. A finished model can then process new inputs. The app you use and the model behind it are not always the same thing: ChatGPT is a product that can use different GPT models.
Example
A weather model takes current measurements as input and predicts tomorrow’s conditions.
Related: Training data · Parameters · Inference
How AI worksTraining dataThe examples an AI system studies while learning patterns.
Training data may include text, pictures, audio, video, code or carefully labelled records. Missing, inaccurate or unfairly balanced data can produce weak or biased results. Training is different from a normal chat: entering information into a product does not necessarily mean it becomes immediate training data.
Example
A speech recogniser may train on recordings from many accents so it works for more people.
Related: Bias · Machine learning · Fine-tuning
How AI worksNeural networkA layered mathematical system that learns complex patterns from examples.
Neural networks are loosely inspired by connections in the brain, but they are not digital brains. Information passes through layers of numerical units, and training changes the connection strengths so useful patterns become more likely.
Example
A neural network can learn which shapes and textures tend to appear in photographs of bicycles.
Related: Deep learning · Transformer · Parameters
How AI worksDeep learningMachine learning using neural networks with many processing layers.
The word ‘deep’ refers to the number of layers, not deeper understanding. These systems can learn very complicated patterns and power modern image recognition, speech tools and language models, but they often need large amounts of data and computing power.
Example
Phone software that recognises objects in a photo commonly uses deep learning.
Related: Neural network · Machine learning
How AI worksTransformerThe neural-network design behind most modern language models.
Transformers use a mechanism called attention to judge which parts of an input are most relevant to one another. This helps a model follow relationships across a sentence or document and makes large-scale language training practical.
Example
In ‘The trophy did not fit in the suitcase because it was too big,’ attention helps connect ‘it’ with ‘trophy’.
Related: LLM · Context window · Attention
How AI worksTokenA small chunk of text an AI model reads and writes.
A token may be a whole short word, part of a longer word or punctuation. Models count tokens rather than pages or words, and usage limits and prices are often based on them. Roughly, 100 English words often become around 130 tokens, though this varies.
Example
‘Unbelievable!’ might be split into several tokens rather than treated as one word.
Related: LLM · Context window
Using AIContext windowThe amount of information a model can consider in one conversation or request.
The context window contains your prompt, conversation history, attached material and the model’s reply. When it becomes full, older details may be removed or overlooked. A large window helps with long documents, but does not guarantee perfect recall or reasoning.
Example
A model with a large context window may compare several reports in one request.
Related: Token · Prompt · LLM
Using AIPromptThe instruction, question or material you give an AI system.
A useful prompt explains the goal, relevant context, audience, constraints and desired format. Prompting is not a magic code: clearer instructions usually help, but the output still needs checking and revision.
Example
Instead of ‘Explain volcanoes,’ try ‘Explain how volcanoes form to a 10-year-old, using one analogy and a five-question quiz.’
Related: Context window · Chatbot
SafetyHallucinationA fluent AI answer that contains invented or incorrect information.
Language models generate likely wording, not guaranteed facts. They may invent dates, quotations, links or sources, especially when asked about obscure information. A polished tone is not evidence, so important claims should be checked against reliable independent sources.
Example
A chatbot may confidently cite a research paper that does not exist.
Related: LLM · RAG · Fact-checking
SafetyBiasA systematic pattern that unfairly favours or disadvantages certain results or people.
Bias can enter through training data, labels, design decisions or how a tool is used. AI can repeat patterns found in society at great scale. Testing with diverse examples and keeping humans accountable are essential safeguards.
Example
A hiring tool trained mostly on past hires could learn to prefer the same backgrounds as before.
Related: Training data · AI ethics
Using AIChatbotA conversational product that lets people interact with software using everyday language.
Some chatbots follow fixed scripts; newer ones use LLMs and can respond flexibly. The chatbot is the interface, while the model is the engine behind it. It may also connect to search, files, image tools or other services.
Example
ChatGPT, Claude, Gemini and Microsoft Copilot are AI chatbot products.
Related: LLM · Agent · Prompt
FoundationsMultimodal AIAI that can work with more than one kind of information, such as text, images, audio or video.
A multimodal model can connect information across formats. It might describe a diagram, answer questions about a recording or create an image from text. It can still misread details, especially tiny text, crowded scenes or unclear audio.
Example
Upload a photograph of a plant and ask the AI to describe visible signs of poor health.
Related: Computer vision · Generative AI
Using AIAI agentAI software that can plan steps and use tools to pursue a goal.
Unlike a chatbot that only replies, an agent may search, read files, update a calendar or run code. More freedom creates more risk: agents need clear permissions, limits, monitoring and human approval before consequential actions.
Example
A travel agent could compare trains, draft an itinerary and ask before making any booking.
Related: Chatbot · Automation · AGI
How AI worksRetrieval-augmented generation (RAG)A method that gives an AI relevant sources before it writes an answer.
RAG searches a selected collection of documents, retrieves useful passages and places them in the model’s context. It can make answers more current and traceable, but poor sources or retrieval can still cause mistakes.
Example
A school assistant searches approved policy documents before answering a question and links to the passages used.
Related: Embeddings · Hallucination · Context window
How AI worksEmbeddingsNumerical representations that place similar meanings near each other.
An embedding turns text, images or other content into a list of numbers. Software can compare those lists to find semantically similar material even when the exact words differ. Embeddings commonly power recommendation and RAG search.
Example
A search for ‘car repair’ can find a document titled ‘how to fix an automobile’ because their meanings are close.
Related: RAG · Model
How AI worksFine-tuningExtra training that adapts an existing model for a narrower task or style.
Developers provide carefully chosen examples so a general model behaves more consistently in a specialist setting. Fine-tuning is different from prompting and from RAG: it changes model behaviour rather than simply adding instructions or sources to one request.
Example
A support team might fine-tune a model on approved examples of helpful replies.
Related: Training data · RAG · Model
FoundationsOpen-source and open-weight AIAI shared so others can inspect or run parts of it, with important differences in what is actually open.
Open-weight usually means the learned model weights can be downloaded. Open-source traditionally also requires usable code and freedoms to study, modify and redistribute it. Licences, training data and commercial restrictions vary, so ‘open’ should never be assumed to mean everything is public.
Example
An organisation may run an open-weight model on its own computers to keep sensitive data in-house.
Related: Model · Fine-tuning
RoboticsComputer visionAI that analyses images and video.
Computer-vision systems can classify objects, locate items, read text and track movement. Their accuracy depends on conditions and training data, and errors can have serious consequences in areas such as medicine or policing.
Example
A robot uses computer vision to identify a box and judge where to grip it.
Related: Robotics · Multimodal AI
RoboticsRoboticsThe field of designing machines that sense, decide and act in the physical world.
A robot combines hardware such as motors and sensors with software that controls it. Some robots use AI for vision, language or planning; others use fixed rules. Physical actions introduce safety issues that do not exist in a text-only chatbot.
Example
A warehouse robot senses obstacles, plans a route and moves shelves while people supervise the system.
Related: Computer vision · AI agent · Automation
RoboticsAutomationUsing technology to carry out a task with less direct human effort.
Automation existed long before modern AI. Rule-based automation repeats known steps; AI automation can handle less predictable inputs but may be harder to inspect. Good automation keeps a person responsible for exceptions and high-impact decisions.
Example
A rule can automatically file invoices by date; AI can also extract the supplier and total from different layouts.
Related: AI agent · Robotics · Algorithm
SafetyDeepfakeAI-generated or altered media that convincingly imitates a real person or event.
Deepfakes may change a face, clone a voice or fabricate an entire scene. They can be used creatively, but also for bullying, fraud and misinformation. Check the original source, independent reporting and visible or audible inconsistencies before sharing.
Example
A scammer clones a family member’s voice and asks for emergency money.
Related: Generative AI · Computer vision · Misinformation
Who makes what?
30 leading AI companies and ecosystems
This is a learning guide, not a league table or endorsement. There is no single objective “top 30”: it includes important model developers, product companies, research organisations and platforms. Products change quickly, so always check current features, age limits, privacy terms and school policy before use.
No.CompanyMain productsStrengthGood for
Google DeepMind
Gemini is available in several sizes and product tiers.
Gemini · Veo · Imagen
Deep multimodal research plus strong integration with Google Search, Workspace and Android.
Long documents, research, media and Google-based workflows.
Anthropic
Known for a strong focus on AI safety and controllability.
Claude
Careful long-form writing, document analysis and increasingly capable coding and agent workflows.
Writing, analysis, coding and working with large documents.
Microsoft
Copilot is a family of products and may use models from Microsoft and partners such as OpenAI.
Copilot · Microsoft 365 Copilot
Brings generative AI into Word, Excel, PowerPoint, Windows, GitHub and enterprise systems.
Workplace productivity and organisations already using Microsoft tools.
Meta
Llama licences are open-weight but are not identical to traditional open-source licences.
Meta AI · Llama
Widely used open-weight models and consumer distribution through WhatsApp, Instagram and Facebook.
Custom deployment, research and AI inside social products.
xAI
Live social content can be useful but should be independently verified.
Grok
Real-time conversational AI closely connected to the X platform and a rapidly developing model family.
Current-event exploration and general chat where live information matters.
Mistral AI
Offers both downloadable and commercial models.
Le Chat · Mistral · Codestral
Efficient European models, flexible deployment and a strong open-weight offering.
Organisations wanting model choice, coding and European deployment options.
Cohere
Primarily business-facing rather than a mainstream consumer chatbot.
Command · Embed · Rerank
Enterprise language AI focused on secure retrieval, search and business data.
Company knowledge assistants and multilingual enterprise search.
AI21 Labs
Jamba uses a hybrid model architecture rather than only a standard transformer.
Jamba · Wordtune
Language models and writing tools with an emphasis on controllability and enterprise use.
Writing support and business language applications.
Alibaba Cloud
Qwen is one of the most significant model ecosystems from China.
Qwen · Tongyi
A large multilingual model family with strong coding, reasoning and open-weight options.
Multilingual applications, coding and custom model deployment.
DeepSeek
Data handling, hosting location and local rules should be checked before school or workplace use.
DeepSeek · R-series · V-series
Efficient reasoning and coding models that accelerated interest in lower-cost open-weight AI.
Reasoning, coding, research and local deployment experiments.
Baidu
Availability and features vary by region.
ERNIE Bot · ERNIE
Chinese-language models connected to a major search, cloud and autonomous-driving ecosystem.
Chinese-language search, enterprise tools and local-market applications.
Tencent
Many capabilities are designed around Tencent’s existing platforms.
Hunyuan · Yuanbao
Multimodal models integrated across a vast communications, games and cloud ecosystem.
Chinese-language consumer and enterprise applications.
Amazon Web Services
Bedrock is a model platform; Amazon Q is an assistant product.
Amazon Q · Nova · Bedrock
Enterprise infrastructure that offers its own models alongside a wide choice of third-party models.
Businesses building secure AI applications with model choice.
NVIDIA
NVIDIA is foundational to the industry even though it is best known for chips rather than a consumer chatbot.
Nemotron · NIM · NeMo
The leading AI-computing ecosystem, plus models and software for building and running AI efficiently.
AI infrastructure, enterprise deployment, simulation and robotics.
IBM
Granite includes open model options for language and code.
watsonx · Granite
Enterprise governance, smaller deployable models and integration with regulated business systems.
Governed AI in large organisations and specialised business use.
Apple
Apple combines its own models with selected external services for some requests.
Apple Intelligence · Foundation Models
Privacy-focused AI integrated into personal devices, with substantial on-device processing.
Everyday writing, communication and assistance across Apple devices.
Databricks
Its core audience is data and engineering teams.
DBRX · Mosaic AI
Connects model development and serving directly to enterprise data and analytics.
Companies building custom AI systems around their own governed data.
Snowflake
Best understood as a data-cloud AI ecosystem.
Arctic · Cortex AI
Brings language models and AI functions close to data already stored in Snowflake.
Enterprise analytics and AI without moving governed data between systems.
Salesforce
Value is strongest for organisations already using Salesforce data.
Agentforce · Einstein · xGen
AI agents grounded in customer, sales and service data inside a major CRM platform.
Sales, customer support and business workflow automation.
Writer
Designed for organisations rather than general consumer use.
Palmyra · Writer
Enterprise generative AI focused on brand control, governance and repeatable business workflows.
Marketing, operations and company-wide writing standards.
Perplexity
A citation is not proof; users still need to open and evaluate the source.
Perplexity · Sonar
Search-first answers with visible web citations and fast research workflows.
Starting research, discovering sources and current information.
Hugging Face
It hosts models from many organisations rather than relying on one model family.
Model Hub · Transformers · Inference
The central open AI community and platform for finding, sharing, testing and deploying models.
Education, research, open models and developer experimentation.
Aleph Alpha
Primarily serves organisations rather than consumers.
Luminous · PhariaAI
European enterprise AI with an emphasis on sovereignty, transparency and regulated use.
Public-sector and enterprise deployments requiring control and explainability.
Zhipu AI
International availability and product names can change by market.
GLM · ChatGLM
A major Chinese foundation-model developer with bilingual, coding and agent capabilities.
Chinese and English applications and custom deployment.
Moonshot AI
Service availability varies outside China.
Kimi
Long-context reading, web research and a popular consumer assistant ecosystem in China.
Reading long material and Chinese-language research workflows.
MiniMax
Different products may have different regional names and access rules.
MiniMax · Hailuo
Consumer-facing multimodal AI spanning language, speech, music and video generation.
Creative media and conversational applications.
01.AI
Model availability and focus continue to evolve quickly.
Yi
Bilingual open-weight and commercial models designed for efficient practical deployment.
Chinese-English applications and teams exploring downloadable models.
Technology Innovation Institute
TII is a research institute rather than a conventional consumer AI company.
Falcon
A prominent open model family backed by a major research institute in Abu Dhabi.
Research and organisations evaluating open-weight language models.
Stability AI
Included as a leading generative-AI ecosystem, not primarily an LLM company.
Stable Diffusion · Stable Audio
Influential open generative-media models, especially for image creation and customisation.
Creative image and audio workflows, research and local generation.
Guide reviewed June 2026. “Strength” describes a widely recognised focus, not a guarantee that one product is best for every task.