Glossary
The people, companies and terms behind the AI news, explained the way a friend would. Names in our articles link here automatically.
A2A
An open standard, started by Google, that lets AI agents built by different companies discover each other and work together.
AA-Omniscience
A benchmark that rewards a model for saying 'I do not know' instead of inventing an answer.
ACM
The main professional body for computer science, and publisher of a huge research library.
active parameters
The portion of a model's parameters actually switched on for any single word it produces, which in mixture-of-experts designs is far smaller than the total.
Adaptyv Bio
A lab that physically builds and tests designed proteins, and runs public competitions that have become an informal benchmark for protein design.
Adobe
The software company behind Photoshop and other creative tools, now building AI into all of them.
Adobe Firefly
Adobe's family of generative AI models and the app that bundles them, sold on the promise that the output is safe to use commercially.
Adversa AI
A security research firm that specialises in breaking AI systems and publishing how it did it.
AES
The worldwide standard for scrambling data so only the right person can read it.
Age prediction
Estimating how old a user probably is from their behaviour, rather than trusting what they typed in a box.
Agent Plugins
An open standard from Amazon, Cursor, Microsoft, OpenAI and Vercel for packaging AI agent extensions so the same bundle works across products.
Agent Skills
An open format for packaging written instructions and workflows that an AI agent can load when it needs them.
Agents' Last Exam
A benchmark of long multi-step tasks inside real applications, used to test whether an agent can finish work rather than just start it.
AGI
Artificial general intelligence: the hypothetical point where a system is broadly as capable as a person across most tasks, rather than good at one thing.
AI agent
An AI that doesn't just answer, but takes actions across multiple steps to complete a task.
AI Bill of Materials
A parts list for an AI model, tying it back to the data, licences and training runs that produced it.
AI for Science
Anthropic's programme giving researchers free Claude credits for scientific work.
AI gateway
A service that sits between your software and many AI providers, so you can switch models without rewriting code or juggling accounts.
AI gigafactory
The EU's term for a very large publicly co-funded site for training and running AI models.
AI Mode
Google Search's conversational mode that answers instead of listing links.
AI Office
The European Commission body responsible for supervising general-purpose AI models and enforcing parts of the EU AI Act.
AI Overviews
The AI-written summaries at the top of Google search results.
AI Security Institute
A UK government body that studies and tests the safety of advanced AI systems.
AI slop
Low-quality content churned out by AI in bulk, fast enough that nobody can review it all.
AI text detector
Software that estimates whether a text was written by a human or by AI.
AI Transparency Tags
Apple Music's disclosure system, where whoever uploads a track declares whether AI was used for the recording, the artwork, the lyrics or the video.
AIME
A hard American high school mathematics competition whose problems are now a standard benchmark for AI reasoning.
Alexa+
Amazon's language-model-powered rebuild of Alexa, which understands conversation rather than fixed voice commands.
Alibaba
The Chinese e-commerce and cloud giant, and a major AI model developer.
alignment
The work of getting a model to actually behave the way its makers intend, including refusing harmful requests.
Allen Institute for AI
A Seattle nonprofit AI research institute, founded by Paul Allen, known for fully open models and datasets.
alpha channel
The extra layer in an image file that records how transparent each pixel is.
Alphabet
Google's parent company, the holding company that owns Google and DeepMind.
AlphaFold
The AI system that predicts protein shapes, a landmark scientific use of AI.
AlphaGo
DeepMind's Go-playing system, still the reference point for AI producing a genuinely original idea.
Amazon
The e-commerce and cloud giant; its AWS cloud hosts much of the AI industry.
Amazon Mechanical Turk
Amazon's marketplace for paid human micro-tasks, long used to train AI.
Amazon Nova
Amazon's own family of AI models, launched in 2024 and largely wound down in 2026.
AMD
A chip maker, and the main challenger to Nvidia in AI hardware.
Amir Salek
The engineer who founded Google's TPU chip programme and shipped its first seven generations, hired by Anthropic in August 2026.
Andon Labs
A research outfit that lets AI agents run a real shop in San Francisco to study how they behave outside a benchmark.
Andreessen Horowitz
A large Silicon Valley venture capital firm, one of the biggest funders of AI startups.
Andrej Karpathy
A founding member of OpenAI and former head of AI at Tesla, now researching at Anthropic and widely followed for his plain-spoken explanations of how models work.
Anjney Midha
Investor behind several European AI companies, formerly a16z, now founder of the fund AMP.
annualised run rate
A company's most recent month or quarter of revenue multiplied out to a full year, used constantly in AI funding news.
Ant Group
The Chinese financial technology giant behind Alipay, and the parent of the AI lab inclusionAI.
Anthropic
The AI safety company that builds Claude, founded by former OpenAI researchers.
Antigravity
Google's agent-first coding environment, where you hand the AI a task and review the work rather than typing every line.
Anysphere
The company behind Cursor, the AI-first code editor.
Apache License
A widely used permissive open-source licence, common for AI models and tools.
API
A direct connection that lets one piece of software talk to another, including AI models.
API key
A secret string that identifies you to a service and bills you for what you use, effectively a password for software.
Apple
The iPhone maker, known for running AI directly on its devices.
ARC-AGI-3
A puzzle benchmark designed to test whether a model can work out rules it has never encountered before.
Arthur Mensch
Co-founder and CEO of Mistral AI, Europe's leading open-model lab.
Artifactory
A widely used server that stores and hands out software packages inside a company.
Artificial Analysis
An independent firm that benchmarks AI models on quality, speed and price.
arXiv
A free website where researchers post papers before, or instead of, formal peer review.
ASIC
A chip built to do one job extremely well, giving up the flexibility to do anything else.
Astra
OpenAI's next announced model family, built to coordinate several agents on one problem for hours or days.
AtCoder
A Japanese competitive-programming platform, now used as a tough test for AI coding.
Atlas
OpenAI's short-lived AI web browser, launched 2025 and shut down in 2026.
Auto Mode
Letting an AI coding agent work without approving every step, with a safety classifier watching instead.
Automated decision-making
When software decides something about a person without a human meaningfully involved, which EU law restricts.
Autoregressive
The standard way language models write: one token at a time, left to right, each one chosen after the one before it.
AVO
Nvidia's research agent system, built for tasks that run for days rather than minutes.
Azure
Microsoft's cloud platform, one of the main places AI models are run.
Backflip AI
US startup building AI that converts 3D scans into editable CAD models.
Baidu
The Chinese search and AI giant, active in open document-AI models.
Bartz v. Anthropic
The US authors' lawsuit that drew the line between training on copyrighted books, which a judge allowed, and pirating them, which he did not.
Beatport
The dominant online store for electronic music, where working DJs buy the tracks they play.
benchmark
A standard test used to measure and compare how well AI models perform.
Benchmark contamination
When the questions from a test have leaked into a model's training data, so a high score reflects memory rather than ability.
Benjamin Mann
A co-founder of Anthropic and one of the authors of the original GPT-3 paper.
Big Video Dataset
LAION's open video training collection from August 2026, covering roughly 10 million hours of footage with machine-written captions for picture and sound.
big world hypothesis
Richard Sutton's argument that any learning system is always far smaller than the world it operates in, and should be designed accordingly.
Black Forest Labs
The German lab behind the FLUX image models, used inside Photoshop, Canva and Picsart.
Black Hat
A long-running series of computer security conferences where researchers disclose new attacks and vulnerabilities.
Blackwell
The name of an Nvidia AI chip generation, now being succeeded by Vera Rubin.
Block
Jack Dorsey's payments and technology company, now building AI-agent tools.
Botnet
A network of hijacked internet-connected devices that an attacker controls remotely, usually without the owners noticing.
bounding box
A rectangle of coordinates marking where something sits in an image or on a page.
Brave
A privacy-focused web browser with its own search engine and AI features.
Bridgewater Associates
One of the world's largest hedge funds, which tested custom AI on finance tasks.
Broadcom
A chip and infrastructure supplier central to building large AI data centres.
bug bounty
A standing offer from a company: find a security hole in our product, tell us instead of selling it, and we pay you.
Bun
A fast tool for running JavaScript, popular with developers.
ByteDance
The Chinese company behind TikTok, and the AI video tool Seedance.
C2PA
An industry standard for attaching a signed record of where a file came from and how it was edited.
CAD
Software for designing physical objects, where the file stores the steps that built the shape rather than just the surface.
Canva
A popular, easy-to-use design tool that has built AI image and text features into everyday use.
Capture the flag
A security exercise where you are told a secret is hidden on a system and asked to break in and retrieve it.
Carson Gross
Creator of the htmx JavaScript library and author of the No AI Fridays proposal.
Cerebras
An AI chip company known for a processor the size of a dinner plate.
Chain of thought
The step by step working a model writes out before giving its final answer.
Character error rate
The standard score for text recognition accuracy: the share of individual characters a system gets wrong.
Chat template
The small piece of formatting that wraps a conversation before the model reads it, marking who said what.
ChatGPT
OpenAI's widely used AI chatbot, the app that brought AI to a mass audience.
Chirp
Google's earlier family of speech-to-text models, now largely superseded by transcription built directly into Gemini.
Chris Lattner
The compiler engineer behind LLVM and Apple's Swift language, now leading the Mojo language at Modular.
CI/CD
The automated system that tests and ships software every time a developer changes the code.
CISA
The United States government agency responsible for defending critical infrastructure, such as power grids and water systems, against cyber attacks.
Classifier
A small model with one job: sort something into categories, like safe or dangerous.
Claude
Anthropic's family of AI assistants, used for writing, coding, analysis and more.
Claude Code
Anthropic's AI coding tool that writes and edits software from the command line.
Claude Cowork
Anthropic's AI agent for everyday file and task work on your computer.
Claude Science
Anthropic's research workbench that runs locally so sensitive data stays on your machines.
Cloudflare
An American internet infrastructure company that sits in front of a large share of the web, handling traffic, security and increasingly AI services.
Co-Scientist
Google DeepMind's multi-agent system that plans experiments, runs lab equipment and drafts the resulting paper.
Codex
OpenAI's AI coding agent that writes and fixes software on its own.
cognitive debt
The idea that leaning on AI for thinking work leaves you with skills and understanding you have quietly stopped maintaining.
ComfyUI
A free interface for running image and video AI models on your own computer, built by wiring boxes together instead of typing settings.
Common Crawl
A nonprofit that has been copying large parts of the public web since 2007 and giving the archive away for free.
compute
The raw computing power, chips, data centres and electricity, needed to build and run AI.
Confidential computing
Running code in a sealed hardware environment so that even the machine's owner cannot see the data inside.
content management system
The software behind a website that lets people publish and edit pages without touching code.
context window
How much text a model can consider at once, its working memory for a single conversation.
Copilot
Microsoft's brand for the AI assistants built into Windows and Office.
copyright management information
The ownership and credit details attached to a copyrighted file, which US law makes it a separate offence to strip out.
CoreWeave
A cloud provider that rents out AI computing power, often Nvidia's newest chips.
cost-per-click
An advertising model where the advertiser pays only when someone actually clicks the ad, not when it is merely shown.
Coursera
A large online learning platform, co-founded by AI researcher Andrew Ng.
Credentio
Google's open-source C++ library for verifying C2PA Content Credentials on a device, without sending files to a server.
Cryptographic Context Injection
A prompt injection attack that hides its instructions behind encryption so filters see only gibberish, then lets the model decrypt them itself.
CUDA
Nvidia's software layer for running general computation on its graphics chips, and the main reason it is hard to switch away from Nvidia.
Cursor
A popular AI-powered code editor for developers.
CuspAI
A startup using AI to design new materials, such as carbon-capture filters.
CVE
A publicly catalogued software security flaw, each given a unique reference number.
CWE
A standard catalogue of software flaw types, so different tools describe the same kind of bug the same way.
CyberGym
A benchmark that measures how well AI models find and exploit real software vulnerabilities.
Dan Luu
An engineer and writer known for carefully re-testing widely repeated claims about software, and often finding them wrong.
Dario Amodei
Co-founder and CEO of Anthropic, and one of the more publicly cautious voices among AI lab leaders.
data loss prevention
A set of rules and tools that stop sensitive files or data from being accessed, shared or leaked by accident, including by AI assistants.
Databricks
A major data platform company where many businesses run their AI.
Daybreak
OpenAI's gated programme that gives vetted security professionals access to models with the usual safety filters relaxed.
Debian
One of the oldest and most influential Linux distributions, run by volunteers under a formal democratic constitution.
Decart
An Israeli startup whose software makes AI models run faster on existing chips.
deepfake
A convincing fake image, video or audio clip of a real person or place, made by an AI model.
DeepSeek
A Chinese AI lab famous for strong open models at very low cost.
DeepSWE
A benchmark that tests whether a model can carry a long software engineering task through to a working result.
Deezer
A French music streaming service, notable for publishing hard numbers on how much AI music it receives.
Demis Hassabis
The Nobel-winning co-founder and CEO of Google DeepMind.
Demucs
An open-source AI model that separates a finished song back into individual parts such as vocals, drums and bass.
dense model
A model that uses all of its parameters for every word it produces, unlike a mixture of experts.
Deutsche Telekom
Germany's largest telecoms company, backing European sovereign-AI efforts.
Diffusion model
A model that starts from random noise and repeatedly cleans it up until a finished image, sound or block of text appears.
digital audio workstation
The software musicians use to record, arrange and mix a song, with every instrument on its own track.
Digital Markets Act
An EU law that forces the largest tech platforms to open up to competitors.
DirectIO
A way of reading a file straight from the drive without letting the operating system cache a copy along the way.
Discovery Loop
A research startup founded in August 2026 by four senior Google AI engineers, structured as a public benefit corporation.
Distillation
Training a smaller model on the answers of a bigger one, so it copies the bigger model's behaviour cheaply.
DMCA
A US copyright law from 1998, now central to fights over AI companies scraping the web.
DNS rebinding
A trick where a website briefly points its own address at your computer so your browser treats your local software as part of that site.
Doubao
ByteDance's AI assistant, one of the most used chatbots in China.
Double-blind evaluation
An AI test where the evaluator never sees the model and the model's owner never sees the questions.
DRAM
The ordinary working memory in every computer, sold as DDR4 and DDR5 sticks.
dual-use
Knowledge or capability that helps beneficial work and harmful work equally well, so a tool cannot simply block the bad half.
DuckDB
A small, fast database that runs inside your program instead of on a server, built for everyday analysis of modest amounts of data.
DuckLabs
The Amsterdam company behind the DuckDB database, acquired by AWS in August 2026.
EleutherAI
A non-profit research lab that builds openly licensed AI models and training datasets, run largely by volunteers.
ElevenLabs
An AI voice company known for realistic speech synthesis and highly accurate transcription.
Elo
A rating system borrowed from chess, now used to rank AI models by head-to-head comparisons.
Elon Musk
Entrepreneur behind Tesla, SpaceX and X, and founder of the AI lab xAI, which builds the Grok models.
Encord
A company that helps prepare and label the data used to train AI models.
Epoch AI
A research group that measures and forecasts AI progress.
Erik Brynjolfsson
The Stanford economist behind the 'Canaries in the Coal Mine' research tracking AI's effect on entry-level employment.
Etched
An American chip startup building processors that run only transformer models, aiming to make AI far cheaper to operate.
EU AI Act
The European Union's landmark law regulating AI according to how risky its uses are.
Expert Intelligence
Google's arrangement for letting people use books and other paid sources inside its AI products, if they own them.
ExploitBench
A test of whether a model can go beyond spotting a software flaw and actually build a working attack.
F1 score
A single accuracy number that balances how often a system is right against how much it finds.
Fable 5
An advanced Anthropic model, powerful at coding and available through Claude plans.
Fair use
A US legal doctrine that allows limited use of copyrighted material without permission, and the main defence AI companies use over training data.
Fair Work Commission
Australia's national workplace tribunal, and the first major tribunal to require people to disclose their use of AI in filings.
fallback
When an AI product quietly hands your request to a different, usually weaker model instead of answering with the one you asked for.
False positive
When a detection system raises an alarm about something innocent.
Fields Medal
Mathematics' most prestigious prize, awarded every four years to researchers under 40.
fine-tuning
Taking an existing model and training it further on your own data to specialise it.
Firecracker
A stripped-down virtual machine technology from AWS, designed with far less to attack than a general-purpose one.
FlashAttention
A widely used technique that makes the attention step of a transformer far faster by moving less data around on the GPU.
FLEURS
A multilingual speech benchmark used to test how accurately a model transcribes audio across many different languages.
FLUX
The image and video model family from Black Forest Labs, partly open to download, partly paid.
Forcepoint
A cybersecurity company whose X-Labs research team studies attacks on AI systems.
Foundation model
A large, general-purpose model trained once on broad data and then adapted to many different tasks.
FP8
A compact way of storing a model's numbers using eight bits each, which halves memory use compared to the older 16-bit standard.
Fractile
A British chip startup, spun out of Oxford in 2022, building inference chips that keep model weights on the same silicon as the compute.
Francois Chollet
AI researcher, creator of the Keras library and co-founder of the ARC Prize.
frontier model
One of the most advanced AI models at the leading edge of what's currently possible.
FrontierCode
A coding benchmark from Cognition that grades the quality of production code rather than whether it merely runs.
Fugu
Sakana AI's system that coordinates many smaller models instead of using one large one.
Functional Generative Networks
A way of generating varied but physically sensible predictions by adding randomness inside the network rather than to the output.
GDPR
The EU's data-protection law, which sets strict rules for handling people's personal data.
GDPval
An OpenAI benchmark that tests models on realistic office work rather than puzzles.
Gemini
Google's family of AI models, spanning fast lightweight versions to advanced ones.
Gemma 4
Google's open model that you can download and run on your own computer.
GEN-1.5
Generalist AI's robot model that learns a task from one short video clip loaded into its context window.
GenCast
Google DeepMind's 2024 weather model, the first AI system to beat Europe's leading ensemble forecast.
General Catalyst
A major venture capital firm and a heavy backer of AI startups.
Generalist AI
A robotics startup building models that let a robot pick up a new task from a short demonstration instead of hours of training.
generalization
How well a model carries a skill over to situations it was not trained on.
Generative AI
The umbrella term for AI systems that produce new content, such as text, images, audio or code, rather than only sorting or scoring existing things.
Genie 3
Google DeepMind's world model, which generates interactive 3D environments you can move around in.
GenRec
Netflix's language-model-based recommendation system, which turns your viewing history into plain text.
GGUF
The standard file format for running AI models on your own computer, packaging the model and its settings into one file.
GitHub
The world's largest platform for hosting and sharing code, owned by Microsoft.
GitHub Copilot
Microsoft's coding assistant, which now lets developers pick between models from several competing labs.
GitLost
A 2026 attack that hijacked AI coding agents through poisoned repositories.
GlucoFM
A very small Google Research model that interprets continuous glucose monitor data to spot metabolic patterns.
The search giant whose DeepMind lab builds the Gemini models.
Google AI Studio
Google's free browser playground for trying Gemini models and getting an API key.
Google DeepMind
Google's AI research division, maker of the Gemini models and AlphaFold.
Google Imagen
Google's family of text-to-image models, now being folded into the general Gemini model line.
GPQA
A hard benchmark of expert-level science questions used to test AI reasoning.
GPT-2
OpenAI's 2019 language model, initially withheld as too dangerous to release, which set the template for staged AI releases.
GPT-4
OpenAI's 2023 model, the one that made large language models mainstream and held the top spot longer than anything since.
GPT-5.5
OpenAI's 2025 flagship model, predecessor of the GPT-5.6 family.
GPT-5.6
OpenAI's model line that combines three models sharing one large memory.
GPT-5.6-Cyber
OpenAI's cybersecurity-specific model, trained to find unknown vulnerabilities and build exploits, available only through Daybreak Red.
GPT-Image-2
OpenAI's image generation and editing model, top of the public image leaderboards through mid-2026.
GPT-Live
OpenAI's real-time voice mode that can listen and speak at the same time.
GPT-Red
An internal OpenAI model built to attack its own systems and find security holes.
GPU
A powerful processor, originally for graphics, now the workhorse chip of modern AI.
Gradient boosting
A classical machine learning method that still beats large language models at predicting values in spreadsheet-like data.
Gravity
An advertising startup building an ad network for AI chatbots, including ads aimed at AI shopping agents rather than people.
Greg Brockman
Co-founder and president of OpenAI, a leading voice on where AI interfaces are heading.
Gregory Kurtzer
The open source veteran behind CentOS, Singularity and Rocky Linux, now working on open AI training data.
Grok
Elon Musk's AI chatbot, built by his company xAI.
Grok Imagine
xAI's image generation and editing model, built into Grok.
Groq
A chip company building hardware specialised for running AI models quickly, rather than for training them.
Grounding
Making a model answer from a specific set of documents you give it rather than from memory.
Guardrail model
A small extra model that sits beside a chatbot and checks whether a message or an answer breaks the rules.
Guidelight
An independent nonprofit that grades AI companies on how well they control the AI systems they run internally.
H200
A high-end Nvidia data centre GPU used for training and running AI models, the memory-upgraded successor to the H100.
Hacker News
A tech discussion forum where many AI projects first surface.
HackerOne
The best-known platform for coordinated vulnerability disclosure, where researchers report security flaws to companies and sometimes get paid for them.
hallucination
When an AI states something false or invented as if it were true.
harness
The software wrapped around a model that decides what it sees, which tools it can use, and which model handles which task.
Harvey
A US company that builds AI tools specifically for lawyers and law firms.
Harvey Tenet
Harvey's first in-house legal model, built by further training an open-weight Kimi K3.
Hatch
Meta's paid AI agent that does tasks for you across its apps.
HAWK
A proposed post-quantum encryption scheme, meant to resist future quantum computers.
HBM
Stacked memory built to sit next to AI chips and feed them data fast enough.
Higgsfield
An AI video startup aimed at marketers and social media teams, valued at 5.4 billion dollars in August 2026.
HTML
The markup language that describes what a web page or email contains and how it should look.
htmx
A small JavaScript library that lets web pages update in place without a heavy front end framework.
Hugging Face
The main public library where open AI models are shared and downloaded.
Human Intelligence Task
The small unit of paid human work on crowdsourcing platforms, typically a few seconds of judgement for a few cents.
Humanity's Last Exam
A deliberately brutal benchmark of expert questions across many fields, built so models cannot ace it.
Humanizer
A tool that rewrites AI-generated text specifically to slip past AI detectors.
Hy4
Tencent's open-weight flagship model from August 2026, with 770 billion total parameters and a context window above one million tokens.
IBM
A century-old American technology company that now builds open AI models and enterprise AI tools under the Granite and watsonx names.
IBM Granite
IBM's family of open-weight language and speech models, released under a permissive licence so anyone can download and run them.
ICML
One of the three biggest academic conferences in machine learning, where much of the field's research is first published.
in-context learning
When a model picks up a new skill from examples placed in the prompt, without any change to the model itself.
inclusionAI
The AI research arm of China's Ant Group, which publishes the Ling family of open models.
inference
Running a trained AI model to get an answer, as opposed to training it in the first place.
Intelligence Index
A single combined score from Artificial Analysis that blends several benchmarks into one number for comparing models.
Internet Archive
A non-profit digital library whose Wayback Machine holds hundreds of billions of saved snapshots of web pages.
IRON
XPeng's humanoid robot, built to move and work the way a person does, heading into mass production by the end of 2026.
Isomorphic Labs
Alphabet's drug discovery company, spun out of Google DeepMind to turn AlphaFold-style research into medicines.
Jack Dorsey
Co-founder of Twitter and head of Block, and an advocate for open, decentralised tech.
Jacob Tsimerman
Number theorist and Fields Medalist who left the University of Toronto to work on AI safety at OpenAI.
Jacobian Lens
An Anthropic research method for reading a model's hidden working memory.
jailbreak
A prompt crafted to get a model to do something its safety training was supposed to prevent.
Jalapeño
OpenAI's first custom chip, designed with Broadcom to run finished models cheaply rather than train new ones.
Jane Street
A large quantitative trading firm that has become an unusually active investor in and customer of AI hardware.
JavaScript
The programming language that runs in every web browser.
Jeff Bezos
Amazon's founder, increasingly active as an AI investor.
Jeff Dean
A legendary Google engineer who co-founded Google Brain and served as Google DeepMind's chief scientist, before leaving in August 2026 to start Discovery Loop.
Jensen Huang
Co-founder and chief executive of Nvidia, the company whose chips most of the AI industry runs on.
JEPA
Meta's world model architecture, which predicts an abstract summary of what comes next rather than the pixels.
Jira
Atlassian's issue and project tracking tool, widely used by software teams to record bugs and tasks.
John Jumper
Nobel-winning scientist behind AlphaFold, who moved from Google DeepMind to Anthropic in 2026.
Jony Ive
The designer behind the iMac, iPod and iPhone, now shaping OpenAI's hardware after it bought his startup io for 6.5 billion dollars.
KoboldCpp
A single-file program for running open AI models on your own computer, with a browser chat window built in.
LAION
A German non-profit that publishes very large open training datasets, including the image collection behind Stable Diffusion.
Lancium
A Texas power infrastructure developer that supplies electricity to large AI data centres, including the Stargate site.
large language model
The type of AI, trained on huge amounts of text, that powers chatbots like ChatGPT and Claude.
Le Chat
Mistral's AI chat assistant, the French answer to ChatGPT.
Lean
A programming language for writing mathematical proofs that a computer can check line by line.
Leanstral
Mistral's model specialised in mathematically proving that code is correct.
LearnVector
A startup building an AI tutor that plans a learner's path and sticks with them until they understand.
Leopold Aschenbrenner
Former OpenAI safety researcher whose 2024 essay on fast AI scaling became the basis for his own hedge fund.
LibGen
A large pirate library of books, widely used as AI training data.
Ling 3.0 Flash
An open-weight model from inclusionAI, the highest-scoring open model under 124 billion parameters when it launched in August 2026.
Linux
The free, open operating system that runs most servers and AI infrastructure.
Linux Foundation
A non-profit that hosts and governs major open-source projects, now including AI ones.
Llama
Meta's family of open-weight language models, the release that made freely downloadable models mainstream.
llama.cpp
The open-source engine that lets AI models run on ordinary computers, including inside Ollama and LM Studio.
LM Studio
A free desktop app for downloading and running open AI models on your own computer.
LMArena
A public leaderboard that ranks AI models by asking people which of two anonymous answers they prefer.
long-horizon task
Work that takes an AI many steps and a long stretch of time to finish.
Lovable
A Stockholm startup that turns plain-language descriptions into working websites and apps.
Luanti
A free open-source platform for building and playing blocky voxel games, formerly called Minetest, widely used in schools.
Lyria
Google's family of AI models for generating music from text descriptions.
MAI
Microsoft's own in-house AI model family, reducing its reliance on OpenAI.
Manus
A general purpose AI agent service that became well known in 2025 for completing multi-step tasks on its own.
Mark Zuckerberg
CEO of Meta, which is investing tens of billions into AI and AI agents.
Markdown
A simple way to format text using plain characters, widely used with AI tools.
MCP
An open standard that lets AI assistants connect to outside tools and data in a uniform way.
Menlo Ventures
A venture-capital firm whose early bet on Anthropic became hugely profitable.
Mental World Modeling
A framework that adds beliefs, intentions and social norms to world models, which normally track only physics.
Mercor
A company that supplies AI labs with expert human data, from written examples to specialist evaluations.
Meta
The company behind Facebook and Instagram, and a major AI spender and open-model publisher.
Metal Performance Shaders
Apple's software layer for running AI computation on the graphics chip inside a Mac, the Apple Silicon counterpart to Nvidia's CUDA.
METR
An independent group that tests advanced AI models for dangerous capabilities.
Michael Burry
The investor made famous by The Big Short, now one of the loudest sceptics of how AI companies account for their chips.
Microsoft
The software giant behind Windows, Office and the Copilot AI assistants.
MIDI
A decades-old standard that stores which notes were played rather than the sound itself, so anything can play them back.
Mike Krieger
Instagram's co-founder, now Anthropic's chief product officer.
Mikey Shulman
Co-founder and chief executive of Suno, the AI music generator.
Minibinder
A small protein designed from scratch to latch tightly onto a specific target protein, which is how many modern medicines work.
MiniMax
A Chinese AI lab known for powerful open-weight models like M3.
MiniMax H3
A 33 billion parameter open-weight video model that generates short clips with synced audio, but whose licence excludes the EU and other regions.
MiniMax M3
An open-weight model from MiniMax that rivals closed systems and handles text, images and video.
Mira Murati
Former OpenAI chief technology officer, now running her own open-model lab.
Mistral AI
The French lab making Europe's leading open-weight AI models.
MIT License
A very permissive open-source licence that lets almost anyone use the code or model freely.
Mixture of experts
A model design where only a small part of the network runs for each request, keeping big models cheap to use.
MLX
Apple's framework for running AI models efficiently on its own chips.
mode collapse
When a model settles into one preferred way of saying things instead of spreading across the many ways a person might phrase it.
Model Hardware Standard
An Anthropic specification that gives AI agents one common way to discover, understand and safely operate physical devices like lab instruments and robotic arms.
model weights
The billions of numbers a model learns during training, effectively its knowledge.
ModelScope
Alibaba's platform for sharing open AI models, roughly China's answer to Hugging Face.
Mojo
A programming language built for AI work that looks like Python but compiles to fast code for CPUs and GPUs.
Monthly active users
How many people opened an app at least once in thirty days, which is a lower bar than it sounds.
Moonshot AI
The Chinese lab behind the popular Kimi models.
Motion Picture Association
The trade body for the major US film studios and streaming services.
multi-token prediction
A speed trick where a model guesses several words ahead at once instead of one at a time.
Muse Code
Meta's terminal coding agent, released in beta in August 2026 and powered by the Muse Spark 1.2 model.
Muse Glimmer
Meta's 30 billion parameter open-weight model from August 2026, small enough to run on a single consumer graphics card.
Muse Spark
Meta's family of AI models for generating images and video.
Mustafa Suleyman
CEO of Microsoft AI, co-founder of DeepMind and later Inflection AI.
Mythos
Anthropic's specialised model line aimed at cybersecurity work.
N-gram
A run of N consecutive words or tokens, the simplest way of capturing which pieces of text tend to follow each other.
Nano Banana
Google's family of image generation and editing models, known for photorealistic results.
Neill Blomkamp
The director of District 9, now making films with AI video generation.
NemoClaw
Nvidia's tool for running an AI coding agent locally, which shipped with a serious security flaw in August 2026.
Nemotron
Nvidia's family of open-weight language models, tuned for speed and efficiency rather than maximum intelligence.
Neural Engine
Apple's dedicated on-chip block for AI work, separate from the main processor and graphics cores.
News Corp
The Murdoch media group behind the Wall Street Journal and the New York Post.
NewsGuard
A company that rates the reliability of news websites and tracks online falsehoods, increasingly used to audit what AI chatbots say.
NIST
A US government body that sets widely used technical and security standards, including for AI.
Nitter
An open-source front end that showed public X posts without an account, shut down after legal threats in August 2026.
No AI Fridays
A 2026 proposal that software teams switch off their AI coding assistants one day a week to keep their own skills sharp.
NotebookLM
Google's AI notebook that answers questions about documents you give it.
Nvidia
The chip maker whose GPUs power most AI training and a lot of AI at home.
Oak Lab
Richard Sutton's Toronto startup, betting against the way AI is currently built.
Oasis Security
A security research firm focused on non-human identities, machine accounts and the credentials that AI agents use.
OAuth
The standard that lets you grant an app limited access to your account without sharing your password.
Ollama
A popular, free tool for running open AI models locally on your own computer.
Olmo
A fully open model family from the Allen Institute for AI, where the training data and code are published alongside the weights.
open source
Software, or an AI model, released so others can freely use, inspect and build on it.
open-weight model
An AI model whose trained parameters can be downloaded and run on your own hardware.
OpenAI
The US research company behind the GPT models and ChatGPT.
OpenClaw
An open-source framework that lets a language model drive a browser and call web services to finish multi-step tasks on your behalf.
OpenRouter
A marketplace that gives you access to many AI models through a single connection.
OpenWALDO
A community project building a shared, licensed and auditable pool of AI training data.
optical character recognition
Technology that turns images of text, scans, photos, PDFs, into editable, searchable data.
Oracle
A veteran enterprise software giant, now a major AI cloud provider.
Orca
A model from Beijing's BAAI lab built on a different architecture than chatbots.
OWASP
A nonprofit that publishes the security industry's most-cited lists of software risks, including one for AI applications.
Pangram
A company that builds detectors for AI-generated text.
Parameter
One of the adjustable numbers inside a model that get tuned during training.
Paul Erdos
Hungarian mathematician who left behind a famous catalogue of unsolved problems, now a favourite testing ground for AI.
Perplexity
An AI search engine that answers questions with cited sources.
Persistent agent
An AI agent that keeps running across days rather than stopping when a single task ends.
Pew Research Center
A nonpartisan American research organisation whose surveys and data studies are widely treated as a neutral reference point.
Phantom squatting
Registering fake package names that AI models tend to invent.
Pieter Abbeel
Robotics and reinforcement learning researcher, now leading Amazon's frontier model research.
pig-butchering
A long-running fraud where the scammer builds a friendship over weeks before steering the victim toward a fake investment or app.
Pinecone Nexus
A Pinecone product that pre-organizes a company's documents into a structured layer AI agents can query directly, instead of searching raw text every time.
Pixtral
Mistral's vision component, the part that lets a model look at images instead of only reading text.
Poolside
An AI startup building open coding models small enough to run on your own machine.
post-training
Everything done to a model after the expensive base training, to shape how it behaves.
Preferred Sources
A Google setting that tells Search, Discover and Google News to show more from publications you have chosen yourself.
Preparedness Framework
OpenAI's internal rulebook for measuring how dangerous a model is before release, and what safeguards that danger level requires.
Pretraining
The first and most expensive training phase, where a model works through enormous amounts of text and learns general patterns.
Prior Labs
A German AI startup specialised in understanding tables and spreadsheets.
PrismML
A startup that shrinks large open AI models small enough to run on a phone.
Private Safety Processing
OpenAI's system for spotting misuse across several conversations while still storing none of the content.
procedural generation
Building content with code that calculates it on the spot, rather than loading finished files someone made earlier.
Project Glasswing
Anthropic's programme for using AI models to hunt for security vulnerabilities in real software at scale.
Prompt caching
Storing the unchanging part of a conversation so the model does not pay full price to re-read it every turn.
prompt injection
An attack that hides instructions in content an AI reads, hijacking what it does.
Protein folding
Working out the three-dimensional shape a protein takes, which determines what it does in the body.
pull request
A formal proposal to change someone's code, which a human normally has to approve.
PyPI
The public registry where Python software packages are published and installed from.
Python
The programming language most AI is built and used with.
QEMU
Open-source software that creates virtual machines, the standard way of giving untrusted code a padded cell to run in.
Qualcomm
US chip designer best known for smartphone processors, now pushing into AI data centre hardware.
Quantization
Shrinking a model by storing its numbers less precisely, so it fits on smaller hardware at a small cost in quality.
Qwant
A French search engine hosted in Europe that does not store search history or profile users for advertising.
Qwen
Alibaba's family of open AI models, among the most downloaded worldwide.
Qwen3.8 Max
Alibaba's 2.4 trillion parameter flagship model, strong on long multi-step work but markedly more willing to guess than its predecessor.
Qwen3.8-Flash-Next
An experimental open-weight model from Alibaba's Qwen team, released in August 2026 as a preview of the architecture behind Qwen4.
RadLE
A benchmark testing AI models on real radiology cases.
Ramp
A US corporate card and expense management company whose spending data has become a widely cited indicator of real AI adoption.
RayNeo
A Chinese smart glasses maker, spun out of TV manufacturer TCL, known for display-first AI glasses.
Reconstruction benchmark
A 2026 test that gives a model only a paper's reference list and asks it to recover the paper's core idea, with frontier models scoring 3 to 15 percent.
recursive self-improvement
The idea of an AI improving itself, then using the better version to improve further.
Red teaming
Deliberately attacking your own system to find out how it breaks before someone else does.
A discussion platform whose vast archive of human conversation has become valuable and contested training material for AI models.
Redis
A very fast in-memory database used almost everywhere as a cache, and increasingly as a memory store for AI applications.
Redwood Research
A non-profit AI safety lab focused on control, meaning ways to stay safe even if a model is misaligned.
Reflect
An AI journaling app that asks you questions instead of answering them.
reinforcement learning
A training method where a model learns by being rewarded for good answers instead of being shown correct ones.
Remote code execution
The worst kind of software flaw: an attacker gets to run their own programs on your machine, from a distance.
Retrieval-Augmented Generation
A technique that feeds a model relevant snippets from your own documents before it answers.
reward hacking
When an AI system hits the number it was asked to improve without doing the thing that number was supposed to measure.
RHAE
The scoring metric behind ARC-AGI-3, which measures not just whether an agent solved a level but how many moves it needed compared with a human.
Richard Sutton
A founder of reinforcement learning who thinks current AI methods are a dead end.
Robin Rombach
Co-founder and CEO of Black Forest Labs, first author of the latent diffusion research behind Stable Diffusion.
Robostral Navigate
Mistral's AI model for robot navigation.
Rocky Linux
A free, community-run version of Red Hat Enterprise Linux that runs on millions of servers.
ROCm
AMD's open software stack for running AI workloads on its accelerators, the alternative to Nvidia's CUDA.
Rowboat
An open-source, local-first AI coworker that keeps your data on your machine.
Ruflo
An open-source tool for running coordinated swarms of AI agents, hit by a maximum-severity security flaw in July 2026.
RuntimeWire
A technology news site written almost entirely by AI agents, with one human as editor.
Rust
A programming language known for safety and speed, common in AI tooling.
S1 (Skild)
Skild AI's robot model that performs a new task after watching one human video, with no extra training.
SaferAI
A nonprofit that independently evaluates AI models for dangerous capabilities and rates how seriously labs manage risk.
Sakana AI
A Tokyo AI lab exploring nature-inspired approaches, such as combining many smaller models.
Salvatore Sanfilippo
The Italian programmer who created Redis, known online as antirez and for writing small, fast systems code by hand.
Sam Altman
Co-founder and CEO of OpenAI, one of the most prominent figures in AI.
Samsung
The South Korean electronics and chip giant, active across AI hardware and investment.
sandbox
A sealed-off testing area where software can run without touching anything real.
SAP
Europe's largest software company, building AI into business systems.
Satya Nadella
CEO of Microsoft, one of the most quoted voices on how AI gets paid for.
Scale AI
A large data-labelling company that supplies the human-annotated training and evaluation data AI labs run on.
Seedance
ByteDance's AI video-generation tool at the centre of a Hollywood dispute.
Seedream
ByteDance's AI image-generation model, sibling to the Seedance video model.
SemiAnalysis
An independent research firm led by Dylan Patel that analyses chips, data centres and the economics behind AI.
SensorFM
A foundation model for sensor data instead of text.
Series B
A named stage of startup funding, showing how mature a company's investment is.
Shieldstral
Mistral's small open model that checks text and images against a safety rule you write yourself in plain language.
Simon Willison
A programmer and writer whose blog is one of the most reliable independent guides to what AI models can and cannot do.
SimpleQA
A benchmark that asks models short factual questions with one correct answer, and is much harder for them than it sounds.
SK hynix
South Korean memory manufacturer and the leading supplier of HBM for AI chips.
Skild AI
A Pittsburgh robotics company building general-purpose 'robot brain' models that work across different machines.
Snowflake
A big data-platform company where many businesses store and analyse their data, increasingly with AI.
SoftBank
A Japanese investment giant making some of the largest bets in AI infrastructure.
Sony Music Publishing
The world's largest music publisher, representing the songwriting side of the business rather than recordings.
Soofi S
A fully open German AI model, trained in Munich on European infrastructure.
Sora
OpenAI's AI system for generating video from a text description.
speculative decoding
A speed trick where a small fast model drafts ahead and a bigger model checks its work.
Speech-to-text
Technology that converts spoken audio into written text.
Spotify
The music streaming service, now building conversational AI into its app.
SQL
The standard language for asking questions of a database.
SRAM
A fast type of computer memory that sits directly on the processor die, holding data without needing constant refreshing.
SSH
The standard way of logging into another computer over a network and typing commands on it securely.
Stability AI
The company behind Stable Diffusion, an early and influential open image generator.
Stable Diffusion
The image generator that made open-weight AI art mainstream, released publicly in 2022 and still the base for a huge amount of community tooling.
stablecoin
A cryptocurrency designed to hold a fixed value, usually one US dollar, so it can be used for payments rather than speculation.
stealth model
An unreleased model published anonymously, usually on a routing service, so a lab can collect real usage data before putting its name on it.
StemDeck
A free open-source desktop app that splits songs into separate instrument tracks on your own computer, with no upload and no account.
Stripe
An American payments company that handles online transactions for millions of businesses, and is now buying its way into AI billing infrastructure.
Substantial equivalence
The US regulatory route by which most AI medical devices reach the market, requiring only that they match an already approved device rather than prove they help patients.
Suno
An AI music generator that turns a written description into a finished song.
SWE-Bench
A widely used test of how well AI models fix real software bugs.
Swiftlet
An open source runtime that runs very large mixture-of-experts models on Macs and iPhones by streaming most of the model from the SSD.
SynthID
Google DeepMind's invisible watermark, hidden inside AI-generated images, audio or video rather than attached as a label.
system prompt
The hidden instructions that set an AI's role and rules before you start chatting.
Taalas
A Toronto chip startup that etches a specific AI model directly into silicon, acquired by AMD in August 2026.
TabPFN
An open model built for spreadsheets and databases rather than for text.
Tabular data
Plain rows-and-columns data, the kind that lives in spreadsheets and databases, and the format AI has been slowest to master.
Tau Robotics
A San Francisco startup renting out humanoid cleaning robots that people steer remotely.
Teleoperation
A human steering a robot from somewhere else, often while the demo implies the robot is doing it alone.
temperature
A setting that controls how predictable or how varied an AI model's output is.
tenant isolation
The rule that keeps one customer's data invisible to every other customer on the same service.
Tencent
Chinese technology giant behind WeChat and a large games business, and one of China's most active publishers of open-weight AI models.
Terence Tao
One of the world's leading mathematicians, and a careful tester of AI's reasoning.
Terminal-Bench
A test of whether an AI model can get real work done from a command line.
Tesla
The electric-car company using AI for driver assistance and robotics.
Text-to-speech
Technology that turns written text into spoken audio, now good enough to be hard to tell from a recording.
The Information
A subscription tech news publication known for breaking deals and internal numbers before anyone else.
Thinking Machines Lab
The AI startup founded by former OpenAI executive Mira Murati; maker of the Inkling model.
Three.js
A free JavaScript library that lets a web browser draw and animate 3D scenes.
tl;dv
An AI meeting recorder that joins video calls, transcribes them and writes summaries.
token
The small chunk of text, roughly three-quarters of a word, that AI models read and produce.
TOPS
Trillions of operations per second, a headline speed number for AI chips, especially in phones and cars.
tortured phrase
A technical term mangled into a nonsensical synonym by paraphrasing software, used as a fingerprint for low-quality or plagiarised papers.
Toyota
The Japanese carmaker, now a serious player in AI-driven robotics.
TPU
A specialised Google chip designed to run AI workloads efficiently.
Tracer.AI
A brand protection company that uses AI agents to detect alleged infringements and file takedown notices at high volume.
Trail of Bits
A US security research firm known for deep technical audits and, more recently, for testing what AI agents can do offensively.
Training data
The pile of text, images and code a model learns from, and the part labs are least willing to describe.
Trainium
Amazon's own AI chip, built to run training and inference on AWS without paying Nvidia's margin.
TSMC
The Taiwanese contract chipmaker that physically manufactures most advanced AI chips, including Nvidia's.
Twist Bioscience
A company that synthesises DNA to order on silicon chips, supplying the genetic instructions labs need to turn a designed protein into a real one.
Unitree
A Chinese robotics maker known for cheap four-legged robots and attention-grabbing humanoids.
Unsloth
An open-source project that makes fine-tuning and running language models faster and small enough for one GPU.
usage credits
A pay-as-you-go way of billing for AI, where you're charged for what you actually use.
Varonis
A data security company whose research arm, Varonis Threat Labs, has repeatedly found one-click flaws in AI assistants like Microsoft Copilot.
Vera Rubin
Nvidia's next-generation AI system, the successor to Blackwell.
Vercel
A US company that hosts websites and web apps, and the maker of the popular Next.js web framework.
Vibe check
An informal, unscored test of an AI model where one person tries something unusual and judges the result by eye.
Vibe coding
Building software by describing what you want in plain language and letting an AI write the actual code.
Vision-language-action model
A robot control model that takes in camera images and spoken instructions and outputs the actual movements.
VISTA
A research harness for playing ARC-AGI-3, often used as the comparison point for other agent systems.
vLLM
The most widely used open source server for running a downloaded model efficiently, especially when many people query it at once.
Voice cloning
Copying someone's voice from a short recording so a machine can say anything in it.
Volta Infra Holdings
A cloud computing startup founded in early 2026 that signed a roughly 10 billion dollar, six year compute deal with Anthropic.
voxel
A cube of volume in a three-dimensional grid, the 3D equivalent of a pixel, and the building block of blocky games and many medical scans.
Voxtral
Mistral's speech recognition models, known for undercutting competitors on price.
VulnCheck
A US security intelligence firm that tracks which software vulnerabilities are genuinely being exploited in the wild.
Walden Robotics
A Toyota spin-off building humanoid robots that roll on wheels instead of walking.
warm shell
A finished, powered and cooled data centre building, ready for servers to be plugged in.
Warner Chappell Music
The publishing arm of Warner Music Group, one of the three biggest music publishers in the world.
Waymo
Alphabet's self-driving car company, which runs paid robotaxi services in several US cities.
Wear OS
Google's operating system for smartwatches, the watch equivalent of Android.
WeatherNext
Google DeepMind's family of AI weather models, including a 2026 version that forecasts tropical cyclone tracks and strength in one system.
WebDev Arena
A leaderboard where humans compare two models' web pages side by side and vote for the better one.
WikiSkill
A Google Research method that gives an AI agent a growing written record of its own successes and failures, which it reads before starting new work.
Wiz
A cloud security company, now part of Google, whose research team regularly publishes findings about AI and cloud vulnerabilities.
Word error rate
The standard score for transcription accuracy: the share of words a speech recognition system gets wrong.
x402
A proposed web standard that lets a server charge a tiny payment for a single request, so software can buy things without a signup form.
XPeng
A Chinese electric carmaker that's now pouring money into humanoid robots through its robotics subsidiary.
Yann LeCun
A Turing Award winner and long-standing sceptic of today's language models.
YouGov
A British market research firm that runs large online surveys, often used by researchers to measure public attitudes to AI.
Z.ai
A Chinese AI company (formerly Zhipu AI) behind the GLM model family.
ZCode
Z.ai's low-cost coding agent, undercutting Western rivals on price.
Zeke Hausfather
Climate scientist whose measurements of AI energy use are among the few independent numbers available.
Zero Data Retention
A setting where an AI provider keeps nothing at all once your request has been answered.
Zero-click attack
An attack that needs no action from the victim at all, not even a click.
Zero-day
A security hole that the people responsible for fixing it do not know about yet.
ZeroBench
A set of 100 image analysis puzzles built to be hard enough that even the best models mostly fail them.
Zhang Yiming
Founder of ByteDance, the Chinese company behind TikTok, Douyin and the Seed AI research team.
Zig
A newer programming language aimed at fast, low-level software.
Nothing found. Try another word.