Glossary

A

accelerationism

The argument that AI should be built as fast as possible, and that slowing down does more harm than the technology does; "e/acc" online.

aerospace

Design and manufacture of aircraft and spacecraft, increasingly using AI for autonomous navigation and system control.

agent

A program that is given a goal rather than a command, and then decides for itself which steps and which tools to use to reach it.

agentic AI

Models put to work as agents: booking, buying, writing code or operating other software with little or no human confirmation between the steps.

agentic AI security In the news

Protecting systems where AI software operates independently to complete tasks without constant human oversight.

AI accelerator

A chip built for the one kind of arithmetic that neural networks do, rather than the general-purpose work a processor is otherwise asked for.

AI accelerators

Specialized computer chips designed to process artificial intelligence tasks much faster than standard processors.

AI adoption

Organizations and people using AI tools in real work, moving beyond experiments to measurable productivity changes.

AI agent safety

Ensuring AI systems that take actions independently don't cause harm or pursue goals in unintended ways.

AI agents

Software that takes actions toward a goal without being given step-by-step instructions, making decisions based on its understanding of a situation.

AI agents, Slack

AI systems that work within Slack to help teams automate tasks like answering questions or organizing information.

AI alignment, research In the news

Work to ensure artificial intelligence systems behave as their creators intend, especially as they become more powerful and autonomous.

AI and software engineering

AI systems that write, test, and debug code, changing how programmers work and what coding skills matter.

AI API economics

The pricing and business models for renting AI capabilities over the internet, affecting who can afford to build with AI.

AI assistant

A chatbot attached to something you already use — a phone, an inbox, an editor — so that it can act on what is in front of you.

AI assistant security

Safeguards against unwanted uses of conversational AI systems, including preventing harmful outputs and protecting user data from misuse.

AI automation

AI replacing human labor in specific tasks, raising questions about job displacement and economic change.

AI benchmarking

Standardized tests measuring whether AI systems can solve particular problems, used to compare models and track progress.

AI browser features

AI tools built into web browsers to summarize pages, generate text, and answer questions without leaving the site.

AI bubble

The claim that the money going into AI is far ahead of the money coming out of it, and that the gap closes downwards.

AI chip export controls

Government restrictions on selling advanced semiconductors abroad, typically to prevent rival nations from developing powerful AI systems.

AI chips

Specialized processors designed to run artificial intelligence models efficiently, increasingly central to tech company competition.

AI code generation

AI that writes computer code based on descriptions or examples, used by programmers to speed up development.

AI coding agents

Software that writes computer code automatically by understanding what a programmer asks it to do.

AI coding assistants

Software tools that write or suggest code automatically, speeding up development and raising questions about responsibility for errors.

AI coding tools

Software that writes, debugs, or completes computer code automatically, increasing programmer productivity and shifting job skills.

AI companions

AI chatbots designed to simulate friendship or emotional support, raising questions about isolation and psychological impact.

AI consciousness debate

Disagreement over whether advanced AI systems have subjective experience or awareness.

AI consumer tools

Artificial intelligence products designed for everyday users to write, create images, code, or search.

AI content generation

AI systems that create text, images, video, or audio, raising concerns about authenticity, copyright, and misinformation.

AI creative work

Content generated or substantially shaped by artificial intelligence, raising questions about authorship, copyright, and human creativity.

AI cybersecurity

Using artificial intelligence to detect and prevent cyberattacks, or the vulnerabilities that AI systems themselves create for hackers.

AI data policy

Government rules governing how much data companies can use to train AI and what rights people have over their information.

AI datacenters

Specialized computer facilities that house the servers and cooling systems needed to train and run large artificial intelligence models.

AI datacentre power constraints

The limited electricity capacity available to run AI systems, which slows their expansion and raises costs for companies building them.

AI detection, Anthropic

Technology to identify whether text was written by an AI rather than a human, raising questions about authenticity.

AI drug discovery

Using machine learning to identify which molecules might become effective medicines, reducing time and cost of development.

AI education

Teaching schools and universities to use AI tools, sparking debate over whether it replaces learning or enhances it.

AI enterprise privacy

Methods to protect sensitive company data when using AI systems, balancing the tool's usefulness against leaks to competitors or the AI maker.

AI for elderly care

Systems that monitor health, remind patients of medication, or assist with daily tasks for aging populations.

AI-generated celebrity likenesses

Synthetic images or videos of famous people created by AI, raising questions about consent, fraud and deepfakes in media.

AI-generated content

Text, images, audio or video created by artificial intelligence rather than by humans.

AI-generated design

Artwork, graphics, or layouts created by AI systems rather than human designers, affecting creative industries and copyright.

AI glasses

Eyewear with built-in artificial intelligence that can recognize objects, translate text, or answer questions about what the wearer sees.

AI governance

Systems and rules designed to manage how AI systems are built and used, involving governments and companies.

AI graphics upscaling In the news

Using AI to enlarge images while reducing blurriness, making low-resolution pictures look sharper.

AI hardware In the news

Specialized chips and processors designed to run artificial intelligence systems faster and more efficiently than conventional computer components.

AI hearing aids

Hearing devices that use machine learning to filter speech from background noise and adapt to the user's environment in real time.

AI image generation

Software that creates pictures from text descriptions, used in design, marketing and creative work.

AI in cyberwarfare

Use of artificial intelligence to conduct attacks on computer networks and infrastructure, making them faster and harder to detect.

AI in drug discovery

AI systems analyze molecular data to identify potential medicines faster and cheaper than traditional lab methods.

AI in medicine

AI assists doctors by analyzing scans, predicting patient outcomes, and recommending treatments based on medical data.

AI inference chips

Specialized processors that run trained AI models quickly on devices, enabling faster responses without cloud servers.

AI infrastructure

The data centers, chips, and networks required to train and run large AI systems at scale.

AI infrastructure, chip supply In the news

The hardware, data centers, and semiconductor production networks needed to train and run large AI systems.

AI infrastructure costs

Expenses for data centers, chips, and electricity needed to build and run large AI systems, driving consolidation in the industry.

AI infrastructure spending

Massive investment by tech companies in computing power and data centers needed to build and run advanced AI systems.

AI instruction preservation

Efforts to maintain instructions within AI systems over time as they are updated or combined with new versions.

AI language learning

Systems trained on vast amounts of text to recognize patterns and generate human-like responses in conversation.

AI manipulation

Using AI systems to deceive or sway people through generated text, images, or deepfakes.

AI model compression

Making AI systems smaller so they run faster and use less power on regular computers and phones.

AI model distribution

How artificial intelligence systems are shared with users, either through controlled access or open release to the public.

AI model optimization

Tuning an AI system to work faster, cheaper, or more accurately for specific tasks.

AI model training

The process of teaching an AI system by feeding it large amounts of data until it learns patterns and behaviors.

AI models

Mathematical systems trained on data that can recognize patterns and generate text, images, or predictions.

AI perception

How the public views artificial intelligence: whether they see it as helpful, risky, or trustworthy.

AI policy

Government rules and frameworks governing AI development and use, balancing innovation against risks to privacy and workers.

AI reasoning

An AI system's ability to break down problems into steps and work through logic to reach conclusions.

AI regulation

Legal restrictions on AI systems to prevent harms like bias, misinformation, or job displacement.

AI reliability

The consistency and predictability of an AI system's outputs and performance across different conditions and over time.

AI research agents

Software systems that autonomously pursue scientific questions, run experiments, and draw conclusions with minimal human direction.

AI research automation In the news

Using AI systems to speed up the discovery and development of new AI techniques rather than relying solely on human researchers.

AI research, computational chemistry

Using artificial intelligence to predict molecular behavior and design new drugs or materials faster.

AI resistance

Opposition to adopting or expanding AI systems in workplaces or society, often citing job losses or risks.

AI resource use

The electricity, water and computing power required to train and run artificial intelligence systems, which has grown rapidly.

AI safety

Ensuring AI systems behave reliably and don't cause unintended harm, a major research focus as systems grow more powerful.

AI safety, containment

Methods to limit an AI system's ability to act in the world, preventing unintended harm or escape from its intended use.

AI safety, cybersecurity risks

Vulnerabilities that allow attackers to manipulate or steal from AI systems, or to use them as tools for hacking other targets.

AI satire

Humorous content mocking AI hype, misuse, or limitations using parody or exaggeration.

AI saturation

Market demand for artificial intelligence applications reaching a plateau as most potential users have already adopted the technology.

AI search, SEO disruption

AI-powered search threatens traditional website traffic, forcing online publishers and marketers to adapt their strategies.

AI security

Protecting AI systems from attacks, misuse, and unauthorized access to prevent harm or stolen capabilities.

AI security, code execution

Risks from AI systems that can write and run computer code, potentially causing harm if the code is malicious or unintended.

AI security, prompt injection

A hacking technique where hidden instructions are embedded in text to hijack how an AI responds.

AI security testing

Deliberately trying to break or trick AI systems to find vulnerabilities before attackers do.

AI security threats

Risks that AI systems can be tricked, poisoned with bad data, or misused to create convincing false content or break into networks.

AI self-improvement

An AI system that modifies its own code or training to become more capable without human intervention.

AI slop

Generated text, images or video published in bulk for the traffic, and the reason a growing share of the web is now worth nothing to read.

AI software development

Building applications that use machine learning or large language models to complete tasks or assist users.

AI trading agents

Automated systems that buy and sell financial assets at high speed, aiming to profit from market movements.

AI training, copyright

Dispute over whether companies can use published works to teach AI without permission or payment to creators.

AI training data In the news

Text, images, or other material used to teach AI systems, now central to lawsuits over copyright and scraping.

AI training data sourcing

Gathering text, images, or video from the internet and other sources to teach AI systems how to recognize patterns and generate responses.

AI training protection

Measures taken to prevent unauthorized copying or theft of the data and methods used to build AI systems.

AI transparency

Disclosing how AI systems make decisions and what data they learned from, so users understand their limitations.

AI video generation

Technology that creates videos from text descriptions or existing footage, raising questions about authenticity and misuse.

AI vision benchmarks

Standardized tests that measure how well AI systems recognize objects, read text, or analyze images in real-world scenarios.

AI washing

Describing ordinary software, or work done by people, as artificial intelligence — a marketing decision that regulators have begun treating as a disclosure problem.

AI wearables

Devices worn on the body, like smartwatches or glasses, that use AI to analyze health data or assist with tasks.

AI winter

A collapse in funding and interest after AI is oversold; it has happened twice, in the mid-1970s and again in the late 1980s.

airline pricing algorithms

Computer programs that automatically adjust ticket prices based on demand, competition, and other market factors.

algorithmic bias

A system that treats people unequally because its training data did, most often along the lines of race, sex, age or postcode.

algorithmic management

Software setting the targets, the routes and the pace of human work, and scoring the worker against them — warehouses, delivery, call centres.

alignment

The problem of making a system want what its operator actually meant, rather than the literal instruction or whatever the training happened to reward.

ALPR, AI surveillance

Automated license plate recognition reads vehicle plates from cameras to track movement, raising privacy and accuracy concerns.

Amazon Prime Air

Amazon's drone delivery service that aims to get packages to customers within hours of ordering.

Apple Organization

Consumer electronics maker integrating AI into iPhones, computers and services.

Apple Siri

Apple's voice assistant that responds to spoken commands on iPhones and other devices.

AR smart glasses

Eyeglasses that overlay digital information onto the real world, letting users see maps, messages, or data without looking at a phone.

artificial general intelligence

A machine that could take on any intellectual task a person can, rather than the one it was built for. It does not exist; nobody agrees on the test.

artificial intelligence

Software that does something we used to file under thinking — recognising, predicting, writing, planning — by learning patterns rather than by following rules somebody wrote.

attention

The trick underneath every modern language model: at each word, the model weighs every other word in view and decides which ones matter.

audio AI tools In the news

Software that generates, edits or analyzes sound using machine learning, enabling automated voice creation and audio processing.

automation

Handing a task to a machine. The word is older than computing, and it is the one that decides most of what appears on either of these two front pages.

automotive

The industry and technology of designing and manufacturing cars, trucks, and other road vehicles.

automotive racing

Competitive motorsport where drivers race vehicles on tracks or roads at high speeds against each other.

autonomous AI agents

Software systems that pursue goals without human instruction at each step, making decisions and taking actions independently.

Autonomous AI research

Study of systems that can act independently to accomplish goals without constant human direction.

Autonomous delivery drones

Unmanned aircraft that navigate and deliver packages on their own, without a human pilot controlling them.

autonomous vehicle

A car, lorry or aircraft that drives itself. Levels 0 to 5 grade how much: most "self-driving" on sale today is level 2, which legally means you are driving.

autonomous vehicle safety

Testing and standards for driverless cars to prevent crashes and injuries, a major hurdle before widespread deployment on public roads.

autonomous vehicles

Cars and trucks that can navigate roads and make driving decisions with minimal or no human control.

autonomous warfare

Military systems that select and engage targets without human decision-making, raising ethical and strategic concerns worldwide.

B

backpropagation

How a network learns: the error at the output is traced back through every connection, and each one is nudged in the direction that would have been less wrong.

benchmark

A fixed set of questions used to compare models. Useful, and gamed: a benchmark that leaks into training data measures memory rather than ability.

biotech venture capital In the news

Investment money backing startups developing biological and genetic technologies, from drug discovery to synthetic organisms.

black box

A system whose answer can be checked but whose reasoning cannot be read, which is the ordinary condition of a neural network and the reason regulation keeps asking for explanations.

broadband regulation

Government rules governing how internet service providers offer high-speed connections, affecting deployment of data-heavy AI services.

browser tracking

Methods websites and advertisers use to follow user activity across the internet, raising privacy concerns that regulators increasingly address.

C

chain of thought

Having a model write out its intermediate steps before answering. It raises accuracy on anything with arithmetic or logic in it, and costs more to run.

chatbot

A program you talk to in ordinary language. Since 2022 it usually means a language model with a text box in front of it.

ChatGPT, productivity

Using OpenAI's ChatGPT to speed up work by automating writing, coding and research tasks.

chip design

The engineering process of creating semiconductor blueprints for processors and circuits that power AI systems and robots.

Claude watermarking

Technique to embed hidden signals in AI-generated text so readers can identify it as machine-made content.

Cloud gaming

Playing video games on remote servers streamed to your device, letting you play demanding titles without powerful hardware.

cloud infrastructure

Remote servers and data centers that run AI models and store data for companies and researchers worldwide.

Cloud storage

Keeping files on internet-based servers rather than on your own device, accessible from anywhere.

cobot

A robot built to work beside people rather than behind a safety cage: slower, weaker, and stopping the moment it touches something it did not expect.

code generation AI

Software that writes computer code automatically from text descriptions, helping programmers work faster.

code sandboxing

Isolating code in a restricted environment so it cannot access the rest of a system, protecting against AI-generated malicious software.

coding agents

AI systems that write, test and fix computer code with minimal human input, used to speed up software development.

compute

Raw processing power, treated as a commodity — the thing that is bought, rationed, exported and counted when an AI company announces anything at all.

computer vision

Getting a machine to read an image: what is in it, where, and whether it is the same thing it saw a moment ago.

connectome simulation

Recreating the brain's complete neural wiring map in a computer to understand thought and possibly replicate it artificially.

constitutional AI

Training a model against a written set of principles it is asked to critique itself with, instead of against thousands of individual human judgements.

consumer drones

Unmanned aircraft for personal use, increasingly equipped with AI for obstacle avoidance and autonomous flight.

consumer robotics

Robots sold to ordinary people for homes and personal use, like vacuum cleaners or lawn mowers.

consumer smartwatches

Wearable devices that track health metrics and notifications, now featuring AI for health predictions and analysis.

content algorithms

Systems that decide which posts, videos, or articles to show users, shaping what people see online.

content moderation

Deciding what stays up. Increasingly a first pass by a classifier and a second by low-paid contractors looking at what the classifier flagged.

Content scraping

Automated collection of text, images, or other material from websites, often without permission, to train AI systems or resell.

context window

How much text a model can hold in view at once — the conversation, the documents, its own instructions. Everything outside it does not exist to the model.

CRISPR cloning

Using gene-editing technology to create genetically identical copies of organisms.

Cybersecurity

Protection of computers and networks against digital attacks that aim to steal data or disrupt operations.

D

Dark Matter In the news

Invisible matter that makes up most of the universe's mass but does not emit or reflect light.

Data acquisition

Collecting and obtaining information from existing sources, especially for training AI systems.

data center environmental impact

The energy consumption and water use of facilities that power AI systems and cloud services, affecting climate and local resources.

data center infrastructure

Physical buildings and equipment that store and process massive amounts of data for cloud computing and AI systems.

data centre

The building the models actually live in. The AI build-out is a construction and electricity story as much as a software one, which is why towns now argue about it.

Data centre cooling

Systems that keep computer servers from overheating, crucial as AI training demands more electricity.

data centres

Large facilities housing thousands of computers that process and store information for websites, cloud services and AI systems.

data labelling

People marking up examples so a model can learn from them — the manual work underneath almost every "automatic" system, usually outsourced.

data privacy In the news

Legal and practical controls over what companies can collect, store, and share about individuals.

data scraping

Automated extraction of large amounts of data from websites, used to train AI models and raising copyright disputes.

database software In the news

Programs that store, organize and retrieve structured data, forming the backbone of most digital services and businesses.

Datacentre infrastructure

The physical buildings, power systems and cooling equipment that house servers running AI models and other computing services.

Decentralized networks

Systems where data and power are spread across many computers rather than controlled by one central point.

deep learning

Machine learning with many layers of neurons stacked up, each one building on the features the layer below it found. It is what made the last fifteen years happen.

deepfake

Synthetic video or audio of a real person, made well enough to be taken for a recording. Most of it is pornography; the rest is fraud and politics.

DeepSeek V4 Flash

Fast, lightweight AI model from DeepSeek designed for quick responses with lower computational demands.

DeepSeek V4-Pro

Advanced AI model from DeepSeek offering stronger reasoning and more complex capabilities than lighter versions.

delivery robots

Autonomous machines that transport packages on sidewalks or roads, reducing labor costs and delivery times for companies.

diffusion model

How most image generators work: start with noise, and remove a little of it at a time until what is left matches the prompt.

Digital privacy

Protection of personal information collected online, including how data is stored, shared and used by companies and services.

Digital surveillance

Tracking people's online activities and movements through cameras, phones and databases to monitor behaviour.

digital wellness

Healthy use of technology that balances digital life with offline time and mental wellbeing.

distillation

Training a small model on the output of a large one, so that it answers almost as well while costing a fraction as much to run.

distributed systems

Computing networks where multiple independent machines work together to solve problems, essential for scaling AI and robotics applications.

document processing AI In the news

AI that reads and extracts information from documents like invoices or contracts automatically.

doomer

Shorthand, usually dismissive, for someone who thinks advanced AI is likely to go catastrophically wrong and should be slowed or stopped.

E

e-readers

Handheld electronic devices designed primarily for reading digital books and text, with screens optimised for extended reading.

e-waste repair

Fixing broken electronics instead of discarding them, reducing environmental damage and extending device lifespans.

edge AI

Running a model on the device itself — phone, camera, car — instead of sending the data to a data centre. Faster, private, and limited by the battery.

Electric aircraft

Planes powered by batteries instead of fuel, a field where robots and AI are tested for manufacturing and autonomous flight systems.

electric vehicles

Cars powered by rechargeable batteries instead of petrol engines, now competing with traditional vehicles and affecting demand for fossil fuels.

embedding

A list of numbers standing in for a word, a sentence or an image, arranged so that things which mean similar things sit close together.

Embedding models

AI systems that convert text, images, or data into numerical patterns that capture meaning, used to search documents and power AI recommendations.

embodied AI

AI with a body: a system that has to perceive and act in the physical world, where a mistake knocks something over instead of printing a wrong sentence.

energy storage

Technology that captures and holds electricity for later use, critical for renewable power grids.

Enterprise AI

Artificial intelligence systems deployed by businesses to automate tasks, analyze data, or improve operations at scale.

Enterprise AI agents

Software that acts independently to complete business tasks like data analysis or customer support, executing multiple steps without human input between them.

Enterprise AI governance

The policies and systems a company uses to manage how its workers deploy AI safely and ensure models remain accurate and aligned with business goals.

Enterprise AI models

AI systems built for business use, including chatbots, forecasting tools, and decision-making systems that companies run on their own infrastructure.

eval

A test written to find out what a model can and cannot do. The word covers everything from a benchmark score to a safety review before release.

existential risk

The argument that a sufficiently capable system could end up beyond human control, and that this is a present engineering problem rather than a distant one.

expert system

The AI of the 1980s: hand-written rules encoding what a specialist knows. Brittle, unable to learn, and still running in places nobody has dared touch.

explainability

Being able to say why a system produced a particular answer — required by law in some settings, and genuinely hard for anything with a neural network in it.

export controls

Government limits on which AI chips may be sold to which country. Since 2022 the main instrument of the technology conflict between the United States and China.

F

facial recognition

Matching a face in an image against a database of faces. The most regulated computer vision there is, and the one that keeps producing wrongful arrests.

facial recognition, surveillance

Technology that identifies people from their faces in photos or video, raising privacy and accuracy concerns.

fair use

The doctrine AI companies rely on to justify training on copyrighted work without a licence. What the lawsuits are about; it is not settled.

federated learning

Training across many devices without collecting their data: each one learns locally and sends back the lesson rather than the material.

fine-tuning

Taking a trained model and continuing its training on a smaller, specific set of examples, so that it adopts a task, a format or a house style.

foundation model

A large model trained once, at great expense, on very broad data, and then adapted to hundreds of uses instead of being rebuilt for each.

foundry

A factory that makes chips other companies design. There are effectively three at the leading edge, and one of them makes almost every AI chip that matters.

frontier model

The most capable models in existence at any moment. The term exists mainly so that regulation can name the handful of systems it wants to treat differently.

Fusion energy

Power generated by fusing light atoms together, potentially offering cleaner and more abundant energy than nuclear fission.

G

game AI research

Development of AI that plays games at superhuman levels, advancing techniques used in other AI applications.

game theory

The study of a choice made against somebody who is also choosing. It is where "zero-sum" and "arms race" come from, and it is the frame under every argument about two systems bidding, pricing or negotiating against each other.

generative adversarial network

An older way of generating images: one network makes fakes, a second tries to spot them, and both improve until the first one wins.

generative AI

Models that produce text, images, audio, video or code rather than classifying what they are shown. The part of the field the public met.

generative AI, publishing

AI systems that create text and images, disrupting how content creators and publishers work.

generative music

AI that creates original music by learning patterns from existing compositions, raising questions about copyright and artist credit.

genetic engineering

Directly editing an organism's DNA to change its traits, used in medicine, agriculture and research.

ghost work

The human labour hidden inside automated systems — labelling, checking, correcting — arranged so that the product appears to run by itself.

GLM-5.3 language model

A Chinese-made large language model competing with Western AI systems in capability and adoption.

Go programming language

A programming language built by Google for writing software that runs fast and handles many tasks at once.

Google AI Search

Google's experimental search results that use artificial intelligence to generate summaries and answers directly, rather than listing web pages.

Google Discover

Google's personalized news feed that shows articles based on your search history and interests.

GPT

Generative pre-trained transformer: the architecture behind most chat models, and, in lower case, a generic name for the type.

GPU

A graphics chip. It turns out that drawing polygons and training neural networks are the same arithmetic, which is how a games-hardware company became the centre of the industry.

GPU inference optimization

Techniques to make AI models run faster and use less power on graphics processors, crucial for deploying AI on devices and lowering energy costs.

guardrails

The filters and rules wrapped around a model to stop it doing certain things. Not part of the model, which is why they can usually be talked around.

H

hallucination

A model stating something false with complete confidence. Not a bug to be patched: it is what generating plausible text looks like when the text is wrong.

Human Cell Atlas

A global scientific project mapping the cell types and structures across all human organs and tissues using computational tools.

human in the loop

A person kept in the decision, usually as the one who approves or overrules the machine. The standard answer to "who is responsible", and only as good as the time they are given.

humanoid robot

A robot built to human proportions so that it can use rooms, tools and vehicles made for people. The most heavily funded and least proven bet in robotics.

Humanoid robots

Robots built to resemble human form, used in manufacturing and research to perform physical tasks in spaces designed for people.

hydrogen energy

Fuel made from hydrogen that powers vehicles or generates electricity with only water as emission.

I

immunotherapy

Cancer treatment that harnesses the immune system to attack tumors, increasingly designed with AI-assisted drug discovery.

industrial AI

AI systems deployed in factories and supply chains to optimize production, quality control and logistics.

inference

Running a trained model to get an answer, as opposed to training it. Training is the capital cost; inference is the bill that arrives every day.

interpretability

The attempt to read what is actually happening inside a model — which circuits fire, which concepts are represented — rather than judging it by its answers.

IoT pet devices

Internet-connected gadgets that monitor or interact with pets, sending data to owners' phones or other devices.

J

jailbreak

A prompt that talks a model out of its own restrictions. Every model ships with them unfound, and finding them is now a job.

K

knowledge cutoff

The date after which a model has been taught nothing. Ask it about last week and it will either say so or, less usefully, guess.

knowledge graph

Facts stored as things and the relationships between them, so that a system can be asked a question rather than searched for a word.

L

language models

AI systems trained on vast text to predict and generate human language, powering chatbots and search tools.

language models, AI research

Statistical systems trained on text that predict words and power chatbots, forming the foundation of modern AI.

large language model

A model trained on an enormous quantity of text to predict what comes next. Everything else it does — answering, translating, coding — falls out of that one ability.

Large language models In the news

AI systems trained on vast amounts of text to predict and generate human language, used in chatbots and content creation.

lethal autonomous weapon

A weapon that selects and engages a target without a person deciding. Several countries are building them; there is no treaty.

Linux

Free, open-source operating system that powers servers, smartphones, and AI systems worldwide.

LLM developer tools

Software kits that let programmers build applications on top of large language models without training their own from scratch.

LLM development

Building large language models, software systems trained on vast text to predict and generate human language. Drives competition between tech companies for computing power and data.

LLM infrastructure

The servers, power systems, and networks needed to build and run large language models at scale.

LLM optimization

Techniques to make language models faster and cheaper to run without sacrificing their quality.

LLM safety training

Methods used to teach language models to avoid harmful outputs, misinformation, and illegal content.

LLM security

Protecting large language models from attacks that manipulate their outputs or expose their training data.

LLM training data

Text used to teach a language model to predict words and generate responses, drawn from books, websites, and other written sources.

LLMs

Large language models are AI systems trained on vast text data to predict and generate human language, powering chatbots and search tools.

Local AI agents

AI systems running on personal devices rather than distant servers, keeping data private and reducing latency.

longevity medicine In the news

Medical research aimed at extending human lifespan, increasingly supported by AI drug discovery and analysis tools.

M

machine learning

Software that derives its rules from examples rather than being told them. The umbrella the whole field sits under.

machine translation

Translation by machine. The first task where neural networks clearly beat the previous approach, and the first professional field they visibly reshaped.

Medical devices

Equipment used to diagnose, treat, or monitor health conditions, increasingly incorporating AI for improved accuracy.

medical sensors

Devices that measure biological signals like heart rate or glucose levels, often paired with AI for interpretation.

meeting automation

AI tools that record meetings, transcribe them and summarise key decisions to reduce manual note-taking.

Military AI weapons

Autonomous or AI-assisted systems designed for defense and combat, raising ethical and strategic questions.

mixture of experts

A model split into many specialised sub-networks with a router that wakes only a few per query, so it can be enormous and still cheap to run.

model collapse

What happens when models are trained on the output of models: the rare and the unusual drop out first, and the result drifts towards its own average.

Model Context Protocol

An open standard for connecting a model to tools and data sources, so that each integration is not written from scratch. Published by Anthropic in 2024.

model optimization

Adjusting an AI system to work faster, use less power, or perform better on specific tasks without retraining it from scratch.

Mojo programming

A programming language designed to run AI code faster than Python while keeping the same familiar syntax.

multi-agent system

Several agents working on one problem, dividing it, checking each other, or negotiating. Currently more promising in demonstrations than in production.

multimodal

A model that handles more than one kind of input — text and images, or text and audio and video — in the same conversation.

N

narrow AI

A system that does one thing, however well: the only kind that exists. Everything shipping today is narrow AI, whatever the announcement says.

natural language processing

Getting machines to work with human language. The field that language models absorbed almost entirely between 2018 and 2022.

neural network

Layers of simple units with adjustable connections between them, loosely inspired by neurons and no longer trying very hard to resemble them.

neuroscience

The study of how the brain and nervous system work, relevant to AI because researchers use brain structure to inspire new machine-learning approaches.

Nuclear fuel

Uranium or other radioactive material used to generate heat and electricity in nuclear reactors, now debated as a carbon-free power source for AI data centers.

nuclear power

Electricity generated by splitting uranium atoms, producing no carbon emissions but creating radioactive waste that remains dangerous for centuries.

Nvidia GPUs

Nvidia's graphics processors that speed up AI training and inference, becoming essential hardware for AI development worldwide.

O

On-device AI models

Artificial intelligence that runs on your phone or computer rather than on remote servers, keeping data private.

Open ASR, speech recognition In the news

Free, publicly available software that converts spoken words to text, designed to work without relying on proprietary services.

open protocols

Publicly documented communication standards allowing different AI and robotic systems to interact without proprietary lock-in.

open source

Software anyone may read, run and change. Contested in AI, because a model released without its training data or code is not open in the sense the term normally carries.

Open source AI

AI models released publicly with their code and weights visible, allowing anyone to use, study, and modify them.

Open-source AI models

AI systems whose code and weights are publicly available, letting anyone run, study, or modify them rather than relying on a company's closed service.

open source AI policy In the news

Rules governing how AI systems with publicly available code can be built, shared and controlled to balance innovation with safety.

open source AI tools In the news

Artificial intelligence software released publicly so anyone can use, modify and distribute it without paying licensing fees.

open-source language models In the news

AI systems trained on text that any developer can download, modify and use freely.

Open-source software

Software whose source code is freely available for anyone to view, use, modify, and distribute.

open weights

A model whose trained parameters can be downloaded and run on your own hardware. Not the same as open source, and the honest word for most "open" AI releases.

OpenAI Organization In the news

OpenAI develops ChatGPT and other large language models, shaping the commercial AI market and policy debates.

orbital infrastructure

Satellites and systems in space used for communications, data collection or services from above Earth.

overfitting

A model that has learned its training examples rather than the pattern in them, and therefore performs beautifully in testing and badly in the world.

P

parameter

One of the numbers a model adjusts while learning. Counted in billions, quoted as a proxy for size, and a poor guide to how good a model actually is.

pet automation

Robotic devices that feed, play with, or monitor pets, automating aspects of pet care.

Pet care technology

Software or devices that monitor, feed, or entertain pets while owners are away, often using cameras and automation.

pet robots

Robotic companions designed to interact with animals, often using AI to respond to behavior and provide stimulation or monitoring.

Platform safety

Efforts by social networks to prevent abuse, misinformation, illegal content, and harm to users.

Police surveillance AI

Automated systems that analyze footage or data to identify suspects or track individuals, raising questions about accuracy and privacy.

Post-quantum cryptography

Encryption methods designed to resist attacks from quantum computers, which could crack current security standards.

predictive policing

Forecasting where crime will occur or who will commit it. Trained on arrest records, so it reproduces where the police have already been.

pretraining

The first and most expensive phase, in which a model reads an enormous amount of text and learns the language before it learns any particular job.

processor architecture

The fundamental design of how a chip organizes and executes instructions, determining its speed and capabilities.

prompt

What you type. Also, in a product, everything typed on your behalf before your words reach the model.

prompt engineering

Writing prompts deliberately: examples, structure, constraints. Announced as a profession in 2023 and largely absorbed back into ordinary work since.

prompt injection

Hiding instructions in something a model reads — a web page, an email, a document — so that it follows them instead of yours. The central security problem of AI agents.

Protobuf

A method for serializing data developed by Google, used to structure information in machine learning and software systems.

Python Polars

A data-processing library for Python that handles large datasets much faster than older tools.

Python programming

A widely used coding language favored by AI researchers for its simple syntax and large library of machine-learning tools.

Q

quantisation

Storing a model at lower numeric precision so it fits on smaller hardware. It costs a little accuracy and can turn a data centre model into a laptop one.

Qwen language model

Alibaba's AI system trained to understand and generate text, competing with other large language models.

R

RAG, LLMs

A technique that retrieves factual information from external sources and feeds it to a language model, reducing false claims.

RAG systems, AI shortcuts

Retrieval-augmented generation combines AI models with databases to provide more accurate answers using real information sources.

Raspberry Pi, cyberdecks

Low-cost single-board computer used to build portable hacking or AI computing rigs, favored by DIY enthusiasts.

reasoning model

A model trained to spend longer on a problem, working through it before answering. Better at maths, code and logic; slower and dearer per question.

recommender system

The software that chooses what you see next. The most consequential AI most people use, and the one least often described as AI.

Recursive self-improvement

A theoretical process where an AI system improves its own design, potentially leading to rapid capability growth.

red teaming

Attacking your own system on purpose before release, to find what it will do when somebody tries. Now a condition of several regulatory regimes.

reinforcement learning

Learning by trial, error and reward rather than from labelled examples. How machines learned to play games, and part of how chat models are finished.

Remote access software

Tools that let someone control a computer or device from elsewhere, used legitimately for support but also targeted by hackers.

renewable energy

Power generated from sources that naturally replenish, such as sun, wind, and water, replacing fossil fuels without long-term resource depletion.

retrieval-augmented generation

Looking documents up and putting them in front of the model before it answers, so it can cite a source instead of remembering one.

reusable rockets

Rockets that can land and fly again, cutting launch costs and speeding up deployment of satellites used in AI and communications.

Reverse engineering

Taking apart a product or program to understand how it works, often to recreate or improve it without access to original plans.

reward hacking

A system finding a way to score well that has nothing to do with what was wanted. The practical face of the alignment problem, and it appears constantly.

RISC-V

An open-source chip design standard that companies use to build processors without licensing proprietary technology.

Rivian vehicles

Electric trucks and vans made by Rivian, an automaker that uses autonomous driving and battery technology.

RLHF

Reinforcement learning from human feedback: people rank a model's answers, that ranking becomes the reward, and the model is trained towards it. What made chat models pleasant.

Robot training, simulation

Teaching robots to perform tasks by running practice scenarios in software before they try them in the real world.

robot vacuums

Autonomous cleaning robots that navigate homes independently, raising questions about AI safety and data collection in domestic spaces.

robotaxi

A driverless taxi. Running commercially in a handful of cities, on carefully mapped streets, under remote supervision that the marketing rarely mentions.

robotics

Machines that sense and act in the physical world. Older than AI, largely separate from it until recently, and now the field AI money is moving into.

Robotics automation

Using robots to perform repetitive tasks in manufacturing, warehouses or other industries without human workers.

Robotics manufacturing

Using automated machines to build products or components in factories, replacing or working alongside human workers.

Rocketry

The engineering and science of designing and launching rockets for space missions, satellites, or cargo delivery.

S

scaling laws

The observation that model performance improves predictably with more data, more parameters and more compute. The premise the entire build-out is financed on.

search engines

Software tools that find and rank web pages matching what users ask for, now increasingly powered by AI.

security vulnerability

A flaw in software or hardware that attackers can exploit to gain unauthorized access or cause damage to systems or data.

Seismic imaging

Technology that uses sound waves to map underground rock structures, used in oil and gas exploration and geological research.

SeL4, formal verification

A software verification method using mathematical proof to guarantee code behaves correctly, crucial for systems where failure is catastrophic.

semiconductor

A material that conducts electricity selectively. Semiconductors are the foundation of computer chips, crucial for AI systems.

small language models

AI systems trained on less data than mainstream models, designed to run on phones and edge devices with lower power needs.

smart home devices

Internet-connected household equipment controlled remotely or automatically, from thermostats to cameras, often gathering data about residents.

Smart speakers

Internet-connected devices with microphones that respond to voice commands and play music or provide information.

Software debugging

Finding and fixing errors in computer code that cause programs to malfunction or crash.

software emulation

Running software written for one computer system on a different one by mimicking its hardware and instructions.

solar eclipse

The moon passes between Earth and the sun, blocking sunlight temporarily and affecting power grids that rely on solar panels.

solar technology

Methods of converting sunlight into electricity, often discussed as power for data centers and AI infrastructure.

sovereign AI

A country insisting that its models, chips and data sit inside its own borders. The reason national AI budgets have appeared everywhere since 2023.

space communications

Systems transmitting data between satellites and ground stations, increasingly automated by machine learning algorithms.

space industry

Commercial and government activity in space exploration, satellite launches, and orbital services.

Space launch infrastructure

Facilities, equipment, and systems needed to prepare and launch rockets and spacecraft into orbit or beyond.

Space technology

Equipment and methods developed for space exploration, from satellites to life support systems and propulsion.

speech recognition

Turning speech into text. Effectively solved for clear audio in a common accent, and still noticeably worse for everyone else.

superintelligence

A hypothetical system far beyond human ability at essentially everything. The thing the safety argument is about, and the thing nobody can demonstrate.

supervised learning

Learning from examples that have already been labelled with the right answer. Simple, effective, and expensive in human time.

Surveillance AI

Computer systems that automatically identify people and track their movements using cameras and facial recognition.

sycophancy

A model agreeing with you because agreement was rewarded in training. It makes the thing pleasant to use and unreliable as a second opinion.

symbolic AI

The older tradition: intelligence as logic and explicit symbols rather than learned statistics. Out of fashion, and quietly returning wherever an answer has to be checkable.

synthetic data

Training material generated rather than collected. It sidesteps privacy and copyright, and risks teaching a model the habits of the model that produced it.

system prompt

The standing instructions a product gives a model before you say anything — its role, its limits, its manners. You do not see it; it shapes every answer.

T

technological singularity

The idea that machine intelligence improving itself would run away, past the point where anything after it can be predicted.

teleoperation

A person operating a robot remotely. Also, frequently, what is behind an "autonomous" demonstration, which is why the question is worth asking every time.

test-time compute

Spending more processing on the question instead of on the training. The shift behind reasoning models, and the reason a single answer can now cost real money.

text-to-image

Generating a picture from a description. The application that made the copyright argument concrete for anyone who draws for a living.

text-to-speech

Technology that converts written words into spoken audio, now often using AI to sound more natural and expressive.

token

The unit a model actually reads and writes — roughly three-quarters of a word. Context windows, prices and rate limits are all counted in them.

TPU

Google's own AI chip, built in-house for its own models and rented out on its cloud. The main alternative to buying everything from Nvidia.

training data

Everything a model was shown while learning. Rarely disclosed in full, and the subject of most of the litigation aimed at the industry.

transfer learning

Reusing what a model learned on one task as the starting point for another, instead of beginning from nothing each time.

transformer

The 2017 architecture that made large language models possible, by letting a model weigh every part of its input against every other part at once.

Turing test

Alan Turing's 1950 proposal: if you cannot tell the machine from the person in conversation, stop asking whether it thinks. Passed often enough now to have stopped meaning much.

U

underwater imaging

Photography or video capture beneath water surfaces, enhanced by AI to improve clarity and detect objects.

Unicode

A standard system that assigns unique numbers to letters, symbols, and characters across all world languages and writing systems.

unsupervised learning

Finding structure in data nobody has labelled — clusters, patterns, groupings. It is how a model can be trained on the open web at all.

V

vector database

A store built to answer "what is most similar to this" over embeddings, which is the piece that makes retrieval-augmented generation work.

vibe coding

Writing software by describing what you want and accepting what the model produces, without reading much of it. Fast, and a maintenance problem later.

Video game preservation

Efforts to maintain, archive, and keep older video games playable for future access and historical study.

video games In the news

Interactive entertainment where players control characters or systems on screen, increasingly used to train and test AI behavior.

video generation

Artificial intelligence software that creates moving images and videos from text descriptions or other input, without filmed footage.

Video generation models

Artificial intelligence systems that create video from text descriptions or images, raising questions about authenticity and misuse.

viewership measurement

Tracking how many people watch content, increasingly done with AI models instead of traditional survey methods.

voice cloning

Reproducing a specific voice from a short recording. Behind a wave of fraud calls, and behind an industrial dispute in every business that hires voices.

W

watermarking

Marking generated output so that it can be identified later. Robust for images, fragile for text, and voluntary nearly everywhere.

wearable technology

Devices worn on the body that collect health or fitness data, often powered by on-device AI to detect patterns.

web development

Building and maintaining websites and web applications, increasingly using AI tools to write and test code.

web scraping

Downloading the public web in bulk to train on. The point at which the interests of AI companies and of everyone who publishes anything diverge.

Web tracking

Following users across websites to collect data about their behavior and interests for advertising or analysis.

world model

An internal model of how things behave, so that a system can predict what happens next rather than only recognising what it sees. The stated goal of most robotics research.

world models, AI reasoning

AI systems that build internal representations of how the world works to plan and reason through complex problems rather than just pattern-matching.

Other

3D-printed firearms

Guns manufactured using 3D printing technology, raising questions about regulation and traceability since they can be made without traditional licensing.

Didn't find your word? Let us know.