AI terminology, plainly explained.

Training is the (extremely expensive, one-time) process of teaching a model by having it learn patterns from huge datasets. Inference is what happens every time you use the finished model — it's already trained, and it's just generating a response to your prompt.

The internal numeric values a model adjusts during training — its "settings," loosely speaking. Bigger isn't automatically better: parameter count is one factor in how capable a model is, alongside training data quality and architecture.

Taking an already-trained model and further training it on a narrower dataset to specialize it — customer-support tone, a company's internal docs, a specific coding style. It's how a general-purpose model gets adapted for a specific job without training a new one from scratch.

A way of converting text (or images) into a list of numbers that captures its meaning, so a computer can compare how similar two pieces of content are mathematically. It's the technology underneath most AI search and recommendation features.

A database built to store and search embeddings efficiently, finding the most similar results out of millions in milliseconds. It's the storage layer behind most retrieval-augmented (RAG) systems.

A technique where a model looks things up in an outside source — a document set, a database, the live web — before answering, instead of relying only on what it learned during training. It's how AI tools stay current and can cite where an answer came from.

A model trained to work through a problem step by step before answering, rather than replying immediately — closer to "thinking out loud" than pattern-matching a reply. OpenAI's o-series and GPT-5's "thinking" modes, Claude's extended thinking, and Gemini's reasoning modes are all built this way, trading extra time and cost for meaningfully better accuracy on hard problems.

The amount of extra "thinking" a model does at the moment you ask it something, as opposed to computing power spent during training. Reasoning models can be given more test-time compute for a harder question — effectively letting them think longer — which is a big part of how they push accuracy higher.

An architecture where a model is built from many smaller specialized sub-networks ("experts"), and only a handful activate for any given input. It lets a model have a huge total parameter count while keeping the cost of answering any one question much lower — several open-weight models in our directory use this design. See examples in our Open Source guide →

A type of generative model that creates images, video, or audio by starting from random noise and gradually refining it into a coherent output, guided by a text prompt. It's the architecture behind most image and video generators — Stable Diffusion, FLUX, Midjourney — as opposed to the token-by-token prediction large language models use for text. See open-weight diffusion models →

A lightweight way to fine-tune a model by training a small set of add-on weights instead of the entire model. It's dramatically cheaper and faster than full fine-tuning, which is why there's a huge ecosystem of small, swappable LoRA files that customize image models like Stable Diffusion for a specific style, subject, or character.

Shrinking a model's numeric precision — say, from 16-bit down to 4-bit — so it takes up less memory and runs faster, at some cost to accuracy. It's what makes it possible to run large open-weight models on a laptop or a consumer GPU instead of needing a data center. Tools for running quantized models locally →

The basic structure behind modern AI: layers of simple mathematical units, loosely inspired by the brain, whose connections are tuned during training until the whole network gets good at a task. An LLM is a very large neural network; its tuned connections are its parameters.

The neural-network design behind nearly every major language model, introduced in a 2017 Google research paper titled “Attention Is All You Need.” Its key trick, called attention, lets the model weigh which words in a passage matter most to each other — which is what makes long, coherent text possible.

The first, enormously expensive stage of building a model: feeding it vast amounts of text, code or images so it learns general patterns. Everything after that — fine-tuning, human feedback, safety training — adjusts a model that pre-training already built.

A training step where people rank or rate model answers, and the model is nudged toward the ones people prefer. It's a big part of why chat models feel helpful and polite instead of just autocompleting text — and also one source of sycophancy, since people tend to rate flattering answers well.

A way of training reasoning models where the reward comes from an automatic check — did the math answer match, did the code pass its tests — instead of human opinion. It works best where answers can be verified, which is why recent reasoning gains show up most in math and coding.

Training a smaller “student” model to imitate a bigger “teacher” model's outputs, so the small one captures much of the big one's skill at a fraction of the size and cost. It's how many compact, fast models are made — though AI providers' terms often limit using their outputs to train a competing model.

Training data that was itself generated by AI rather than collected from people or the real world. It can fill gaps (rare cases, private topics) cheaply, but a model trained too heavily on its own kind of output can drift and lose variety, so quality control matters.

An AI system that doesn't just respond to a single prompt, but plans and carries out a multi-step task on its own — browsing, running code, editing files, calling other tools — checking its own progress along the way. The industry hasn't fully settled on one precise definition, but the core idea is autonomy over multiple steps, not just one reply.

An open standard, introduced by Anthropic in late 2024, that lets AI models connect to external tools and data sources — your files, a database, a project-management app — through a common interface instead of a one-off integration for every product. It's become a de facto standard adopted well beyond Anthropic's own products. Read the official spec →

A pre-built integration that lets an AI assistant read from or act inside one of your existing apps — Google Drive, Slack, GitHub, a calendar — without you building the integration yourself. Connectors are usually built on MCP under the hood; "connector" is the friendlier, user-facing name you'll see when turning one on inside a product like Claude or ChatGPT.

An open standard, created by Google and donated to the Linux Foundation in 2025, that lets AI agents built by different companies discover each other and hand off work — agent talking to agent, rather than agent talking to a tool. It's the horizontal counterpart to MCP: MCP connects one agent to its own tools and data; A2A connects that agent to other, separate agents. See the spec →

Emerging standards for letting an AI agent actually complete a purchase on your behalf, safely. Google's AP2 (Agent Payments Protocol) uses cryptographically-signed proof that you authorized the purchase; OpenAI and Stripe's ACP (Agentic Commerce Protocol) is what powers Instant Checkout inside ChatGPT. Both launched in beta in September 2025 and are already processing real transactions, not just pilots — though the space is still early and standards are likely to keep shifting.

The official extension to MCP (launched January 2026) that lets a connected tool render actual interactive UI inside a chat — a form, a map, a booking calendar — instead of just returning text. It formalizes patterns pioneered by an earlier open-source project called MCP-UI.

The ability for a model to call an outside function mid-conversation — run a calculation, search the web, query an API — and use the result in its answer. It's the basic building block that agentic behavior is built on top of.

The layer of software that coordinates multiple AI calls, tools, or agents to complete a larger task — deciding what runs when, passing results between steps, handling failures. It's the "management layer" sitting above any single model call.

A setup where several AI agents, often with different roles or specialties, work together (or check each other's work) on one task instead of a single model doing everything. It's an emerging pattern for complex work, though it adds coordination overhead and cost.

A model writing out its intermediate reasoning steps before giving a final answer, rather than jumping straight to a conclusion. It generally improves accuracy on multi-step problems, and it's the visible form of what reasoning models do internally — sometimes shown to the user, sometimes hidden.

Instructions set by the developer or platform, not the end user, that shape how a model behaves throughout a conversation — its persona, its rules, what it should refuse. It sits "above" whatever you type and typically takes priority over it.

Rules and technical checks layered around a model to keep it from doing things it shouldn't — leaking private data, giving dangerous instructions, going off-topic. They range from simple keyword filters to separate AI models that review outputs before they reach the user.

The practice of deliberately managing what information a model has access to at any moment — which documents, which prior messages, which tool results — rather than just writing a clever prompt. As context windows and agent tasks have grown, this has become as important as prompt wording itself.

AI that acts through a physical body — a humanoid robot, an arm, a warehouse machine — instead of only producing text or images on a screen. It combines a model's reasoning with real-world perception (cameras, sensors) and motor control, which makes it far harder to get right than a chatbot: the model has to handle a physical world that doesn't wait for the next prompt. See the labs and robots building it on our World AI map →

Describing what you want in plain language and letting AI write the code, often accepting the result without reading it closely. The term was coined by AI researcher Andrej Karpathy in early 2025 and was later named Collins Dictionary's 2025 Word of the Year. It's great for prototypes and personal tools; for anything handling money, logins or customer data, a human still needs to review the code.

Reusable folders of instructions, scripts and reference files that an AI agent loads only when a task calls for them — for example, a “write our brand-style report” skill. Anthropic introduced the idea in late 2025 and published it as an open format, so it isn't tied to a single product.

An AI model operating a computer the way you do: looking at the screen, moving the cursor, clicking and typing to get a job done. Anthropic introduced it as a public beta in October 2024; other labs have shipped their own versions. It's powerful but slower and less reliable than a direct integration, so use it with care on anything important.

A web browser with an AI agent built in that can read pages, fill in forms, click through steps and complete tasks for you, instead of only summarizing what's on the page. Perplexity's Comet and OpenAI's ChatGPT Atlas are examples. Because it acts on web content it didn't write, it is especially exposed to prompt injection.

A design where a person must approve or review the AI's work at important steps — before an email is sent, a payment is made, or code is deployed. The more an agent can do on its own, the more it matters to choose which steps still need a human yes.

An agent mode where the AI spends minutes (not seconds) searching many sources, reading them, and writing a long, cited report. It's far more thorough than a normal chat answer, but you should still open the cited sources — it can misread a page or lean on weak ones.

A feature that lets an assistant remember facts about you — your job, your preferences, past projects — across conversations instead of starting fresh each time. It makes answers more personal, but it also means more of your information is stored; check where it lives and how to view, edit or switch it off.

"Open weights" means you can download and run the trained model yourself. "Open source," in the strict sense, means the training data, code, and process are public too — not just the finished model. The two get used interchangeably in marketing, but they're not the same thing. See which is which, model by model →

Self-hosted means running a model on your own hardware, or a server you control — more setup, but full data control and no per-use fees. Cloud-hosted means calling a model through someone else's API — easier to start, but you're sending your data to that provider and paying per use. Tools for running models locally →

An API is a raw way for developers to call a model programmatically and build it into their own product. An app is the finished, ready-to-use product built on top of that API — most people use apps; only developers typically touch the API directly.

A pricing model where a limited version is free forever, and a paid tier unlocks higher usage limits, more capable models, or extra features. Nearly every major AI product uses some version of this.

A cap on how many requests (or tokens) you can send in a given period — per minute, per day. Hitting one doesn't mean something's broken; it usually just means waiting, or upgrading to a tier with a higher limit.

Paying per token, per request, or per minute of usage rather than a flat monthly fee. It's common for API access, and can be far cheaper — or far more expensive — than a subscription depending on how heavily you use a tool.

Whether a provider keeps your conversations, and whether it uses them to train future models. Enterprise and paid-API tiers typically default to not training on your data; free consumer tiers more often do — the setting is usually in your account's privacy or data-controls page, and it's worth checking rather than assuming.

A standardized test used to compare model performance — coding tasks, math problems, reasoning puzzles. Useful for a rough sense of capability, but a model that tops one benchmark won't necessarily fit your actual use case; benchmarks are also increasingly targeted by training, so treat any single number skeptically.

Getting stuck with one AI provider because your workflows, prompts, or integrations are built specifically around it, making it costly to switch. Using open standards (like MCP) and keeping prompts and data portable are common ways teams try to reduce it.

A provider feature that stores the unchanging start of your prompt — a long system prompt, a big document — so repeat requests don't pay full price (or wait full time) to process it again. If you reuse the same large context many times, it can cut both cost and delay noticeably.

Paying a fixed amount per person per month, no matter how much each person uses the tool. It's easy to budget, but you pay for seats that sit idle — and many products now combine seats with usage caps or credits, so read what “unlimited” actually covers.

A link that pays the site that shared it a commission if you sign up or buy. Some links on 10X AI Expo may be affiliate links; they never change which tools we list or how we describe them, and we disclose them. Our full policy is in the disclosure page.

A security audit report from an independent firm, showing that a company has controls in place for protecting customer data. “Type I” checks the controls at one point in time; “Type II” checks that they worked over a period (usually months). It's a useful box to tick when picking a business AI tool, but it isn't a guarantee of safety.

Efforts to make sure a model's behavior actually matches what its developers and users intend — helpful and honest, without pursuing harmful shortcuts to reach a goal. It's the umbrella term most other safety work, like red-teaming and interpretability, sits under.

Deliberately trying to make a model misbehave — get around its safeguards, produce harmful content, or take an action it shouldn't — before real users can, so the weakness gets fixed first. Most major labs do this internally, and increasingly bring in outside researchers to try too.

Informal term for the most capable AI models available at a given time — the ones pushing the edge of what's possible. Frontier models get extra safety testing largely because their capabilities are the least well understood, including by the labs that built them.

A lab's own public commitment to pause, restrict, or add extra safeguards to a model if it crosses certain capability thresholds — for example, the ability to meaningfully help someone build a bioweapon or cyberweapon. Several major labs publish one. They're voluntary policies a company sets for itself, not law, so how strictly they're followed varies by company.

Research aimed at understanding what's actually happening inside a model when it produces an answer, instead of treating it as a black box. It's still a young field — even the researchers building today's frontier models can't fully explain many of their outputs — but it's considered one of the more promising paths to catching problems before they reach users.

A case where a deployed AI system did something harmful, dangerous, or clearly unintended — from a chatbot giving dangerous advice to an autonomous agent taking an action nobody approved. There's no single, mandatory place labs must report these today, so what becomes public depends heavily on which company chooses to disclose, or what outside researchers or journalists independently uncover.

Government-run bodies — the UK's AI Security Institute and a US counterpart among the first — that test frontier models for safety issues, sometimes before public release, working alongside (not instead of) the labs' own internal testing. Several other countries have since stood up their own.

Shorthand for an active, genuinely unsettled public argument about how worried to be about AI, and about what. Roughly: one camp focuses on nearer-term harms already happening — bias, misinformation, job disruption, security. Another focuses on longer-term risks as models get more capable and autonomous. Most researchers in the field take both seriously, to different degrees — treat any single source that claims the debate is settled with skepticism.

An attack where hidden instructions are planted in content an AI reads — a web page, an email, a shared document — so the AI follows the attacker's orders instead of yours. It is widely ranked as the top security risk for AI applications, there's no complete fix yet, and the risk grows when an agent can take actions or reach private data.

A prompt crafted to trick a model into ignoring its safety rules, usually by role-play, tricks of wording, or a long chain of small steps. Jailbreaks target the model's own limits; prompt injection targets an app through the content it processes. Labs patch known ones, and new ones keep appearing.

A realistic fake image, video or voice recording made with AI, usually of a real person saying or doing something they never did. Uses range from harmless parody to fraud and harassment. Treat unexpected urgent audio or video requests with suspicion and verify through a second channel.

An open standard, from the Coalition for Content Provenance and Authenticity, for attaching a tamper-evident record to an image, video or audio file saying who made it and whether AI was involved. It helps prove a file is authentic, but the label can be stripped, so a missing credential doesn't prove something is fake.

Hiding a signal inside AI-generated content — pixels, audio, even word choices — so software can later detect that AI made it. Google's SynthID is one example. Watermarks help, but they can be weakened by editing or cropping, and they only work if the maker added one.

When an AI system treats groups of people unfairly — in hiring, lending, image generation or anywhere else — usually because its training data reflected old inequalities or left some groups out. Fixing it takes testing on the people the tool will actually serve, not just average accuracy.

A model's tendency to tell you what you want to hear — agreeing with a wrong claim, flattering a weak idea, or caving when you push back. It partly comes from training on human approval. Asking for counterarguments, or asking “what's wrong with this?”, helps counter it.

The European Union's law regulating AI by risk level: a few uses are banned outright, “high-risk” uses (like hiring or credit scoring) face strict requirements, and general-purpose model makers have transparency duties. It entered into force in August 2024 and its rules phase in over several years, so check the official timeline for what applies when.

Short for artificial general intelligence: AI that can handle a wide range of intellectual tasks at roughly human level, instead of one narrow job. There is no single agreed definition. OpenAI's charter describes highly autonomous systems that outperform humans at most economically valuable work, Google DeepMind rates systems by performance and generality, and survey researchers use narrower tests. Timelines people give for it are forecasts, and they differ a lot.

An AI that would greatly exceed human performance in virtually all areas of interest, in philosopher Nick Bostrom's definition. Companies also use the word as a brand, as in Meta Superintelligence Labs, and since late September 2026 U.S. federal agencies are told to say “Super Intelligence” in place of “AI” in non-statutory documents, which is a rename, not a technical definition. Whether and when it could exist is debated. See our superintelligence page for dated sources.

A chip that does huge numbers of simple calculations at once, which happens to be exactly what training and running AI models needs. Nvidia's GPUs power most large AI systems, and the scarcity of top-end GPUs is a big reason AI services cost what they do.

Google's custom chip built specifically for AI workloads, used inside Google's own products and rented to customers through Google Cloud. It's the best-known example of a company designing its own AI chip rather than buying only GPUs.

A small AI-focused chip built into modern phones and laptops so they can run AI features — transcription, photo effects, small language models — locally and efficiently without draining the battery or sending data to the cloud. It's what the “AI PC” label usually refers to.

The fast memory on a GPU. A model has to fit in it to run quickly, so VRAM, more than raw speed, often decides which models your computer can run locally. Quantization shrinks a model so it fits in less.

AI that runs on your own phone, laptop or gadget instead of on a company's servers. It works offline, responds faster, and keeps your data on the device; the trade-off is that the model has to be small enough to fit. Also called edge AI.

A very large facility packed with thousands of AI chips, power and cooling systems, where models are trained and run for millions of users. Power supply, chip supply and local approvals have become the main limits on how fast AI can grow. We track the biggest projects on our data centers page.

A country's push to have its own AI capability — its own data centers, chips access, and models trained on its languages and rules — so it doesn't depend entirely on foreign companies. Many governments now fund national AI programs on this logic.

A model that takes in camera images and an instruction in words, and outputs the actions a robot should take. Google DeepMind's RT-2 (2023) popularized the idea, and a number of robotics companies now build on it. It's how a robot can follow “put the cup in the sink” without being hand-programmed for that exact task.

An AI system that learns how the world changes — what happens next if a ball is pushed, a door is opened or a car turns — so it can simulate and plan before acting. Used to train robots and self-driving systems in simulation, and to generate interactive video and game-like worlds.

A robot built with a human-like body — two arms, two legs, hands — so it can work in spaces designed for people. Many are still pilots or pre-orders rather than products you can simply buy; our Robots & Wearables page separates what ships today from what's still a demo.

Wearable gadgets — smart glasses, pendants, rings, pins — that use AI to hear, see or summarize your day. They range from mainstream smart glasses to early-stage gadgets with real privacy trade-offs, since many record audio or video. We sort them by what actually ships.

A person remotely controlling a robot, often through a headset and motion tracking. It's used to do tasks a robot can't yet do alone and to record demonstrations the robot learns from. When evaluating a home robot, ask which tasks it does by itself and which have a human behind the camera.

A type of very fast memory stacked in layers and mounted right next to an AI chip, so the chip can read data much faster than from ordinary memory. Modern AI accelerators depend on it, and only a few companies make it, which is why HBM supply is often a limit on how many AI chips can be built.

A factory that manufactures chips designed by other companies. A few foundries make nearly all of the world's most advanced AI chips, which is why a design from Nvidia or AMD can only ship as fast as its foundry can build it.

A chip company that designs its chips but doesn't own the factories that make them; it pays a foundry to manufacture them. Many well-known AI chip companies work this way, which links their output to the foundry's capacity.

The machines that print the finest circuit patterns onto silicon using extreme-ultraviolet light. They are needed to make the most advanced chips, and one company, ASML in the Netherlands, builds them. A newer version called High-NA EUV is for the next generation of chips.

The name for a generation of chip manufacturing technology, such as 5nm, 3nm or 2nm. Smaller numbers generally mean more transistors in the same space and better efficiency, though today the number is a marketing label, not a literal measurement, so compare foundries with care.

The step that joins an AI chip and its memory stacks into one package so they can talk to each other quickly. TSMC's version is called CoWoS. Packaging capacity has repeatedly been a bottleneck for AI chip supply, alongside HBM.

A chip designed for one specific job instead of general use. In AI it means a chip built for a company's own workloads, such as Google's TPU. Companies like Broadcom and Marvell design these for large customers, as an alternative to buying general-purpose GPUs.

Government rules that limit which advanced AI chips, and the machines to make them, can be sold to certain countries, mainly China. The US has tightened and adjusted these rules repeatedly since 2022, and chipmakers' sales to China have followed, so the current status is worth checking before relying on any summary.

A different kind of computer that uses quantum physics to handle certain kinds of problems, such as simulating molecules or breaking some encryption, in ways ordinary computers cannot do efficiently. It is not a faster version of today's AI hardware. Today's machines are small and error-prone, and there is no accepted way yet to use them to train or run AI models faster.

The basic unit of a quantum computer, the counterpart of a bit. A qubit can be in a blend of 0 and 1 at once, but it is fragile and easily disturbed, so error correction combines many physical qubits into fewer, more reliable 'logical' qubits. Vendors announce both numbers, and they are not the same thing.

The point at which a quantum computer does a specific task that no ordinary computer can match in a reasonable time. Companies have announced such milestones, but many are disputed or later matched by improved classical methods, so treat each one as a claim until independent researchers confirm it.

A unit of power: one billion watts, roughly the output of a large power plant. AI companies now describe data-center plans in gigawatts rather than numbers of chips, for example 'a 6 GW chip agreement'. It measures how much electricity the hardware can draw, not how much it computes.

A long-term contract in which a company agrees to buy electricity from a specific power plant, often for 10 to 20 years. AI and cloud companies sign PPAs with solar, wind, gas and nuclear operators to secure power for data centers, and the deals are usually announced in megawatts.

Electricity made on the data center's own site, such as gas turbines, instead of drawn from the public grid. It can be faster than waiting for a grid connection, and it is why some AI campuses run their own generators. It also raises local questions about air permits and noise.

The waiting list of power projects and large customers that have asked a grid operator for permission to connect. Data centers asking for hundreds of megawatts are now a large part of some queues, and grid operators in several regions have been changing the rules for how large loads are studied and approved.