AI, translated to human

Software that performs tasks that once required human intelligence: understanding language, recognizing images, making decisions. Modern AI learns patterns from massive amounts of data instead of following hand-written rules.

A way of building AI where computers learn from examples instead of explicit instructions. Show a system thousands of spam emails and it learns to spot spam on its own — scale that idea up and you get today's AI.

Machine learning using neural networks with many stacked layers, each learning progressively more abstract patterns. The approach behind image recognition, voice assistants, and large language models.

A computing structure loosely inspired by the brain: layers of simple units passing signals to each other. During training, the connections between units are adjusted millions of times until the network produces useful outputs.

The trained AI itself — a file of learned numbers that turns inputs into outputs. ChatGPT is an app; the model is the engine inside it. When people compare AI systems, they're usually comparing models.

The process of building a model by feeding it huge amounts of data and gradually adjusting it until its outputs improve. Training is expensive and happens once; using the finished model is called inference.

A model doing its actual job: taking your input and producing an answer. The computing cost of serving millions of answers is why AI companies obsess over making inference cheaper and faster.

A step-by-step set of rules a computer follows to solve a problem. In AI discussions the word is often used loosely to mean "the system making decisions," as in "the recommendation algorithm."

Systematic unfairness in what an AI produces, usually inherited from its training data. A hiring tool trained on historical hiring data can repeat historical discrimination — one of the most important practical problems in deployed AI.

AI that creates new content — text, images, music, video, code — instead of just analyzing existing data. The technology behind ChatGPT, Midjourney, and most AI tools launched since 2022.

Machine learning from labeled examples — the AI is shown inputs paired with correct answers, like photos tagged "cat" or "dog," until it learns to label new ones. The most common way models are trained for classification and prediction tasks.

Machine learning without labeled answers — the AI finds structure in raw data on its own, like grouping customers by behavior or topics in a pile of documents. Used for clustering, anomaly detection, and discovering hidden patterns.

Training an AI through trial and error with rewards and penalties, like training a pet — the system tries actions, gets scored, and gradually learns the strategy that earns the most reward. Behind game-playing AIs and parts of chatbot alignment.

When a model memorizes its training data instead of learning general patterns — perfect scores on examples it has seen, poor performance on anything new. The classic sign of a model that learned the answers rather than the subject.

The collection of data a model learns from — text, images, audio, or other examples. A model's knowledge, abilities, and biases all trace back to its dataset, which is why "what was it trained on" is one of the most important questions in AI.

The field of AI focused on understanding images and video — recognizing objects, faces, scenes, and actions. Powers everything from phone face unlock to medical scan analysis to self-driving car perception.

The field of AI focused on computers understanding and producing human language — translation, sentiment analysis, chatbots, summarization. Large language models absorbed much of this field, but the term still covers the underlying techniques.

The human work of tagging raw data with correct answers — marking spam emails, outlining objects in photos, rating responses — so models have examples to learn from. The invisible labor behind nearly every AI system.

An AI that doesn't just answer questions but takes actions toward a goal: browsing the web, writing files, or using software on your behalf. The difference between asking a friend for directions and asking them to drive you there.

AI that plans, uses tools, and completes multi-step work with minimal supervision — the industry's biggest shift since chatbots. If a chatbot answers "how do I get a refund," agentic AI goes and gets the refund.

The structured way a model uses tools: instead of writing text, it outputs "call this function with these inputs" — check the weather, query a database, send an email — and software executes it. Also called tool calling; the plumbing under most agent behavior.

An open standard letting AI assistants plug into outside tools and data — your files, calendar, company database — the way USB lets devices plug into a computer. When an app "supports MCP," assistants can actually use it, not just talk about it.

Using AI to run multi-step tasks end-to-end without a person driving: triaging inboxes, processing invoices, publishing content. Where classic automation followed rigid rules, AI automation handles the messy middle that needs judgment.

Keeping a person inside an automated process to review, approve, or correct the AI's work at key checkpoints. The standard answer to deploying AI safely for anything that matters.

Several AI agents working together with different roles — researcher, writer, critic, coordinator — passing work between them. Often outperforms one agent doing everything, for the same reason teams beat solo generalists.

A model's ability to operate external tools — search, calculators, code execution, APIs — during a task. Tool use is what turns a chatbot that only talks into an agent that can act.

Rules and checks wrapped around an AI system constraining what it can do: content filters, spending caps, forbidden actions, approval requirements. Guardrails live in the surrounding software, not inside the model.

An AI assistant embedded directly inside a workflow — the code editor, the spreadsheet, the inbox — suggesting and drafting while you stay in control. Also Microsoft's brand name for exactly that.

An AI agent that runs on its own toward a goal without step-by-step human direction — researching a topic, writing code, or managing tasks while you do other things. The "set it and forget it" end of the agent spectrum.

Breaking a big goal into smaller subtasks an AI can handle one at a time — research, then outline, then draft, then edit. The core planning skill that lets agents tackle complex work instead of choking on it whole.

An agent reviewing its own work for errors and improving it before delivering — drafting, critiquing, and revising in a loop. One of the simplest ways to raise agent reliability without a bigger model.

The step where an agent figures out its approach before acting: what tools it needs, what order to do things in, and what could go wrong. Good planning separates agents that complete goals from ones that wander.

In a multi-agent system, the coordinator that assigns work, tracks progress, and merges results from specialist agents. The manager layer that keeps a team of AIs pulling in the same direction.

An AI agent that operates a real web browser — clicking, scrolling, filling forms — to complete tasks on websites the way a person would. The technology behind automated research, booking, and web testing agents.

An agent's ability to retain information across tasks and sessions — what it learned, what worked, your preferences — instead of starting from zero every time. Distinct from the context window, which only covers the current conversation.

A model invoking external tools during its work — search, calculators, code execution, APIs — as structured function calls rather than plain text. The capability that turns a chatbot that only talks into an agent that can act.

Many agents deployed at once on a big job — dozens researching, testing, or migrating code in parallel while a supervisor coordinates. The industrial-scale version of multi-agent systems.

The scaffolding around a model that turns it from something that talks into something that does: feeding it tasks, handing it tools, running what it asks for, looping until done. When people say "AI agent," the harness is doing the orchestrating.

An agent running repeatedly on its own — checking something, doing work, reporting back — on a schedule or until a goal is met, instead of waiting for a prompt each time. "Set it and forget it" for AI.

Add-ons giving an AI app extra powers — connecting it to email, bookings, design tools. Same idea as browser extensions: the core stays the same, plugins extend its reach.

A sealed environment where AI-generated code or actions run safely without touching anything real. Why an AI can execute code for you without an error — or an attack — wrecking your machine.

AI suggesting the next lines of code as you type, like autocomplete for programming. The gateway feature of AI coding tools, trained on billions of lines of public code.

An AI that codes alongside you: suggesting, explaining, and debugging in real time inside your editor. The name nods to the human practice of pair programming, with the AI as your partner.

Automated review of code changes for bugs, security issues, and style problems before a human looks. Catches the obvious so reviewers can focus on design and logic.

Plain-language description of program logic, written as a stepping stone to real code. AIs are good at turning pseudocode into working programs and vice versa.

An AI that helps find and fix bugs: reading error messages, tracing likely causes, suggesting fixes. Often faster than manual debugging for common errors.

Converting code from one programming language to another — Python to JavaScript, for example. Useful for migrations, though the output always needs testing.

AI writing test cases for your code automatically: inputs, expected outputs, edge cases. Expands test coverage with far less manual effort.

Repetitive starter code every project needs — setup, configuration, standard patterns. AI generates it instantly, which is a big chunk of the productivity gain from coding assistants.

Software Development Kit: a package of tools, libraries, and docs for building on a platform. AI companies ship SDKs so developers can integrate models with a few lines of code.

A library that simplifies calling an API, hiding raw HTTP requests behind friendly functions. Most developers use wrappers rather than hand-crafting API calls.

Building working software by describing it in plain language and letting AI generate the code. Early-stage but improving fast — the plain-language dream of app building made concrete.

Building AI-powered apps without writing code, using visual builders and templates. Lets non-programmers ship chatbots, automations, and tools.

An AI coding tool reading and mapping your entire project so its suggestions understand your code's structure, not just the open file. The difference between generic help and genuinely useful help.

AI editing code directly where your cursor is — rewriting a function, applying a fix — instead of just suggesting text in a chat panel. The interaction model of modern coding assistants.

AI is the broad goal: machines doing intelligent tasks. Machine learning is the main method used to achieve it — learning from data rather than hand-written rules. All ML is AI, but not all AI is ML.

Machine learning is the field of learning from data; deep learning is its most powerful subfield, using multi-layered neural networks. Deep learning handles images, language, and other complex data where classic ML struggles.

Training builds the model (expensive, done once); inference runs it (cheap per use, done constantly). Confusing the two leads to wildly wrong cost estimates.

Parameters are the model's learned internal settings (its size); tokens are the chunks of text it processes (its usage). You pay for tokens; capability roughly scales with parameters.

A chatbot answers questions in conversation; an agent takes actions to complete goals. The chatbot tells you how to get a refund; the agent gets it.

GPT is the model family (the engine); ChatGPT is the app built on it (the car). People compare models when they mean GPT, and products when they mean ChatGPT.

Open weights means you can download and run the model; true open source also means open training data and code. Most "open" models are open-weights, not open-source.

Fine-tuning bakes knowledge into the model through extra training; RAG looks knowledge up at answer time. Fine-tuning changes behavior; RAG grounds answers in your documents. They're often combined.

Both control randomness in AI output. Temperature scales how adventurous word choices are; top-p (nucleus sampling) limits choices to the most probable words. Most users only ever need temperature.

Supervised learns from labeled examples (photos tagged cat/dog); unsupervised finds structure in unlabeled data (grouping customers by behavior). Labels are the difference.

Generative AI creates new content (text, images); discriminative AI classifies and predicts (spam or not, price forecasts). The last few years' revolution is almost entirely generative.

Automation follows fixed rules to repeat tasks; AI handles tasks needing judgment, language, or adaptation. A script that sends invoices is automation; a system that reads and triages them is AI.

Your prompt is what you type; the system prompt is the hidden instruction the app gave the AI before your conversation. Yours asks the question; it sets the rules of engagement.

Embeddings turn text into searchable meaning-vectors without changing the model; fine-tuning changes the model itself through training. Retrieval uses embeddings; behavior change needs fine-tuning.

Searching by meaning rather than exact words. "Affordable ways to stay warm" can match "budget heating options" because embeddings capture intent, not just keywords.

Traditional search matching exact words and phrases. Fast and precise when you know the right terms, but it misses paraphrases and synonyms that semantic search catches.

Combining keyword and semantic search to get the best of both: exact matches when terms matter, meaning matches when they don't. The standard approach in modern RAG systems.

Splitting documents into small pieces before storing them for retrieval. Good chunks are the right size — big enough to carry meaning, small enough to be precisely relevant.

A curated collection of documents an AI can draw on when answering — help articles, manuals, company wikis. The "knowledge" in retrieval-augmented generation.

Producing answers tied to retrieved sources rather than pure memory. Grounded responses can cite where claims came from, making them more trustworthy and verifiable.

A reference to the source behind an AI's claim. The best AI search and RAG tools cite sources inline, letting you check facts instead of taking the model's word.

The automated series of steps that moves data from raw sources to usable form: collect, clean, transform, load. Every serious AI product runs on pipelines, not manual spreadsheets.

Automatically extracting information from websites. A common way to build datasets and feed RAG systems — and a legal gray area depending on terms of service and copyright.

Information organized in a fixed format: tables, spreadsheets, databases with defined fields. Easy for machines to query precisely, unlike free-form text.

Information without a fixed format: raw text, images, audio, video. Most of the world's data is unstructured, and AI's great strength is making sense of it.

A second, more careful relevance pass over retrieved documents before showing them to the model. Retrieval casts a wide net; reranking picks the best catches.

The date up to which a model was trained — anything after that, it doesn't know from training. RAG and web browsing exist largely to work around cutoffs.

Extracting clean text and structure from files like PDFs, slides, and scanned pages so AI systems can actually use them. The unglamorous first step of most document AI.

Graphics Processing Unit: a chip designed for massive parallel math, originally for rendering games. Turned out to be perfect for AI training, making GPUs the most important hardware in the AI boom.

Tensor Processing Unit: Google's custom chip built specifically for AI workloads. Powers much of Google's AI research and cloud offerings as an alternative to NVIDIA GPUs.

NVIDIA's platform for running general-purpose computing on its GPUs. A huge part of NVIDIA's moat: much of the AI software ecosystem is built on it.

Video RAM: the fast memory on a GPU where the model lives while running. VRAM size is the practical limit on which models you can run locally — bigger models need more.

Software and hardware set up to serve a model to users at scale: handling requests, batching, and keeping latency low. Running a model is easy; serving millions of requests is engineering.

Renting GPU power by the hour from a cloud provider instead of buying hardware. How most startups train and serve models without a data center.

Running AI directly on devices — phones, cameras, sensors — instead of in the cloud. Faster, more private, works offline; limited by the device's modest computing power.

The delay between a request and a response. In AI apps, latency is the difference between a tool that feels instant and one that feels broken.

How much work a system completes per unit of time — tokens per second for a model server, requests per second for an API. The capacity number that decides whether your app survives launch day.

Floating-point operations per second: the standard measure of computing speed for AI hardware. Model training is discussed in petaflops; inference in teraflops.

Many GPUs networked together to act as one giant machine for training large models. Frontier models train on clusters of tens of thousands of GPUs.

Application-Specific Integrated Circuit: a chip designed for one job, like Google's TPUs for AI. More efficient than general chips at that job, useless at everything else.

The engineering of putting a trained model into production: packaging, scaling, monitoring, updating. Training gets the headlines; serving is where products live or die.

The delay when an idle AI service wakes up to handle its first request — loading the model into memory. Noticeable in serverless setups; providers work to minimize it.

The technique behind most AI image and video generators. The model learns by watching images dissolve into noise, then reverses the process: starting from pure static and refining step by step into a picture matching your prompt.

Generating pictures from a written description. Type "a cozy cabin in the snow at dusk" and get an image — the capability behind Midjourney, DALL-E, and Stable Diffusion.

A synthetic on-screen presenter: a digital human that speaks your script with lip-synced video. Used for training videos, localized content, and faceless channels without cameras or crews.

Recreating someone's voice with AI from a short audio sample, then making it say anything. Powers dubbing, audiobooks, and voiceovers — and raises real consent and fraud concerns.

Using AI to enlarge and sharpen an image or video beyond its original resolution, reconstructing detail that wasn't there. Common for restoring old photos and improving AI-generated art.

Editing part of an image by having AI regenerate just the selected region: remove an object, change a background, or fix a flaw while the rest of the picture stays untouched.

Adjusting a speaker's mouth movements in video to match new audio — most often used with AI dubbing so translated speech looks natural instead of badly dubbed.

Music composed or produced with AI: full songs from a text prompt, backing tracks, or stems separated from existing recordings. Tools in this space range from fun toys to serious production assistants.

Low-Rank Adaptation — a lightweight way to fine-tune an image model on a specific style, character, or concept using just a handful of images. The reason you can download tiny files that teach Stable Diffusion to draw in any artist's style.

A saved version of a trained image model, shared as a downloadable file. The AI art community trades thousands of checkpoints, each with its own style — photorealistic, anime, cinematic — built on the same base architecture.

Generating video clips from a written description — "a drone shot over a coastal town at sunset." One of the fastest-moving frontiers in AI media, going from gimmicky to genuinely useful within a couple of years.

Converting written text into spoken audio with AI voices — the technology behind voiceovers, audiobooks, and virtual assistants. Modern TTS is nearly indistinguishable from human narration.

Transcribing spoken audio into written text — meeting notes, subtitles, voice commands. The counterpart to text-to-speech and the foundation of every voice assistant.

AI-generated or altered video and audio that convincingly impersonates a real person. Enables creative uses like dubbing and de-aging actors, but also fraud and misinformation — which is why detection and watermarking matter.

Extending an image beyond its original borders by having AI imagine what surrounds it — turning a portrait into a full scene. The reverse of cropping, and a staple trick in AI image editing.

Applying the visual style of one image to the content of another — your photo reimagined as an oil painting or anime still. One of the earliest popular AI art techniques, now a standard feature in creative tools.

A technique guiding image generation with structural inputs — a pose skeleton, a depth map, an edge sketch — so you control composition precisely instead of gambling on the prompt. A pillar of professional AI image workflows.

Generative Adversarial Network: two neural networks trained against each other, one generating images and the other judging them. Powered the first wave of AI image breakthroughs before diffusion models took over.

Convolutional Neural Network: the classic architecture for image recognition, scanning pictures in small patches to learn edges, shapes, then objects. Drove the 2012 deep-learning breakthrough; still everywhere in computer vision.

Keeping the same character recognizable across many generated images or video shots. Historically one of the hardest problems in AI art, and a headline feature whenever a model improves at it.

A visual glitch in generated content — warped text, melted edges, six fingers. Artifact-spotting is still the fastest way humans clock an image as AI-made.

AI generating in-between frames to make video smoother or slow motion from regular footage. Turns 24fps into buttery 60fps by imagining what happened between frames.

Re-recording video dialogue in another language — now AI-powered with voice cloning and lip-sync, collapsing the cost of releasing content globally.

A number anchoring the randomness in generation: same prompt plus same seed gives nearly the same output. Essential for reproducing an image you liked or debugging behavior.

Classifier-Free Guidance scale: how strongly an image model should follow your prompt versus improvise. Low values wander creatively; high values obey literally, sometimes at the cost of quality.

The model's internal "map of meaning," where similar concepts sit near each other. Image generators navigate this space to turn prompts into pictures; small moves morph one concept toward another.

Neural Radiance Fields: a technique that builds 3D scenes from a set of 2D photos, letting you render new viewpoints. The bridge between photography and 3D.

An AI model trained on enormous amounts of text until it becomes remarkably good at predicting what comes next. That single skill unlocks writing, summarizing, translating, coding, and reasoning. ChatGPT, Claude, and Gemini are all powered by LLMs.

Generative Pre-trained Transformer — the model family behind ChatGPT. Generative (creates new text), Pre-trained (learned language from huge datasets first), Transformer (the architecture that made it work). The name became shorthand for this entire class of AI.

The neural network architecture behind nearly every modern AI model. Its key idea, "attention," lets a model weigh how every word in an input relates to every other — a design that scaled spectacularly with more data and computing power.

The internal settings a model adjusts during training — billions of numbers encoding everything it "knows." Parameter counts like 7B or 70B roughly indicate a model's size: bigger usually means more capable, but also more expensive to run.

The small chunks of text a model actually reads and writes, usually pieces of words a few characters long. Tokens are also the unit AI usage is billed in — roughly 750 words per 1,000 tokens.

How much text a model can consider at once: your conversation, pasted documents, and instructions combined, measured in tokens. When a conversation exceeds the window, the earliest parts drop out — which is why long chats sometimes make an AI forget what you said at the start.

The input you give an AI to get it to do something: a question, an instruction, or a half-written paragraph. Clear, specific prompts with context and examples consistently produce better results than vague ones.

When an AI confidently states something false or invented — fake citations, wrong facts, imaginary events. It happens because models predict plausible-sounding text rather than checking facts, so important claims should always be verified.

Taking a pre-trained model and giving it extra training on a smaller, specialized dataset to adapt it for a specific job — medical Q&A, legal drafting, or a company's support tone. Cheaper than training from scratch, more tailored than prompting alone.

A technique where the AI looks up relevant documents before answering, then grounds its response in what it found. RAG reduces hallucinations and lets models answer from your own files and up-to-date sources instead of only their training data.

An AI that handles more than one type of input and output — text, images, audio, video — in a single model. Show it a photo and ask questions about it, or talk to it out loud: most flagship models are now multimodal.

A model whose trained files are publicly downloadable, so anyone can run, study, or modify it. Different from fully "open source" — training data and code may stay private — but it lets developers build without depending on a company's API.

A way of turning words, images, or documents into lists of numbers that capture their meaning, so a computer can measure similarity mathematically. Embeddings are how AI search finds "similar" things and how RAG retrieves relevant documents.

A database designed to store embeddings and find the closest matches at speed — the memory layer behind RAG systems and semantic search. When an AI "looks up" your documents, a vector database is usually doing the finding.

The first, massive phase of building a language model: learning general language skills from huge amounts of text before any specific tuning. Pre-training is what makes a model broadly capable; everything after adapts that capability.

Reinforcement Learning from Human Feedback — training a model to be helpful and safe by having humans rank its answers and rewarding the better ones. A key reason chatbots feel conversational and aligned rather than just autocomplete engines.

Shrinking a model by storing its numbers with less precision — like compressing a photo. Done well, it cuts memory and cost dramatically with barely any quality loss, which is how big models end up running on phones and laptops.

A standardized test used to compare AI models — math problems, coding challenges, reasoning puzzles — producing the scores you see in every model announcement. Useful but gameable: models can be tuned to ace benchmarks while feeling no better in daily use.

The mechanism that lets a model weigh which parts of the input matter for each part of the output — when generating the word after "the capital of France is," attention focuses on "France." The breakthrough that made transformers, and modern AI, work.

Application Programming Interface — a set of rules letting different software talk to each other. An AI model's API lets developers send prompts and receive answers programmatically, which is how thousands of apps are built on top of models like GPT and Claude.

Training a small, fast model to imitate a large, powerful one — transferring the big model's skills into a cheaper package. How "mini" and "lite" model versions deliver surprisingly good performance at a fraction of the cost.

The effort to make AI systems behave in line with human values and intentions — helpful, honest, and safe rather than merely capable. Techniques like RLHF and system prompts are alignment tools; the broader challenge is one of AI's deepest research problems.

An architecture where a model contains many specialist sub-networks but activates only a few per input — like a hospital routing you to the right department instead of consulting every doctor. Big-model capability at a fraction of the running cost.

A model that understands images and text together — show it a chart, a screenshot, or a photo of your fridge and ask questions. The capability behind "upload an image" in modern chatbots.

A model built to think before answering, spending extra computation working through problems step by step. Dramatically better at math, code, and planning; slower and pricier per answer.

A deliberately compact language model — cheaper, faster, small enough for a phone or laptop. Less capable than giants, but for focused jobs often all you need.

A model fresh out of pre-training, before being taught to be a helpful assistant. It completes text but doesn't follow instructions well — chat models are base models plus instruction tuning.

A model at the leading edge of capability — the best anyone has shipped. "Frontier lab" means the handful of companies building them.

A language model specifically tuned to follow instructions, as opposed to a raw base model. The practical category covering nearly every chatbot and assistant product.

A model whose job is turning text into embeddings rather than generating answers. The quiet workhorse behind semantic search, recommendations, and RAG retrieval.

A model with an unusually large context window — able to consider whole books or codebases at once. Unlocks document analysis and long conversations that smaller windows can't handle.

A model that activates only part of itself for each input, like mixture of experts. The efficiency direction of model design: capability without paying full compute on every token.

The practice of writing inputs that get the best results from AI: being specific, giving examples, assigning a role, breaking tasks into steps. Less about magic words, more about clear communication — like writing a good brief for a freelancer.

Asking for something with no examples: "Translate this to French." Modern models handle most everyday tasks zero-shot; add examples when the format or style really matters.

Including a couple of examples inside your prompt so the model copies the pattern — two or three sample inputs with the outputs you want. One of the most reliable ways to improve results without any technical work.

Getting the AI to work through a problem step by step — “first this, then that, therefore…” — instead of jumping to an answer. Showing its work measurably improves accuracy on math, logic, and planning.

Hidden instructions an AI receives before your conversation starts: its role, personality, and rules. When two apps built on the same model behave completely differently, the system prompt is usually why.

A setting controlling how predictable or adventurous an AI's answers are. Low temperature gives consistent, factual output; high temperature gives varied, creative output with more risk of errors.

Forcing a model to answer in an exact machine-readable format, usually JSON, instead of free prose — so software can reliably use the answer. The unglamorous feature that makes AI dependable inside real products.

Designing everything the model sees when it works — instructions, documents, tools, memory, examples — not just the wording of a single prompt. As AI systems grew more complex, managing context became the real craft.

A reusable prompt with blanks to fill in, like "Summarize [ARTICLE] for [AUDIENCE] in [LENGTH]." How teams and apps keep AI output consistent instead of reinventing the request every time.

A general-purpose AI helper for answering questions, drafting text, summarizing, and planning. The word increasingly implies it can also take actions for you — which is where assistants start becoming agents.

Starting a prompt by assigning the AI a role — "You are a senior copywriter…" or "Act as a Python expert…" — to steer its tone, knowledge, and behavior. A simple trick that reliably improves specialized answers.

Tricking an AI into ignoring its instructions by hiding malicious commands in content it reads — a webpage, document, or image telling it to "disregard previous instructions." One of the most important security concerns for AI apps that process untrusted content.

A prompt crafted to bypass an AI's safety rules and get it to do something it was designed to refuse. An ongoing cat-and-mouse game between attackers finding loopholes and developers closing them.

Giving the model a single example of what you want before asking — halfway between zero-shot (no examples) and few-shot (several). Often enough to lock in a format without a long prompt.

In image generation, telling the AI what to avoid — "no blurry, no extra fingers, no watermark" — alongside what you want. As important as the main prompt for getting clean, professional results.

A prompting pattern where the AI alternates between reasoning ("I should check the price first") and acting (calling a tool), in a visible loop. The foundation of many AI agents: think, do, observe, repeat.

Improving answers by having the AI solve the same problem several times independently and going with the most common result. Costs more computation but noticeably raises accuracy on reasoning tasks.

Markers like triple quotes or XML tags that separate different parts of a prompt — instructions vs. examples vs. user content. Delimiters reduce confusion and are a basic defense against prompt injection.

Feeding one AI response into the next prompt as an assembly line: outline first, then draft each section, then edit. Usually beats asking for everything in one giant prompt.

The output length limit — the most a model may generate in one response. When an answer cuts off mid-sentence, this ceiling is usually why.

AI output appearing word by word as it's generated instead of arriving all at once. The answer takes the same total time, but watching it type feels faster.

The documented tendency of models to focus on the beginning and end of a long context while glossing over the middle. Practical tip: put key instructions first or last.

A robot with a human-like body: head, torso, arms, legs. The bet is that human-shaped robots can work in human-designed spaces — warehouses, homes — without rebuilding the world for them.

AI that exists in a physical body and learns by interacting with the real world, not just data. Robots that learn by doing represent AI's next great leap beyond screens.

Simultaneous Localization and Mapping: how robots figure out where they are while building a map of unfamiliar surroundings. The core trick behind robot navigation, from warehouse bots to Mars rovers.

The component that makes a robot move — motors, hydraulics, artificial muscles. AI decides what to do; actuators do it.

Combining data from multiple sensors — cameras, lidar, touch — into one coherent picture of the world. No single sensor is reliable enough alone; together they're trustworthy.

Controlling a robot remotely, with a human in the loop. Used for dangerous environments and for collecting the demonstration data that trains autonomous robots.

Training robots in simulation, then transferring the skills to physical hardware. Simulation is cheap and safe; the challenge is the "reality gap" where simulated physics differs from the real world.

Robots improving their skills through experience and data rather than hand-programmed routines. The application of machine learning — especially reinforcement learning — to physical machines.

The field devoted to making sure AI systems behave safely and as intended — from preventing chatbots from giving dangerous advice to the long-term challenge of keeping highly capable systems under human control.

Deliberately attacking an AI system to find its weaknesses before real attackers do — probing for jailbreaks, injection vulnerabilities, and harmful outputs. Standard practice before any major model release.

The policies, laws, and organizational rules controlling how AI is built and used — from company review boards to national regulations. As AI spreads into hiring, healthcare, and finance, governance decides who is accountable when things go wrong.

How openly an AI system's workings are disclosed — what data it trained on, how it makes decisions, when you're talking to a machine. A core demand of regulators and a growing expectation of users.

The practice of building and deploying AI with fairness, safety, and accountability designed in from the start — testing for bias, documenting limitations, and keeping humans in charge of consequential decisions.

The unresolved legal question of whether AI companies may train on copyrighted text, images, and music without permission. Lawsuits from authors, artists, and publishers are actively shaping the answer, with major implications for the industry.

The degradation that happens when AI models are trained on increasing amounts of AI-generated content instead of human-created data — each generation a slightly worse copy of the last. A real concern as synthetic content floods the web.

Embedding invisible markers in AI-generated content so it can later be identified as machine-made. A promising tool against deepfakes and misinformation — though today's watermarks can still be removed or evaded.

The discipline of ensuring AI is developed and used fairly, safely, and accountably — covering bias, transparency, and who bears responsibility when systems cause harm.

Whether people agreed to have their data, voice, or likeness used to build or power AI systems. From training-data lawsuits to voice cloning, consent is the fault line of the AI ethics debate.

Tools and techniques for identifying AI-manipulated media. An arms race: generators improve, detectors catch up, and the stakes are elections and fraud.

Using AI to review user content at scale — flagging hate speech, spam, and explicit material. Fast and tireless, but still needs human judgment for edge cases.

The principle that organizations must answer for what their AI systems do — explain decisions, fix harms, face consequences. The policy backbone of AI regulation.

Protecting personal information in AI systems: what's collected, how it's used, whether it leaks through models. Privacy regulations now collide directly with how models train.

Content fully or partly generated by AI — images, video, audio, text. The umbrella term covering everything from helpful voiceovers to malicious deepfakes.

Who pays when AI causes harm: the developer, the deployer, or the user. Largely unsettled law, and one of the biggest brakes on AI adoption in high-stakes fields.

A controlled program letting companies test AI products under regulator supervision with relaxed rules. How governments try to encourage innovation without abandoning oversight.

Independent examination of an AI system for bias, safety, and compliance — like a financial audit for algorithms. Increasingly required by regulation for high-risk AI.

A mechanism to shut down an AI system quickly if it misbehaves. Simple in theory; for deeply integrated systems, genuinely hard to build — and a staple of AI safety debates.

Low-effort, mass-produced AI content flooding the internet: generic articles, fake images, spammy videos. The term of contempt for AI output nobody asked for and nobody wants.

Building software by describing what you want in plain language and letting AI write the code, guiding by feel rather than technical specs. Fun, fast, and genuinely useful for prototypes — risky for anything critical.

Slang for someone convinced AI will lead to catastrophe. Originally an insult, partly reclaimed: AI safety researchers sometimes wear it as a badge.

The belief that technological progress, especially AI, should be pushed as fast as possible — associated with the e/acc movement. The philosophical opposite of the doomers.

Over-the-top flattery or hype, often describing how AI assistants agree with everything you say. Sycophancy as a service, and a running joke about chatbot personalities.

Someone with an almost intuitive knack for getting great results from AI through well-crafted prompts. Half skill, half mystique — and increasingly a real job description.

The idea that much of the web is now bots talking to bots, with humans a minority audience. AI-generated content at scale made a conspiracy theory feel uncomfortably plausible.

Internet slang for a robot or AI, often used dismissively. Part of the cultural backlash vocabulary as AI systems become more visible in daily life.

The meme depicting AI as a monstrous tentacled creature wearing a smiley-face mask — raw, alien capability barely contained by a friendly chatbot interface. The AI community's favorite self-portrait.

A thought experiment about an AI told to make paperclips that converts the entire world into paperclips. The classic illustration of why "just follow the goal" is dangerous without aligned values.

A young company building its product around artificial intelligence. The current boom has produced thousands, from model labs to thin wrappers — sorting signal from noise is the market's ongoing work.

A company that trains and offers large general-purpose models — OpenAI, Anthropic, Google, Meta. Everyone else in the AI economy builds on, competes with, or distributes their work.

AI built for one specific industry or job — legal document review, medical imaging, construction estimating — rather than general use. Where much of AI's real business value is being captured.

A product that's essentially a nice interface on someone else's model, with little proprietary technology. Easy to build, hard to defend — the most crowded and fragile corner of the AI market.

AI deployed inside large organizations: copilots for employees, automation for operations, governed and secured. The biggest budgets in AI software are here.

How companies evaluate and buy AI tools: security reviews, data handling, accuracy testing, cost modeling. The unglamorous gatekeeper deciding which AI products actually get adopted.

Charging by how much of the AI you consume — per token, per minute, per task. The natural pricing for AI, and a shock for buyers used to flat subscriptions.

Charging per user, the classic SaaS model applied to AI tools. Simple to understand; increasingly strained as AI agents do the work of many seats.

Free basic tier with paid upgrades — the dominant go-to-market for AI consumer apps. The free tier buys distribution; the trick is converting enough users before inference bills pile up.

What model providers charge developers per unit of use, usually per million tokens, differing for input and output. The meter running underneath thousands of AI products.

A durable competitive advantage in AI: proprietary data, distribution, workflow lock-in, or cost structure. Hotly debated, because model capabilities alone have proven surprisingly easy to copy.

The industry's defining strategic split: freely downloadable weights anyone can run versus proprietary models behind an API. Each side claims the future; the market uses both.

Firms that help companies adopt AI: strategy, implementation, training. Booming, because most organizations know they need AI and don't know where to start.

The perennial decision, sharpened by AI: build custom models and tools in-house, or buy off-the-shelf products. The right answer depends on how core the capability is to your business.

Return on investment from AI spending: the metric every enterprise pilot is now judged on. After the experimentation phase, "did it pay?" decides what survives.

Employees using unauthorized AI tools at work — pasting company data into public chatbots. The security headache of the AI era, driving demand for governed alternatives.

Artificial General Intelligence: AI that can do any intellectual task a human can, not just narrow specialties. Nobody agrees how far away it is — estimates range from years to decades — but it's the north star of the leading labs.

AI that surpasses the smartest humans at every intellectual task. Hypothetical for now; the subject of both utopian hopes and existential-safety research.

The hypothesized point where AI improvement becomes self-sustaining and progress explodes beyond human comprehension or control. More philosophy than forecast, but it shapes how seriously people take AI safety.

The observed pattern that bigger models, more data, and more compute reliably produce better AI. The empirical bet behind the largest training runs — and the industry's guiding faith.

Improving AI by throwing more computing power at training and inference. The strategy that delivered the last decade of progress; the open question is how far it goes.

The concern that AI is running out of high-quality human-created training data. If true, future progress needs synthetic data, new architectures, or far better data efficiency.

AI powerful enough to reshape the economy and society the way electricity or the internet did — regardless of whether it's "general" intelligence. The scenario most forecasters actually plan around.

The practice of predicting AI capabilities and timelines — expert surveys, trend extrapolation, prediction markets. Unreliable but unavoidable: everyone building on AI is implicitly betting on a forecast.

Abilities a model already has that nobody has discovered or elicited yet. Each jump in prompting or tooling reveals the overhang was larger than assumed.

The overall advance of AI capabilities over time, measured in benchmarks, economic impact, and real-world deployment. Faster than most experts predicted a decade ago.

Research on keeping highly capable future AI systems aligned with human interests — the technical agenda behind much of the AI safety field, from interpretability to alignment theory.

Technology reading signals directly from the brain, or writing to it — the famous examples being companies building neural implants. Still early, but a candidate for how humans and AI eventually merge rather than compete.

One complete pass through the entire training dataset. Models usually train for many epochs, seeing the same data repeatedly until performance stops improving.

A small chunk of training data processed together in one step. Instead of updating the model after every single example, training uses batches — faster and more stable.

How big a step the model takes when adjusting itself during training. Too high and learning becomes unstable; too low and training crawls.

The score a model tries to minimize during training: a mathematical measure of how wrong its predictions are. Lower loss means better predictions; the whole training process is one long effort to push this number down.

The optimization algorithm behind most AI training: repeatedly nudging the model's settings in the direction that reduces its error. Imagine finding the bottom of a valley in fog by always stepping downhill.

The technique that figures out how much each part of a neural network contributed to an error, so training knows what to adjust. The workhorse calculation that makes deep learning possible.

Techniques that discourage a model from memorizing training data, keeping it simpler so it generalizes better. A model with good regularization performs well on new data, not just data it has seen.

A regularization trick where random parts of a neural network are temporarily switched off during training. It forces the network to not rely on any single part, making it more robust.

Reusing a model trained on one task as the starting point for another. Instead of training from scratch, you take a model that already understands language or images and adapt it — faster and cheaper.

Extra training that teaches a model to follow instructions and respond helpfully, using examples of good instruction-response pairs. The reason modern chatbots answer questions instead of just continuing text.

Direct Preference Optimization: a simpler alternative to RLHF for aligning models with human preferences, learning directly from pairs of preferred and rejected answers without a separate reward model.

Parameter-Efficient Fine-Tuning: methods that adapt a large model by training only a tiny fraction of its parameters. LoRA is the most famous PEFT technique — big-model customization on a small budget.

When learning something new makes a model forget what it knew before. A major challenge in training AI on sequential tasks, and part of why models are usually trained on everything at once.

Training a model the way you'd teach a student: easy examples first, harder ones later. Can speed up learning and improve final performance on difficult tasks.

Artificially expanding training data by creating modified copies — rotating images, paraphrasing text, adding noise. More varied examples make models more robust without collecting new data.

Artificially generated training data, often created by another AI. Useful when real data is scarce or private, but over-reliance risks model collapse as synthetic content compounds its own flaws.

Halting training when the model stops improving on held-out data, even if it could keep memorizing the training set. A simple, effective defense against overfitting.

A setting chosen before training starts rather than learned by the model: learning rate, batch size, number of layers. Tuning hyperparameters is part science, part art, and a big factor in model quality.

Data held back from training and used to check progress along the way. If performance on the validation set stalls while training loss keeps dropping, the model is probably overfitting.

Data the model never sees until final evaluation — the honest exam after all studying is done. Good practice keeps it strictly separate from training and validation.

Everything done to a model after pre-training: instruction tuning, RLHF, safety training. Pre-training builds raw capability; post-training shapes it into a useful, safe product.

More pre-training on new data after the initial run — fresh news, a specialized domain, another language. A way to update or specialize a model without starting over.

Supervised Fine-Tuning: training a model on labeled examples of desired behavior — sample instructions paired with ideal responses. The standard first step of post-training.

Training a model to prefer better answers over worse ones, using human rankings or comparisons. RLHF and DPO are both preference-tuning methods with different machinery.