The AI Glossary

Short for 'effective accelerationism,' an online movement arguing that AI development should be pushed forward as fast as possible. It is often positioned as the optimistic counterpoint to AI doomers who urge caution.

Marketing a basic chatbot or scripted automation as a fully autonomous 'AI agent.' The agentic-era cousin of AI washing, it inflates how much independent planning a product actually does.

Autonomous systems that don't just answer questions but independently plan and execute multi-step tasks across apps, tools, and the web to reach a goal. Instead of one response per prompt, an agent breaks work into steps and iterates toward a result.

AI for IT Operations — using machine learning to monitor systems, detect anomalies, correlate alerts, and automate incident response across complex infrastructure.

The research field focused on making AI systems reliable, controllable, and free of harmful or unintended behavior — especially as models become more capable and autonomous.

Derisive slang for low-quality, mass-produced, soulless AI-generated content — generic blog posts, malformed images, robotic video — flooding the web. Merriam-Webster named 'slop' its 2025 Word of the Year.

Slapping an 'AI-powered' label on legacy products that use little or no real artificial intelligence, purely to ride the hype. A deceptive marketing practice that exaggerates AI's actual role.

The field of safety research focused on keeping an AI system's goals, actions, and outputs faithfully matched to human values and intent — so a capable model does what we actually want.

A standardized way for software to talk to other software. AI providers expose models through APIs so developers can send prompts and receive completions without hosting the model themselves.

A theoretical milestone where an AI can understand, learn, and apply knowledge across any intellectual task at or above human capability — not just the narrow tasks it was trained for.

The broad field of building systems that mimic human cognitive functions like learning, reasoning, perception, and problem-solving.

A hypothetical intelligence that vastly exceeds the best human minds across virtually every domain — a stage beyond AGI and a central subject of long-term safety debate.

The process inside a transformer that lets a model weigh the relevance of surrounding tokens when interpreting each one, so a word's meaning is shaped by its context rather than read in isolation.

A graded scale describing how much an AI system acts on its own — from suggesting actions a human approves, up to planning and executing entire workflows with minimal oversight.

The core algorithm for training neural networks: it measures how much each weight contributed to the model's error and nudges every weight to reduce that error on the next pass.

A standardized test set and scoring method used to compare models on a capability — reasoning, coding, math, safety. Increasingly run as 'evals' tailored to a specific product or task.

Systematic skew in a model's outputs, usually inherited from imbalanced or prejudiced training data. Fairness work aims to detect and reduce harm to particular groups.

A system whose internal reasoning is opaque — you see inputs and outputs but not how the decision was made. A central motivation for interpretability and explainable AI.

Internet-culture shorthand for low-value, attention-eroding digital media — increasingly applied to endless feeds of cheap AI-generated content.

When fine-tuning a model on new data causes it to lose previously learned skills. A key challenge in continually updating models without retraining from scratch.

Prompting a model to 'think out loud,' breaking a hard problem into explicit step-by-step reasoning before giving a final answer — which sharply improves accuracy on logic and math.

A program that converses in natural language. Modern chatbots are powered by large language models, a major leap over the rigid, rule-based scripts of earlier generations.

A playful-to-pointed slang term for robots and AI, borrowed from Star Wars. It captures a mix of affection and anxiety about machines spreading through everyday life.

The raw processing power — typically GPU or TPU hours — needed to train or run a model. Compute is one of the defining costs and bottlenecks of modern AI.

A training approach where a model critiques and revises its own responses against a written set of principles (a 'constitution'), reducing reliance on large amounts of human-labeled feedback.

The 2026 successor to prompt engineering: deliberately architecting everything a model sees — instructions, memory, retrieved data, tools, and state — as a dynamic system rather than a single prompt.

The maximum amount of text (measured in tokens) a model can consider at once. Larger windows let a model reason over long documents, codebases, or conversations without losing the thread.

A neural network design specialized for grid-like data such as images, using filters that detect local patterns (edges, textures) and build them up into higher-level features.

An AI assistant embedded inside a tool — an editor, IDE, or office app — that suggests, completes, or generates work alongside the user rather than replacing them.

An attack that corrupts a model by slipping malicious or misleading examples into its training data, causing hidden failures or backdoors that surface later.

The structured collection of examples a model learns from. Its size, quality, and diversity largely determine what the resulting model can and cannot do.

A branch of machine learning using multi-layered neural networks. The 'deep' refers to those stacked hidden layers, each extracting progressively more complex features from the data.

A generative model that learns to create images, audio, or video by starting from random noise and iteratively removing it until a coherent result emerges. The basis of most modern image generators.

Models that classify or label existing data — spam vs. not-spam, cat vs. dog — rather than generating new content. The complement to generative AI.

Training a smaller, cheaper 'student' model to imitate a larger 'teacher' model, capturing much of its capability at a fraction of the size and cost.

Slang for someone who believes advanced AI poses a serious, possibly existential, risk and argues for slowing down. Usually framed against accelerationists.

Running AI directly on local devices — phones, cameras, sensors, vehicles — instead of in the cloud, for lower latency, offline capability, and better privacy.

Numerical representations of data (text, images, audio) plotted in a high-dimensional space, where similar meanings sit close together. They let machines measure semantic relationships mathematically.

Capabilities that appear only once a model crosses a certain scale, without being explicitly trained for — a surprising and debated property of large models.

The part of a model that compresses input into a dense internal representation. Encoder-based designs excel at understanding and search; decoders specialize in generation.

One full pass of the training algorithm over the entire dataset. Models typically train for many epochs, refining their weights a little more each time.

Techniques and tooling that make a model's decisions understandable to humans — critical in regulated domains like finance, healthcare, and law where 'because the model said so' isn't enough.

Steering a model by giving it a handful of examples directly in the prompt, so it infers the desired pattern without any retraining.

Taking a broad, pre-trained model and training it further on a narrow, specialized dataset to customize its behavior — for example legal review or medical triage.

A large model trained on broad data that serves as a reusable base for many downstream tasks via fine-tuning or prompting, rather than being built for one purpose.

The most capable, largest-scale models at the leading edge of the field — the ones that define the current state of the art and draw the most safety scrutiny.

A model's ability to invoke external tools or APIs in a structured way — querying a database, sending an email, running code — turning a text generator into an agent that acts.

A generative design pitting two networks against each other — one creating fakes, one detecting them — until the generator produces convincingly realistic output.

AI that creates new content — text, images, audio, video, code — rather than only classifying or predicting. The category that brought AI into mainstream daily use.

The AI-era evolution of SEO: optimizing content so it gets favorably cited and synthesized by AI search assistants and language models, not just ranked by traditional search engines.

A family of transformer-based language models pre-trained to predict the next token. The architecture popularized by OpenAI that brought LLMs to the mainstream.

The optimization method that trains most models: it repeatedly adjusts weights in the direction that most reduces error, gradually 'descending' toward a better-performing model.

A chip originally built for rendering graphics whose massively parallel design makes it the workhorse for training and running neural networks.

Tying a model's responses to verifiable external sources — documents, databases, search — so answers are based on real evidence rather than the model's parametric guesswork.

The rules, filters, and checks placed around a model to keep its behavior safe and on-policy — blocking disallowed outputs, constraining tool use, and validating actions.

When a model produces confident, fluent output that is factually wrong or invented. A core reliability challenge, mitigated by grounding, retrieval, and verification.

The scaffolding around a model that turns it into a working agent — the loop, tools, memory, and control logic that manage how it plans, acts, and recovers from errors.

A design where a person reviews, approves, or corrects an AI system's outputs at key points — balancing automation with human judgment and accountability.

A setting chosen before training that shapes how a model learns — learning rate, batch size, number of layers — as opposed to the weights the model learns on its own.

A model's ability to pick up a new task purely from instructions and examples in the prompt, with no change to its underlying weights.

The act of running a trained model to produce an output. Inference is the recurring, at-scale cost of serving AI, distinct from the one-time cost of training.

Research into understanding what is actually happening inside a model — which internal features and circuits drive its behavior — so we can trust, debug, and steer it.

A crafted prompt or technique that tricks a model into bypassing its safety guidelines and producing restricted output. A constant cat-and-mouse for model providers.

A curated store of documents or facts an AI system retrieves from to answer accurately within a specific domain — the backbone of most retrieval-augmented setups.

The date after which a model has no built-in knowledge, because its training data ends there. Anything more recent must be supplied via tools, search, or retrieval.

A structured network of entities and the relationships between them. Pairing knowledge graphs with LLMs adds precise, queryable facts to fluent generation.

Examples tagged with the correct answer — the fuel for supervised learning. Labeling is often expensive and slow, which is why self-supervised methods have grown so important.

A massive neural network trained on vast text to predict the next token, enabling it to write, reason, summarize, and code. Examples include GPT, Gemini, and Claude.

The next evolution of LLMs, built for complex, multi-step problem solving, logic, and long-horizon planning — deliberating before answering rather than reacting instantly.

The delay between a request and the model's response. Low latency is essential for interactive products like chat, voice, and real-time agents.

The formula that scores how wrong a model's predictions are. Training is the process of minimizing this number, step by step, across the dataset.

An efficient fine-tuning method that trains a small set of added parameters instead of the whole model, dramatically cutting the cost and memory of customization.

A subset of AI in which systems learn patterns from data and improve with experience, rather than following rules that a programmer wrote out by hand.

Mechanisms that let an agent retain information across steps or sessions — short-term scratchpads and long-term stores — so it can stay coherent over extended tasks.

An architecture that routes each input through a few specialized sub-networks ('experts') instead of activating the whole model, delivering strong performance at lower compute cost.

An open standard from Anthropic for connecting AI models to external tools and data sources — a kind of 'USB-C for AI' that lets any compatible model reach databases, APIs, and files consistently.

A layer that directs each request to the most suitable model — balancing cost, speed, and quality by sending easy queries to small models and hard ones to frontier models.

A setup where several specialized agents collaborate — planning, delegating, and checking each other's work — to tackle problems too big for a single agent.

Models that understand and generate across multiple data types — text, images, audio, and video — within a single system, much closer to how humans perceive the world.

AI built to do one specific task well — translation, recommendation, face recognition. Essentially all AI in production today is narrow, as opposed to general.

The discipline of getting computers to understand and produce human language — the lineage that led to today's large language models.

A model loosely inspired by the brain, made of interconnected layers of simple units ('neurons') whose weighted connections are tuned during training to recognize patterns.

A model whose trained parameters are released publicly, so anyone can run, inspect, or fine-tune it locally — distinct from closed models accessible only through an API.

Coordinating multiple models, tools, and steps into a reliable workflow — deciding what runs when, passing results between stages, and handling failures along the way.

When a model memorizes its training data instead of learning general patterns, so it performs well in testing but poorly on new, unseen inputs.

The internal numbers a model adjusts during training to encode what it has learned. Parameter count (often in the billions) is a rough proxy for a model's capacity.

AI embodied in the physical world — powering advanced robotics, autonomous drones, self-driving systems, and smart manufacturing rather than living purely in software.

The first, most expensive training stage, where a model learns broad patterns from enormous amounts of unlabeled data before any task-specific tuning.

The craft of writing precise, structured inputs that steer a generative model toward accurate, well-formatted, on-target output.

An attack that hides malicious instructions in content a model reads — a web page, a document, an email — hijacking its behavior. A top security risk for tool-using agents.

Compressing a model by storing its numbers at lower precision, shrinking memory and speeding up inference with only a small hit to accuracy — key to running models on modest hardware.

An earlier architecture for sequential data that processes inputs one step at a time while carrying state forward. Largely superseded by transformers for language tasks.

Deliberately attacking a model to find its failures, jailbreaks, and harmful behaviors before release — adversarial testing borrowed from cybersecurity.

Agents that pause to critique and revise their own work mid-task — checking results, catching mistakes, and re-planning — a defining capability trend of 2026.

Training by trial and error: an agent takes actions, receives rewards or penalties, and gradually learns a strategy that maximizes long-term reward.

A technique that aligns a model to human preferences by training it against ratings of its outputs — a major reason modern assistants feel helpful and well-behaved.

An organizational practice of building and deploying AI ethically — covering fairness, transparency, privacy, accountability, and real-world impact.

A technique that lets a model fetch relevant documents or data at query time and ground its answer in them, instead of relying only on what it memorized during training.

Learning by predicting hidden parts of the data itself — like the next word in a sentence — so models can train on raw internet text without costly human labels.

Search by meaning rather than keywords, using embeddings to find results that are conceptually related to a query even when they share no exact words.

Running models through a service that provisions compute on demand and scales to zero when idle — so teams pay per use instead of maintaining always-on GPU servers.

Employees using unsanctioned AI tools at work without IT approval — a fast-growing governance and data-security headache for organizations.

A hypothesized point at which AI self-improvement becomes so rapid that technological change outpaces human ability to predict or control it.

A compact language model designed to run cheaply and fast — often on a phone or edge device — trading some breadth for efficiency, privacy, and low latency.

Slang for someone who offloads most of their thinking to a chatbot, leaning on AI even for tasks they could easily handle themselves. Linked to the broader 'AI slop' discourse.

A skeptical metaphor arguing that language models merely remix statistical patterns from their training data without genuine understanding — 'parroting' rather than reasoning.

Training on labeled examples where the correct answer is provided, so the model learns to map inputs to known outputs — the most classic machine-learning setup.

Artificially generated training data, often produced by other AI models, used to augment scarce real data or cover rare and sensitive cases.

The hidden, high-priority instructions that set a model's role, tone, and rules for a conversation before the user's first message ever arrives.

A setting that controls randomness in a model's output — low values make responses focused and deterministic, high values make them more varied and creative.

Google's custom chip purpose-built to accelerate the math behind neural networks — an alternative to GPUs for large-scale training and inference.

Giving a model extra thinking steps at inference rather than at training — letting it deliberate longer on hard problems to boost accuracy. A foundation of large reasoning models.

The basic unit a model reads and writes — a word, subword fragment, or character. Tokens are how context length is measured and how API usage is billed.

The cost model of running AI, where every input and output token carries a price. Designing products around token usage is central to making AI features profitable at scale.

Breaking input into tokens — words, subword pieces, or characters — so a model can process it numerically. The first step in turning text into something a model can read.

The total processing power consumed to train a model, often measured in GPU-hours or FLOPs. Frontier models require enormous, capital-intensive training runs.

The foundational design behind modern LLMs. It processes sequences in parallel using attention rather than step-by-step, unlocking the scale and efficiency that made today's models possible.

Alan Turing's 1950 proposal that a machine could be deemed 'intelligent' if a human couldn't reliably tell it apart from another human in conversation. A historical, much-debated benchmark.

Finding structure in data that has no labels — clustering similar items or spotting patterns — without being told the right answers in advance.

A database optimized to store and search embeddings, finding the items closest in meaning to a query. The retrieval engine behind most RAG and semantic-search systems.

A conversational, experimental way of building software where developers describe intent in natural language and steer an AI assistant — prioritizing high-level direction and rapid prototyping over hand-writing every line.

A multimodal model that jointly understands images and text — able to describe a photo, read a chart, or answer questions about a screenshot.

Embedding a hidden, detectable signal in AI-generated text, images, or audio so it can later be identified as machine-made — a tool against misinformation and impersonation.

AI slop in the workplace — superficially polished but hollow AI-generated reports, slides, or messages that shift the real work of sense-making onto whoever receives them.

Shorthand for explainable AI — methods that surface why a model produced a given output, so humans can audit, trust, and contest automated decisions.

A landmark real-time object-detection model family that identifies and locates objects in an image in a single pass, widely used in vision and robotics.

A model performing a task it was never explicitly trained or given examples for, relying on its general knowledge to generalize from the instruction alone.