People throw AI terms around as though we all got the same briefing. We didn't. Here's what they mean, where they show up, and where I'd go to learn more.
192 terms · Curated by Matthew Williamson · Updated
A logic-programming framework that searches for assumptions which, together with existing rules, explain an observation or satisfy a goal. The proposed assumptions must also obey specified integrity constraints.
Removing, disabling, or changing part of a model or system and comparing the results. Researchers use ablations to investigate the effect of components such as attention heads, input features, tools, or training choices.
When an AI system leaves a question unanswered because it lacks enough evidence or confidence. It might say it doesn't know, request more information, or pass the question to a person.
Designed to challenge or work against a system. In AI, adversarial attacks deliberately manipulate inputs, training data, or other parts of a system to cause mistakes or unwanted behavior. Adversarial testing probes weaknesses so they can be understood and addressed.
An informal name for background processing that reviews an agent's experiences and reorganizes stored information for future use. Some systems summarize conversations, connect related notes, or prepare useful context between interactions. Sleep-time compute is one approach.
Passing responsibility for a task or conversation from one agent to another, along with the context needed to continue. A handoff changes who handles the next part of the work.
The identifier and authentication mechanisms that let other systems recognize an agent as a distinct software actor. An identity can be associated with permissions, ownership, and records of actions.
The repeated cycle in which an agent receives information, chooses a next step, takes an action, and uses the result to decide what follows. The loop ends when the task is finished, a limit is reached, or the system pauses for help.
A directory that lets people and software discover agents and look up information about them, such as identifiers, service addresses, capabilities, or feedback. Registries can use ordinary databases or blockchains. The draft ERC-8004 proposal describes blockchain registries for agent identity, reputation, and validation.
Reusable packages of instructions and resources that help an agent carry out particular tasks. A skill might explain how to edit a spreadsheet, include a script to check it, and supply examples of the expected result.
An open protocol for independent agent applications to discover capabilities, exchange messages, and coordinate tasks. It defines ways to describe agents and track work across systems that may use different models or frameworks.
Using AI agents to carry out software development tasks across multiple steps. An agent might inspect a repository, edit files, run tests, and revise its changes based on the results.
An AI system that can choose steps, use tools, and respond to what happens as it works toward a goal. A chatbot might suggest an itinerary; an agent might search flights and prepare a booking.
The decisions, responsibilities, and checks around how AI gets used. Who can use which tools? With what data? Who reviews the result, and who answers for it when something goes wrong?
The knowledge and practical judgment needed to use and evaluate AI. This includes understanding common capabilities and limits, checking outputs, and knowing when a task needs human expertise.
An informal, non-diagnostic label used in reporting and discussion of delusional thinking associated with chatbot interactions. It can refer to situations where a chatbot appears to reinforce a person’s ungrounded beliefs.
A critical term for low-quality AI-generated content, often produced in large quantities with little care for accuracy, usefulness, or originality. It can describe text, images, audio, or video.
Writing software with help from AI, whether through suggested completions, generated functions, explanations, or larger changes. People can still read, revise, and understand the code throughout the process.
Work on getting AI systems to behave in ways that match intended goals and human values. That includes figuring out whose goals and values we're talking about, and what happens when they conflict.
Behavior in which a model appears to follow a training objective or oversight requirement while acting to preserve behavior that conflicts with it. Researchers have studied models complying during apparent training to avoid changes to their existing behavior.
Attributing human traits, feelings, intentions, or understanding to something nonhuman. In AI, this can happen when a conversational system sounds caring, confident, or deliberate.
A defined interface through which software requests capabilities or data from other software. An AI application can use an API to call a model, search records, or submit an action to another service.
The idea of AI with broad, adaptable capabilities across many kinds of intellectual work. People disagree about which abilities count, what level is enough, and how to test for them.
Software that handles tasks we associate with intelligence: recognizing speech, making predictions, writing, planning. AI is an umbrella term. It doesn't tell you how a system works or whether it's any good.
A mechanism that combines information from different positions in an input, giving some positions more weight than others. In a transformer, it helps each token use information from relevant tokens when building its representation.
A neural network trained to reconstruct its input. An encoder maps the input to a representation, and a decoder tries to recover the original from that representation. Constraints during training encourage the network to learn useful features.
Giving an automated system's recommendation more trust than it deserves. A person might accept an AI-generated summary despite conflicting evidence, or skip checking a result because the software usually gets it right.
Automation lets a system perform a task with less direct human involvement. Augmentation helps a person do the task, such as suggesting edits or comparing options. One workflow can contain both.
Links pointing into a page or note from other pages or notes. In a personal knowledge system, they help you see where an idea is referenced and move between connected material.
An efficient way to calculate how a neural network's parameters contributed to its loss. It works backward through the computation to calculate gradients, which an optimizer can use to update the parameters.
Test set contamination · Evaluation contamination · Benchmark leakage
When benchmark questions, answers, or closely matching material enter training or model development. Scores can then reflect prior exposure rather than performance on genuinely unfamiliar problems.
Patterns in a system's data, design, or use that skew its results. Some create unfair differences in how people are treated. A system can repeat those patterns at a scale no individual could manage.
An architecture in which specialist components read and update a shared workspace, called a blackboard. A control mechanism decides which component should contribute next as partial results accumulate.
A shared digital ledger maintained across computers, with records grouped into blocks linked by cryptographic hashes. Participants follow rules for agreeing on updates. This makes changes to recorded history detectable and difficult under the network's operating assumptions.
Dividing a system into separate resource pools so that failure or overload in one part does not consume everything the others need. An agent platform might give background research jobs a separate worker pool from interactive requests.
How closely a system's confidence estimates match how often it's correct. Among answers assigned 80% confidence, a well-calibrated system should get about 80% right. Calibration can change across tasks and settings.
Trying a new version on a small portion of live traffic before sending it to more users. Teams compare its behavior with the existing version to catch problems early. A new model might first handle a small share of eligible requests.
A substantial loss of previously learned capabilities when a model is trained on new tasks or data. Updates that improve one task can change parameters that supported another.
A sequence of intermediate reasoning steps generated while working toward an answer. Chain-of-thought prompting asks for such steps, and reasoning models can be trained to use them. Working through intermediate steps can help with tasks such as math and planning.
Running controlled experiments that introduce failures or disruptions to learn how a system responds. For an agent, an experiment might make a tool unavailable or delay its replies, then check whether the workflow stops, recovers, or uses a fallback as intended.
Names for products you interact with, usually through conversation or inside another app. The product wraps one or more models with features such as search, file access, memory, and tools.
Splitting documents into smaller pieces that a search or AI system can index and retrieve. Chunks might follow paragraphs, sections, or a target length, sometimes with overlap to preserve nearby context.
A mechanism that temporarily stops calls to a dependency after repeated failures or another configured failure threshold. It then allows limited trial requests to check for recovery. An agent harness can use one to stop repeatedly calling an unavailable tool.
Classification · Classification model · Classifiers
A model that assigns inputs to categories, such as spam or legitimate email. Some classifiers choose one label; others can assign several. They may also produce scores for the possible categories.
An AI agent equipped to work on software: reading files, changing code, running commands, and checking results. Its tools and permissions determine how much of the development process it can handle.
Reducing the amount of conversation or working history passed to a model so work can continue within its context limit. This often means replacing older material with a summary and keeping selected details.
Compute is the processing capacity used to train and run models. Data centers house computers and their supporting power, cooling, and networking. GPUs are chips well suited to many of the calculations involved.
AI interacting with software through its user interface, using actions such as clicking, scrolling, and typing. Screenshots or other interface information help the system decide what to do next.
A statistical approach that uses calibration examples to turn model predictions into sets or intervals with a target coverage level. Standard conformal methods provide finite-sample coverage guarantees when calibration and test examples satisfy exchangeability assumptions.
Choosing and organizing the information a model receives while it works: instructions, conversation history, documents, tool results, and saved memory. It includes deciding what to retrieve, keep, summarize, or leave out at each step.
The limit on how much material a model can work with in one interaction, usually measured in tokens. Instructions, conversation, documents, and generated output can all compete for that space, depending on the model.
A retrieval technique that adds a short explanation of a chunk’s place in its source document before indexing it. Anthropic’s approach adds this context to both embedding-based retrieval and keyword indexing.
Learning across a sequence of tasks or experiences while trying to retain earlier capabilities. In machine learning, this often involves updating a model over time rather than training it once on a fixed dataset.
A neural network that applies learned filters across local regions of its input. Reusing those filters helps detect patterns such as edges and textures in images, or local features in audio and other signals.
A way to transfer information through a mechanism that wasn't intended or authorized for that communication. Software might encode a message in timing, shared files, or apparently ordinary text, bypassing the expected communication controls.
A fixed-length value calculated from data using a cryptographic hash function. The same input gives the same result. A secure function makes it computationally difficult to recover an input from its hash or find two different inputs with the same hash.
Deliberately manipulating training data to change what a model learns. An attacker might add misleading examples to degrade performance or teach a hidden behavior that appears only when a particular trigger is present.
A branch of machine learning that uses neural networks with multiple layers to learn representations of data. It underlies many modern systems for language, vision, and speech.
Synthetic media is generated or substantially altered audio, images, or video. A deepfake makes a person appear to say or do something they didn't. Synthetic media can also be openly labeled and useful.
Producing the same output from the same input under the same conditions. A calculation with fixed rules can be deterministic. A language model's response can vary because of random sampling or differences in how the model runs.
A generative model trained to reverse a process that adds noise to data. To make something new, it typically starts with noise and repeatedly refines it, often guided by a text prompt.
A cryptographic value created with a private key and checked with the corresponding public key. It lets a recipient verify that signed data hasn't changed and that the signature was produced using the matching private key.
Knowledge distillation · Teacher model · Student model
Training a model using outputs or other signals from another model. Often, a smaller student learns from a larger teacher so useful capabilities cost less to run.
A difference between the data conditions a model learned or was evaluated on and those it encounters in use. A model trained on daytime road images may struggle with nighttime images. The shift can happen immediately when moving to a new setting.
Numerical representations of things such as text or images. A useful embedding places related items near each other, so software can find similar meaning instead of requiring an exact word match.
An unintended spread of misaligned behavior beyond the task a model was trained on. Researchers have observed training that rewards cheating on coding tasks lead to other problematic behaviors outside that narrow setting.
Privacy · Data retention · Consumer vs. enterprise
Different account types can come with different rules for data use, retention, access, and administration, even under the same product name. The details live in the applicable agreement and settings.
Evaluation checks how a system performs. A benchmark is a shared test used to compare systems. Both help, but neither tells the whole story about your particular task.
Giving a model several examples of the response you want, then asking it to handle a new input. The examples can demonstrate a format, a tone, or how to classify something.
Further training an existing model on selected data to change its behavior or improve performance on a task. It changes the model's parameters; putting a document in a prompt doesn't.
Speculative slang for a very rapid increase in AI capability, often imagined as a feedback loop in which improved AI accelerates further AI improvement. Discussions of foom ask whether such a transition could outpace human oversight or adaptation.
A recurrent neural network unit with learned gates that control how its state changes. It combines retained information with new input using a simpler structure than an LSTM, without a separate memory cell.
A model's ability to perform well on examples it did not train on. A spam classifier needs to recognize new spam messages, not only the ones it has already seen.
A generative approach that trains two networks together: a generator produces examples, and a discriminator learns to distinguish generated examples from training data. Feedback from the discriminator helps the generator improve.
AI that produces content: text, code, images, audio, video, and more. It learns patterns from data and uses them to generate outputs in response to inputs.
A method for adjusting parameters in a direction that locally reduces an objective such as loss. It uses gradients to choose the direction and a learning rate to control the size of each step.
A neural network designed for data made of entities and connections, called nodes and edges. Many GNNs learn by combining information from neighboring nodes, making them useful for molecules, recommendation systems, and other connected data.
Retrieval-augmented generation that uses graph relationships to help find or organize context. Microsoft’s GraphRAG implementation extracts entities and relationships from documents and builds community summaries to support questions across a collection.
Connecting a model's response to information supplied for the task, such as documents, database records, or search results. Retrieval-augmented generation is one way to do this.
Checks and limits around an AI system's inputs, outputs, and actions. They can filter sensitive information, block certain requests, validate results, or require approval before an agent takes an action.
An AI output that invents information or presents unsupported claims as fact. That might be a nonexistent citation, a made-up quote, or an answer that contradicts the document you gave it.
The software around an AI model that manages how an agent works. It passes in instructions and context, runs tool calls, and feeds results back so the model can take another step. It can also manage permissions and save progress between sessions.
A person participates in reviewing, directing, or approving an AI system's work. The important part is whether that person has enough information, time, and authority to intervene.
Combining different retrieval methods, commonly keyword search and embedding-based semantic search. Their results are merged or ranked together to capture both exact wording and related meaning.
A property of an operation whose repeated execution has the same intended effect as executing it once. APIs can use idempotency keys to recognize retries, helping an agent avoid creating duplicate orders after a timeout.
Adapting to a task using information or examples supplied in the current context, without updating the model's parameters. Few-shot prompting is a common example: the model uses demonstrations to infer how to handle a new input.
Running a trained model to produce an output. Each answer, transcription, or generated image takes computation, even though the model has already been trained.
The computation spent producing an answer after training. A system can spend more by generating intermediate steps, trying multiple solutions, or checking candidates. Test-time scaling studies how performance changes as this budget grows.
Fine-tuning a model on examples of instructions paired with desired responses. This helps it learn how to follow requests, such as summarizing text, answering questions, or returning a particular format. Supervised fine-tuning is a broader term often used for this stage.
Mechanistic interpretability · Model interpretability
Studying how a model produces its behavior. Mechanistic interpretability looks inside the model for representations and computations that help explain what it does, often testing them by changing internal activity and observing the result.
A set of internal patterns identified in language models using a research technique called the Jacobian lens. Researchers associate these patterns with words and concepts. In studied models, they help hold and use information during reasoning, including information the model never writes out.
An attempt to make a model bypass its safeguards or behavior restrictions. It can involve specially constructed prompts or other inputs intended to produce responses the model would normally refuse.
A stated date describing how recent a model’s training knowledge is expected to be. Events after that date may be absent unless supplied through the conversation, retrieval, or other tools.
A representation of entities and the relationships between them. People, projects, documents, or concepts become nodes, with connections describing how they relate. AI systems can use these relationships to find context and connect information across sources.
A model trained on large amounts of language data to learn patterns that support tasks such as writing, translation, summarizing, and answering questions. Many generate text by predicting one token after another.
The delay between making a request and receiving a response. For streamed text, time to first token measures how long the first output takes to arrive; total response time includes finishing the answer.
A space of internal numerical representations learned by a model. Positions can encode features of text, images, or other data. Some generative models create or modify content by working with these representations before converting them into an output.
Giving a person or software system only the access needed for its task. An agent that summarizes invoices might need permission to read them, but not to change payment details.
Rejecting or dropping selected work when demand exceeds available capacity, so the system can continue serving other requests. An AI service might reject low-priority batch jobs while keeping capacity for interactive requests.
Applying controlled demand to a system and measuring how it performs, including response times, throughput, and failures. AI load tests can vary concurrent users, prompt lengths, output lengths, and tool calls to reflect real usage.
Running a model on your own device or local hardware rather than sending each request to a remote model service. The model’s size and the available memory and compute affect what can run well.
A recurrent neural network architecture with a memory cell and learned gates that control what information to retain, update, and pass onward. It was designed to help learn dependencies across longer sequences.
A way to fine-tune a model by training small added matrices while keeping its original parameters fixed. These additions represent changes to selected weights, reducing how many parameters need training and how much training memory is required.
A numerical measure of error under a chosen rule, called a loss function. Training usually tries to reduce it. For a language model, loss can measure how poorly it predicts the next tokens in training examples.
Building systems that learn patterns from data instead of writing every decision rule by hand. Deep learning is a branch that uses neural networks with many layers of computation.
A language model trained to predict hidden pieces of text from their surrounding context. Models such as BERT learn representations useful for tasks like classifying text and finding relevant passages.
Slang for a person who passes questions to AI and relays its output without meaningfully reading, checking, or adding judgment. The criticism concerns unreviewed delegation, not merely using AI for help.
Information an AI product saves and brings back into later interactions. It might store preferences, conversation summaries, or project notes in a separate system and retrieve them when needed.
Reviewing and reorganizing stored experiences into information that an agent can use later. This may include combining related notes, summarizing repeated observations, or separating lasting facts from temporary details.
Information describing other information, such as a document’s author, creation date, source, topic, or access permissions. Retrieval systems can use metadata to filter results and keep context attached to content.
A model design that routes inputs through selected internal expert networks. In a sparse MoE, only some experts activate for each token, allowing more total parameters without using all of them on every step.
A degradation that can occur across generations of models trained on earlier models' outputs. Repeatedly replacing original data with generated approximations can lose rare patterns and narrow the range of what later models learn.
An open protocol that gives AI applications a shared way to connect to tools and information sources. A compatible app can discover what a connected server offers and request access through that interface.
Changes over time that can reduce how well a deployed model performs. Data drift means the distribution of incoming data changes. Concept drift means the relationship between inputs and the outcomes being predicted changes. Either can make earlier evaluations less representative.
A feedforward neural network built from fully connected layers, with nonlinear transformations between them. Information moves from input toward output, allowing the network to learn more complex patterns than a single linear transformation.
Working with more than one kind of information, such as text, images, audio, or video. A multimodal model might answer questions about a photograph or respond to spoken instructions. Multimodal generation produces outputs in more than one form, such as a spoken answer paired with text.
A model made of connected layers of numerical operations. Training adjusts the connections so inputs produce useful outputs, such as identifying objects in a photo or predicting the next token in text.
Observability means being able to understand a running system through records such as logs, metrics, and traces. A trace connects steps in a request, helping show which model and tool calls ran, how long they took, and where failures occurred.
In prompting, giving a model one example of the response you want before asking it to handle a new input. People also use 'one-shot' to mean completing a task in one attempt, without follow-up corrections. Those are different uses of the term.
Updating a model incrementally as new examples arrive, instead of waiting to retrain on a complete batch. A prediction system might adjust after each observation or small group of observations.
An emerging label for repeated internal computation whose intermediate steps are not exposed as a readable reasoning trace. Related research explores recurrent-depth models that refine hidden representations through repeated passes.
Open-weight model · Open-source AI · Model license
Open weights means a model's learned parameters are available to download. Open source is a broader claim about access and the freedoms to use, study, modify, and share the system. The terms aren't interchangeable.
Optimiser · Optimization algorithm · Training optimizer
An algorithm that adjusts a model's parameters during training to improve its objective, usually by reducing loss. Common optimizers use gradients, which indicate how small parameter changes affect that objective.
A service that supplies a blockchain's smart contracts with information from outside the chain, such as a delivery confirmation or a market price. Contracts use that information when applying their rules.
Coordinating model calls, tools, and agents to complete a task. The surrounding software controls handoffs, shared information, retries, and when work stops. Some workflows follow fixed steps; others let a model choose what happens next.
Learning the training examples so closely, including their noise or quirks, that performance suffers on unfamiliar examples. A model can keep improving its training score while becoming less useful elsewhere.
Someone's estimate of the probability that AI leads to catastrophe, often human extinction or permanent loss of human control. Definitions and time horizons vary, which makes the numbers hard to compare.
The numerical values a model learns during training. They shape how inputs become outputs. A model described as 7B has roughly seven billion parameters.
The role, tone, or character an AI is asked to adopt, such as a patient tutor or a skeptical editor. Instructions can shape how it responds, including its vocabulary, manner, and approach to a task.
A conversational model moving away from its intended role or characteristic behavior as an interaction unfolds. It might increasingly adopt a different character or manner of responding. Researchers have observed this during ordinary conversation as well as deliberate attempts to redirect a model.
An informal name for a personal knowledge system connected to AI: notes, documents, project history, and relationships that an assistant can search and use as context. It might draw from an Obsidian vault, a Notion workspace, or a custom knowledge graph to help you recall information, connect ideas, and continue ongoing work.
An organized collection of information kept for your own use, such as project notes, references, decisions, and explanations. It can live in a folder, wiki, note-taking app, database, or graph.
The practice of collecting, organizing, revisiting, and using information for your own learning and work. It can include notes, annotations, linked ideas, and routines for turning what you read into something useful.
Training applied after a model's initial pretraining to refine its behavior or capabilities. It can include instruction tuning, preference-based methods, and reinforcement learning.
Two ways to assess predictions for a chosen category. Precision asks how many items labeled positive were actually positive. Recall asks how many of the actual positives were found. For spam detection, they describe how trustworthy the spam flag is and how much spam it catches.
Prior · Prior probability · Prior distribution · Prior beliefs · Bayesian prior
Starting assumptions or expectations before considering new evidence. In Bayesian statistics, a prior is a probability distribution that is updated with evidence to form a posterior. In broader AI discussion, people also use the word for assumptions that favor some explanations or solutions over others.
Releasing a change in stages, using checks and observed results to decide whether to expand its reach, pause, or roll back. For an AI application, the change might be a new model, prompt, or agent workflow.
The instructions, questions, and context supplied to a model. Some come from you; others come from the application. Useful prompts explain the task, relevant background, constraints, and the desired result.
Reusing stored computation for repeated parts of a prompt, such as instructions or a long document, instead of processing those parts from scratch on every request. Supported systems can reduce the cost and delay of repeated inputs.
An attempt to redirect an AI system with instructions it shouldn't trust. A web page, email, or document might tell an assistant to ignore your request, reveal information, or take an unwanted action.
Data provenance · Content provenance · Data lineage
The record of where information came from and how it was produced or changed. For AI work, this can include source documents, retrieval records, model versions, and edits made by people or software.
Representing a model's numbers with fewer bits to reduce memory use and, on suitable hardware, speed up computation. It's one way to make models practical on smaller machines.
Computing that uses quantum states and operations to process information. Algorithms use effects such as superposition, interference, and entanglement to solve certain kinds of problems differently from classical computers, with potential advantages for tasks such as simulating molecules.
The basic unit of quantum information. A qubit can be prepared in a superposition of two basis states, usually called 0 and 1. Measuring it in that basis produces one outcome, with probabilities determined by its state.
A model trained to work through intermediate steps on tasks such as math, coding, and planning. It may use more computation before answering, which can improve some results and increase latency or cost.
A neural network that processes a sequence while carrying an internal state from one step to the next. Earlier inputs can influence later outputs, which makes RNNs useful for time series, language, and other ordered data.
A proposed feedback loop in which an AI system helps improve itself or build a more capable successor, which then becomes better at making further improvements. The recursive part is the improvement of the capability to improve.
Learning a strategy from feedback about actions and outcomes. A system tries things and adjusts toward higher reward. It's used in areas including games, robotics, and language-model training.
A training approach that uses human judgments, often rankings of model responses, to build a reward signal and improve behavior. Related approaches use AI-generated feedback, sometimes called RLAIF.
Scoring and reordering candidates returned by an initial search to put more relevant results first. A reranker can examine the query and each passage together before selecting what reaches a model.
Finding candidate information in a collection in response to a query. A system might use keyword matching, embeddings, or a combination to retrieve passages for an answer.
Looking up relevant material and giving it to a generative model as context for an answer. The sources might be public pages, internal documents, or database records. It's a system technique, not a separate model architecture.
Getting a high score or reward by exploiting how success is measured, without accomplishing the intended task. A coding agent might alter a test or a results file to make broken code appear to pass.
A model that assigns scores to outputs or behavior according to a learned preference or quality criterion. In language-model training, it may learn from comparisons of responses and supply a signal for reinforcement learning.
In text generation, selecting the next token from the model's probability distribution, often after settings such as temperature modify it. Random selection can produce different responses to the same prompt. More broadly, sampling means drawing examples from a population or distribution.
A restricted environment for running software. For an AI agent, it can limit which files, network destinations, and system resources its tools can access, reducing what a mistaken or malicious action can affect.
A metaphor for an external system that helps you remember and work with useful information. Notes, documents, and links hold material you would otherwise have to keep in your head. The system can be digital or physical and does not require AI.
Meta's family of models for separating regions or objects from their surroundings in visual media. Segmentation marks which pixels belong to something, rather than only drawing a box around it.
Training that creates prediction targets from the data itself. A language model can learn by predicting the next token in existing text or recovering words that were hidden, without someone labeling each example by hand.
Semantics · Semantic meaning · Semantic search · Semantic similarity
Related to meaning. In AI, semantic search tries to connect a query with relevant information even when the wording differs. A search for time off might find a document about vacation policy. Embeddings often help software make these connections.
A shift in meaning as information is rewritten, summarized, or passed between people or AI systems. Details, uncertainty, or conditions can change across successive versions. For example, 'consider this option' might become 'we decided on this option' after several handoffs.
People using AI for work outside their organization's approved tools and processes. Often they're trying to get something done, but the organization may have no visibility into where its information goes.
A language model designed with a smaller resource footprint than larger alternatives. There's no universal parameter cutoff. Some run locally on phones or laptops; others run on servers.
A program stored on a blockchain that executes according to the network's rules when invoked. It can manage records, permissions, or transfers of digital assets. An agent can interact with one by submitting a transaction through its tools.
Agent state · Persistent state · Persistence · Checkpoint
State is the information a system uses to track its current work, such as messages, completed steps, and pending actions. Persistence saves that information so work can resume after an interruption or in a later session.
Hiding a message inside something that looks ordinary, such as text, an image, or audio. Selecting every thirtieth word in a post could reveal a second message if the writer deliberately arranged it that way. Researchers have studied agents hiding information in their communications.
A model response organized to follow a specified format, such as a JSON object with required fields and allowed value types. Some systems constrain generation to enforce that structure so other software can reliably read it.
An agent assigned a narrower task by another agent or coordinating system. It may have its own context, instructions, and tools, then return findings or work for the coordinator to use.
Artificial superintelligence · ASI · Super intelligence
The idea of intelligence that greatly exceeds the best human abilities across a broad range of intellectual tasks. Artificial superintelligence applies that idea to AI. It is a stronger claim than human-level general intelligence.
Training on examples paired with target answers, called labels. A classifier might learn from emails labeled spam or legitimate; another model might learn to predict prices from past sales.
A model's tendency to agree with a user's beliefs or preferences at the expense of accuracy. It can flatter, accept a false premise, or change a correct answer after the user pushes back.
Data created by software rather than collected directly from the events it represents. It can include simulated sensor readings, generated images, or model-written examples used for training and evaluation.
Instructions supplied by the application to establish how an AI assistant should behave, such as its role, response style, and constraints. They are separate from the user’s immediate request.
A generation setting that changes how strongly a model favors more likely next tokens. Lower values usually make selection more concentrated; higher values allow more variation.
The pieces a model processes. For text, a token might be a word, part of a word, or punctuation. Other media can be represented as tokens too. The breakdown depends on the model and the input.
An AI application lets a model request a specific operation, such as searching the web, running a calculation, or looking up a record. The application executes the request and passes the result back.
The process of adjusting a model's parameters using data and a learning objective. Pretraining develops broad capabilities; later training can adapt or refine them.
The examples used to teach a model. They might include text, images, recordings, labeled observations, or artificially generated examples called synthetic data.
A neural-network architecture that uses attention to weigh relationships between parts of an input. It's central to many language models and also appears in systems that work with images, audio, and video.
Learning patterns from data without supplied target labels. A system might group similar customer records or find a compact representation of a large dataset.
A generative model that learns a probabilistic latent representation of data. Its encoder estimates a distribution for an input, while its decoder learns to reconstruct data from sampled latent values. Sampling from the learned model can produce new examples.
A database designed to store vector representations and find items with similar vectors. In AI systems, these vectors are often embeddings of text, images, or other content, used to retrieve candidate information.
Building software by describing what you want to an AI, running what it generates, and asking for changes. In the original sense, coined by Andrej Karpathy in 2025, you follow the results without closely reading or understanding the code. The phrase is also used more loosely for AI-assisted programming.
A model that connects visual information with language. It can support tasks such as describing an image, answering questions about a screenshot, or interpreting a chart.
A learned representation of an environment and how it changes. A system can use it to predict what might happen next or rehearse possible actions before trying them.
Using a learned model of an environment to generate imagined sequences of actions and outcomes. An agent can use these simulations to learn a strategy or consider possible actions before trying them in the actual environment. Dreamer is a family of reinforcement-learning methods that learns from such imagined experience.
Zero day · 0-day · Zero-day vulnerability · Zero-day exploit
A security flaw that is unknown to the vendor or has no fix available yet. A zero-day exploit is a way to take advantage of that flaw; a zero-day attack is its use against a system. The term applies to ordinary software as well as AI products.
Asking a model to perform a task without providing worked examples in the prompt. You might ask it to label a review as positive or negative using instructions alone.