AI Terms, Explained
Clear, authoritative definitions of the AI terms you'll actually run into — from everyday basics to the words showing up in 2026's headlines. Tap any term to read the full explanation.
Agent (AI Agent) An AI system that takes actions on its own to reach a goal, not just answer. ↗ An AI agent is a system that does more than respond to a single question. Give it a goal and it can plan steps, call tools, check its own work, and keep going until the task is done. A regular chatbot answers what you ask. An agent decides what to do next.
In practice, an agent might pull data from a file, run a calculation, draft a summary, and flag the result for your review — all from one instruction. The trade-off is control: because agents act, they need clear limits and a human checking the important steps.
Agentic AI The broader shift from AI that generates content to AI that performs work. ↗ Agentic AI describes the move from tools that produce text or images to systems that carry out multi-step work on their own. The key trait is a feedback loop: the system tries something, notices when a step fails, and looks for another path instead of stopping.
This is one of the defining themes of AI in 2026. It raises the value of AI for real tasks, and it also raises the need for approvals, oversight, and clear responsibility when an automated system makes a decision.
AGI (Artificial General Intelligence) A hypothetical AI that matches human ability across almost any task. ↗ AGI refers to an AI that could handle nearly any intellectual task a person can, rather than being good at one narrow thing. Today's systems are narrow by comparison — strong in specific areas, limited outside them.
Whether AGI is close, far, or even well-defined is debated among researchers. Treat confident predictions in either direction with caution; the term is used loosely in marketing and precisely almost nowhere.
AI (Artificial Intelligence) Software that performs tasks usually thought to need human intelligence. ↗ AI is the broad field of building software that can do things we associate with human thinking — recognizing patterns, understanding language, making predictions, generating content. It is an umbrella term that covers many methods, from older rule-based systems to today's large language models.
When people say "AI" in everyday conversation now, they usually mean the generative tools that write, summarize, and answer questions. That is one branch of a much larger field.
Alignment The work of making AI systems behave in line with human intent and values. ↗ Alignment is the effort to make sure an AI does what people actually want, including the parts they did not spell out. A system can follow instructions literally and still produce a harmful or useless result if its goals are poorly matched to yours.
This matters more as systems act with less supervision. Alignment work spans how models are trained, how they are tested, and what guardrails sit around them in real use.
API (Application Programming Interface) A connection that lets one piece of software talk to another. ↗ An API is a defined way for two programs to exchange information. For AI, it usually means a service that lets your software send a prompt to a model and get a response back, so you can build the model into your own app or workflow.
You do not need to run the model yourself — the API handles that. You send input, you get output, and you are typically billed by how much you use.
Attention The mechanism that lets a model weigh which words matter most to each other. ↗ Attention is the core idea behind modern language models. As the model reads text, it learns which words to focus on when interpreting each other word — connecting "it" to the noun it refers to, for example, even across a long sentence.
This mechanism, introduced in the transformer architecture, is what allowed models to handle long, complex passages far better than earlier approaches and made today's systems possible.
Automation Using technology to carry out tasks with little or no manual effort. ↗ Automation is having software handle work that a person would otherwise do by hand. It predates AI by decades, but AI widens what can be automated — moving from rigid, rule-based steps to tasks that involve language, judgment, and messy inputs.
The practical question is rarely "can this be automated" but "should it be, and with how much human oversight." AI raises the ceiling on the first and the stakes on the second.
Benchmark A standard test used to compare how well AI models perform. ↗ A benchmark is a shared test that measures models on the same task so their results can be compared. Benchmarks cover things like reasoning, coding, math, and reading comprehension.
They are useful but imperfect. A high benchmark score does not always mean a model is better for your specific use, and models can be tuned to perform well on popular tests without being more capable in general. Treat scores as one signal, not the whole story.
Bias Systematic skew in an AI's outputs, often inherited from its training data. ↗ Bias in AI means the system consistently leans a certain way in a manner that is unfair or inaccurate. It usually comes from the data the model learned from — if that data reflects human prejudices or gaps, the model can repeat them.
Bias is hard to remove fully because it is baked into patterns the model absorbed. Awareness, testing, and human review are the practical defenses, especially for decisions that affect people.
Bot A software program that runs automated tasks, sometimes powered by AI. ↗ A bot is any program that performs tasks automatically. Many bots are simple and rule-based — they follow fixed scripts. Others are powered by AI and can hold a conversation or handle varied requests.
The word covers a wide range, from a basic auto-reply to a sophisticated AI assistant. "Bot" alone tells you something is automated, not how capable it is.
Chatbot A program you converse with in natural language to get answers or help. ↗ A chatbot is software designed for back-and-forth conversation in plain language. Early chatbots followed scripted menus. Modern ones, built on large language models, can understand free-form questions and respond flexibly.
The quality varies enormously depending on the underlying technology. The same word covers a phone-tree menu and a capable AI assistant, so judge by what it can actually do.
Context Window How much text a model can consider at once, including your input and its reply. ↗ The context window is the amount of text a model can hold in mind during a single exchange. It covers your prompt, any documents you include, and the response. Measured in tokens, it sets a hard limit on how much the model can work with at one time.
A larger window lets you feed in long documents or keep a longer conversation coherent. Once you exceed it, the model starts losing track of the earliest material.
Context Engineering Deliberately shaping what information a model sees to get better results. ↗ Context engineering is the practice of carefully choosing and arranging what you put in front of a model — instructions, examples, retrieved documents, and conversation history — so it has what it needs and not what distracts it.
It is a step beyond writing a single prompt. As tasks get more involved, deciding what context the model receives, in what order, often matters as much as the wording of the request itself.
Data The information AI systems learn from and operate on. ↗ Data is the raw material of AI. Models learn patterns from large collections of text, images, or other information, and the quality and breadth of that data shape what the model can do.
The saying "garbage in, garbage out" holds firmly here. A model trained on narrow or flawed data inherits those limits, no matter how advanced the method.
Dataset A structured collection of data used to train or evaluate a model. ↗ A dataset is an organized body of data assembled for a purpose — training a model, testing it, or fine-tuning it for a specific job. It might be millions of web pages, labeled images, or records specific to one field.
The makeup of a dataset has direct consequences. What it includes, what it leaves out, and how it is labeled all influence the behavior of any model built on it.
Deep Learning A method using many-layered neural networks to learn complex patterns. ↗ Deep learning is a branch of machine learning that stacks many layers of artificial neurons, letting a system learn increasingly abstract patterns from data. The "deep" refers to the number of layers.
It is the approach behind most modern AI breakthroughs, including image recognition and large language models. Its strength is handling messy, high-dimensional data; its cost is needing large amounts of data and computing power.
Diffusion Model An AI that creates images by gradually turning noise into a picture. ↗ A diffusion model generates images by starting from random visual noise and refining it step by step until a coherent picture emerges, guided by your text description. It learned this by studying how images break down into noise and reversing the process.
This approach powers many popular image generators. It is why you can describe a scene in words and receive an original image built to match.
Distillation Training a smaller model to copy the behavior of a larger one. ↗ Distillation is a technique where a compact model learns to imitate a bigger, more capable model. The large model acts as a teacher, and the smaller "student" model captures much of its ability at a fraction of the size and cost.
The payoff is efficiency: a distilled model can run faster and cheaper, which matters for putting AI on phones, laptops, or high-volume services.
Embedding A numeric representation of meaning that lets AI compare and search by similarity. ↗ An embedding turns a word, sentence, or document into a list of numbers that captures its meaning. Items with similar meaning end up with similar numbers, so software can measure how related two pieces of text are.
Embeddings power semantic search and recommendation. They are also the backbone of retrieval systems that let AI find relevant information in a large collection before answering.
Edge AI Running AI directly on a local device instead of a remote server. ↗ Edge AI means the model runs on the device in your hand or on-site — a phone, camera, or laptop — rather than sending data to a distant data center. The processing happens at the "edge" of the network, close to where the data is created.
The benefits are speed, working offline, and keeping data local for privacy. The trade-off is that local hardware limits how large a model you can run.
Evaluation (Eval) The process of measuring how well an AI system performs a task. ↗ Evaluation, often shortened to "eval," is how teams check whether an AI is doing its job well. It usually means running the system against a set of test cases with known good answers and grading the results.
Good evaluation is harder than it sounds, especially for open-ended tasks where there is no single right answer. It is essential, though — without it, you are guessing whether a change made the system better or worse.
Fine-Tuning Further training a general model on specific data to specialize it. ↗ Fine-tuning takes an already-trained model and trains it further on a focused dataset so it performs better on a particular task or in a particular style. Instead of building a model from scratch, you adapt a capable one.
It is useful when you need consistent behavior or domain knowledge that prompting alone cannot reliably produce. It costs more effort than prompting but less than training a new model.
Foundation Model A large, broadly trained model that serves as a base for many uses. ↗ A foundation model is a large model trained on a wide range of data, built to be adapted for many downstream tasks rather than one. The big language models behind today's AI tools are foundation models.
The idea is leverage: one heavily trained base can be fine-tuned, prompted, or connected to tools for countless specific jobs, instead of training a separate model for each.
Few-Shot Learning Guiding a model with a handful of examples inside the prompt. ↗ Few-shot learning means giving the model a few examples of what you want directly in your prompt, so it can follow the pattern. Show it two or three input-output pairs, and it often picks up the format and intent.
It is a simple, powerful prompting technique. When a task is hard to describe in words, demonstrating it with examples frequently works better than explaining it.
Generative AI AI that creates new content — text, images, audio, code — rather than only analyzing. ↗ Generative AI produces original output instead of just classifying or predicting. Give it a prompt and it writes, draws, composes, or codes something new based on patterns it learned during training.
This is the category most people now mean by "AI." It powers the writing assistants, image generators, and coding tools that brought AI into everyday use.
GPT (Generative Pre-trained Transformer) A family of language models built on the transformer architecture. ↗ GPT stands for Generative Pre-trained Transformer — a type of large language model. "Generative" means it creates text, "pre-trained" means it learned from a vast amount of data before you use it, and "transformer" is the underlying architecture.
The name became widely known through popular chatbots, but it describes a technical approach that many models share.
Grounding Tying an AI's answers to real, verifiable sources rather than memory alone. ↗ Grounding means connecting a model's responses to actual sources — documents, databases, or search results — so its answers rest on real information rather than only what it absorbed in training. A grounded answer can point to where it came from.
This is a main defense against confident but wrong responses. It is why systems that retrieve and cite sources tend to be more trustworthy for factual work.
Guardrails Rules and limits that keep an AI system's behavior within safe bounds. ↗ Guardrails are the constraints placed around an AI to stop it from producing harmful, off-topic, or unsafe output. They can filter inputs, restrict what the model is allowed to do, and check responses before they reach you.
As AI systems take more autonomous action, guardrails matter more. They are the difference between a capable tool and one that can act in ways you did not intend.
Hallucination When an AI states something false or invented with apparent confidence. ↗ A hallucination is when a model produces information that sounds plausible but is wrong or made up — a fake citation, an invented statistic, a confident answer to something it does not know. The model is not lying; it is filling gaps with patterns that fit, regardless of truth.
This is a core limitation of current systems. It is why important AI output should be verified, and why grounding answers in real sources helps.
Hyperparameter A setting chosen before training that shapes how a model learns. ↗ A hyperparameter is a configuration value set before training begins, controlling aspects of how the model learns rather than what it learns. Examples include how fast it adjusts and how large its internal structure is.
Choosing good hyperparameters is part skill, part experiment. They can make a meaningful difference in how well the final model performs.
HITL (Human in the Loop) A setup where a person reviews or approves an AI's actions. ↗ Human in the loop means a person stays involved in an AI process — reviewing outputs, approving steps, or correcting mistakes — rather than letting the system run fully on its own. The human acts as a checkpoint.
For consequential decisions, this is a sensible default. It keeps the speed of automation while preserving judgment and accountability where they matter.
Inference The stage where a trained model produces an answer from your input. ↗ Inference is the model at work — taking your prompt and generating a response using what it learned during training. Training is the learning phase; inference is the using phase.
Every time you get an answer from an AI tool, that is inference. It is also where the ongoing cost lives for AI services, since each response takes computing power to produce.
Instruction Tuning Training a model to follow instructions phrased in natural language. ↗ Instruction tuning is a training step that teaches a model to understand and follow directions written the way people naturally phrase them. It is part of what turns a raw text-predictor into something that responds helpfully to requests.
It is a major reason modern assistants feel cooperative — they were specifically trained on examples of instructions paired with good responses.
Jailbreak A prompt crafted to bypass an AI's safety restrictions. ↗ A jailbreak is an attempt to trick a model into ignoring its guardrails and producing content it is meant to refuse. It usually works by disguising the request or framing it in a way the safety training did not anticipate.
Understanding jailbreaks matters for anyone deploying AI, because it shows that safety measures can be probed and worked around. Defending against them is an ongoing part of building responsible systems.
Knowledge Cutoff The date after which a model has no built-in knowledge of events. ↗ A knowledge cutoff is the point in time where a model's training data ends. The model knows nothing about events after that date unless you provide the information or connect it to live sources.
This is why an AI may be unaware of recent news. Always consider the cutoff when asking about current events, and supply up-to-date information when accuracy on recent matters counts.
Knowledge Graph A structured map of facts and how they relate to each other. ↗ A knowledge graph organizes information as connected entities — people, places, concepts — and the relationships between them. Rather than loose text, it is a network where the links carry meaning.
It gives AI a structured way to reason about facts and connections, and it is sometimes combined with language models to make their answers more precise and traceable.
LLM (Large Language Model) A model trained on vast text that generates and understands language. ↗ A large language model is an AI trained on enormous amounts of text to predict and produce language. From that training it picks up grammar, facts, reasoning patterns, and style, which lets it write, answer, summarize, and converse.
LLMs are the engine behind most of today's AI assistants. "Large" refers to both the data they learned from and the number of internal parameters they hold.
Latency The delay between sending a request and receiving the AI's response. ↗ Latency is how long you wait for an answer after submitting a prompt. Lower latency feels snappier; higher latency means more waiting. It depends on the model's size, the length of the response, and the system serving it.
For interactive tools and real-time uses, latency is a real design concern. A more capable but slower model is not always the better choice for the job.
LoRA (Low-Rank Adaptation) An efficient way to fine-tune a model by adjusting a small set of values. ↗ LoRA is a method for fine-tuning a large model without retraining all of it. Instead of changing every parameter, it adds and trains a small number of new values, capturing the adaptation cheaply.
The benefit is that you can specialize a big model with far less computing power and storage, which makes custom AI more accessible to smaller teams.
Machine Learning Teaching computers to learn patterns from data instead of explicit rules. ↗ Machine learning is the approach of building systems that improve at a task by learning from examples, rather than being programmed with fixed rules for every case. You give it data, it finds patterns, and it applies them to new inputs.
It is the foundation under most modern AI. Deep learning and large language models are branches of it.
MCP (Model Context Protocol) An open standard for connecting AI models to tools and data sources. ↗ The Model Context Protocol is a shared standard for letting AI systems connect to external tools, data, and services in a consistent way. Instead of building a custom integration for every connection, developers can use a common method.
It matters because it makes it easier to give models safe, structured access to the information and actions they need — a building block for more capable, connected AI.
Mixture of Experts (MoE) An architecture that activates only the relevant parts of a model per input. ↗ Mixture of Experts is a model design that splits the network into specialized sub-sections, or "experts," and routes each input only to the ones that apply. Rather than running the whole model every time, it uses the relevant slice.
This lets builders create very large, capable models that stay efficient to run, because only part of the model fires for any given request. It is a key technique behind several leading systems.
Multimodal AI AI that works across more than one type of input, such as text and images. ↗ Multimodal AI can handle multiple kinds of data together — reading text, interpreting images, sometimes processing audio or video — within one system. You might show it a picture and ask a question about it in words.
This broadens what AI can do, moving from text-only tools toward systems that perceive and respond across the formats people actually use.
Model Collapse Quality decay when models are trained too much on AI-generated data. ↗ Model collapse is a degradation that can happen when AI models are trained heavily on content produced by other AI rather than original human data. Over successive rounds, the output loses diversity and drifts away from reality.
It is a real concern as AI-generated material fills the web. It underlines why quality, human-grounded data remains valuable for training reliable systems.
Neural Network A model loosely inspired by the brain, built from layers of connected nodes. ↗ A neural network is a system of interconnected nodes, organized in layers, that passes signals forward and adjusts its connections as it learns. It is loosely inspired by how neurons connect in the brain, though it is a mathematical structure, not a biological one.
Neural networks are the core building block of deep learning and the large models behind modern AI.
NLP (Natural Language Processing) The field of getting computers to understand and produce human language. ↗ Natural language processing is the area of AI focused on language — helping computers read, interpret, and generate human text and speech. It covers tasks like translation, summarization, sentiment analysis, and question answering.
Large language models are the latest and most capable products of this field, but NLP as a discipline is decades old.
Open Source (Open Weights) AI models whose internals are publicly available to use and build on. ↗ Open-source or open-weight AI refers to models whose trained parameters are released publicly, so anyone can run, study, or adapt them. This contrasts with closed models accessed only through a provider's service.
Open models offer transparency, control, and the ability to run locally. Closed models often lead in raw capability. Which fits depends on your needs around cost, privacy, and customization.
Overfitting When a model memorizes its training data and fails on new examples. ↗ Overfitting happens when a model learns its training data too closely — including the noise and quirks — and then performs poorly on data it has not seen. It has memorized rather than generalized.
It is a common pitfall in machine learning. The goal is a model that captures the underlying pattern, not one that aces the practice set but stumbles in the real world.
Prompt The input or instruction you give an AI to get a response. ↗ A prompt is what you type or feed into an AI to direct it — a question, a command, a description, or a block of context. The model's response depends heavily on how the prompt is worded and what it includes.
Because of this, writing effective prompts is a practical skill. Clear, specific prompts with the right context reliably produce better results than vague ones.
Prompt Engineering The craft of writing prompts that get reliable, high-quality results. ↗ Prompt engineering is the practice of designing prompts deliberately to draw out the best response from a model. It includes being specific, giving examples, setting a role or format, and supplying the context the model needs.
It is less about secret tricks and more about clear communication. The better you specify what you want, the better the output tends to be.
Parameters The internal values a model adjusts during training to store what it learns. ↗ Parameters are the adjustable values inside a model that get tuned during training. They are where the model's learned knowledge effectively lives. Modern large models hold billions of them.
Parameter count is often cited as a rough measure of a model's scale, though more parameters does not automatically mean better — design and training quality matter just as much.
Prompt Injection An attack that hides malicious instructions inside content an AI reads. ↗ Prompt injection is a security risk where harmful instructions are slipped into data the AI processes — a web page, a document, an email — tricking the model into following them. The attack rides in on content the system was asked to read.
It matters especially for AI agents that act on outside information. Defending against it is an active area of AI security work.
Quantization Shrinking a model by storing its numbers at lower precision. ↗ Quantization reduces a model's size by representing its internal values with less numerical precision — fewer bits per number. The model gets smaller and faster to run, with usually only a modest loss in quality.
It is a key technique for running capable models on limited hardware, such as a laptop or phone, where the full-precision version would be too large.
RAG (Retrieval-Augmented Generation) Letting an AI look up information before answering instead of relying on memory. ↗ Retrieval-augmented generation connects a model to an external source of information — documents, a database, search results — so it retrieves relevant material first, then uses it to answer. Without it, the model can only draw on its training; with it, the model can use current or private information it was never trained on.
RAG is the main way organizations give AI access to their own knowledge without retraining a model. It also reduces invented answers by grounding responses in real, retrieved content.
Reasoning Model A model designed to work through problems in steps before answering. ↗ A reasoning model is built to think through a problem in stages — breaking it down, working intermediate steps, checking itself — rather than producing an immediate answer. This step-by-step process improves performance on harder tasks like math and logic.
The trade-off is time and cost: more thinking means slower, more expensive responses. For complex problems that is often worth it; for simple ones it is overkill.
RLHF (Reinforcement Learning from Human Feedback) Training a model to be more helpful using human ratings of its answers. ↗ RLHF is a training method where people rate a model's responses, and those ratings are used to steer the model toward outputs humans prefer. It is a big part of why modern assistants feel helpful and aligned with what users want.
It shapes tone and behavior, not just raw knowledge. The human judgments guide the model toward being useful, honest, and safe in its replies.
SLM (Small Language Model) A compact language model built to run efficiently, often locally. ↗ A small language model is a more compact alternative to a large one, designed to run efficiently — sometimes on a single laptop or even a phone. It trades some raw capability for speed, lower cost, and the ability to run privately on local hardware.
For focused tasks, a well-chosen small model can be more practical than a giant one. The principle of keeping data on your own device also strengthens privacy.
Semantic Search Searching by meaning rather than exact keyword matches. ↗ Semantic search finds results based on what you mean, not just the literal words you typed. It uses embeddings to compare the meaning of your query against the meaning of stored content, so it can surface relevant results that share no exact keywords.
It is what makes AI-powered search feel like it understands intent, and it underpins many retrieval systems behind modern AI tools.
Synthetic Data Artificially generated data used to train or test models. ↗ Synthetic data is information created by a computer rather than collected from the real world, used to train or evaluate AI when real data is scarce, sensitive, or expensive. It can fill gaps and protect privacy.
It is useful but must be handled carefully — over-reliance on AI-generated data can degrade model quality over time, a problem known as model collapse.
Shadow AI Employees using AI tools without official approval or oversight. ↗ Shadow AI describes people in an organization using AI tools on their own, outside any sanctioned process or policy. It often happens because the tools are genuinely useful and easy to access.
The concern is risk: sensitive data may go into unvetted services, and there is no oversight of accuracy or security. It is a governance challenge organizations increasingly have to address.
Token The small chunk of text a model reads and generates, like a word piece. ↗ A token is the unit of text a model processes — roughly a word or part of a word. Models break your input into tokens, work with those, and produce output token by token. A sentence might be a dozen tokens.
Tokens matter practically because context limits and pricing are measured in them. Longer inputs and outputs use more tokens, which affects both capacity and cost.
Training The process of teaching a model by exposing it to large amounts of data. ↗ Training is how a model learns. It processes huge amounts of data, adjusting its internal parameters to capture patterns, until it can perform its task. This is the resource-heavy phase that happens before you ever use the model.
Training a large model from scratch takes enormous data and computing power, which is why most applications adapt an existing model rather than build one.
Transformer The neural network architecture behind most modern language models. ↗ The transformer is the architecture that made today's language models possible. Its key innovation, the attention mechanism, lets the model weigh the relationships between all parts of the input at once, handling long and complex text far better than earlier designs.
Nearly every major language model is built on it. Its introduction was a turning point that set off the current wave of AI progress.
Temperature A setting that controls how random or focused a model's output is. ↗ Temperature is a dial that adjusts how varied a model's responses are. Lower temperature makes output more focused and predictable; higher temperature makes it more diverse and creative, at the risk of wandering.
It is a practical control: factual, consistent tasks suit a low setting, while brainstorming or creative writing can benefit from a higher one.
Unsupervised Learning Learning patterns from data that has no labels or answers attached. ↗ Unsupervised learning is a method where a model finds structure in data on its own, without being given labeled examples of the right answer. It looks for patterns, groupings, and relationships that are already present.
It is useful for discovering hidden structure — clustering similar items, for instance — and much of how large language models learn from raw text falls under this broad idea.
Vector Database A database built to store and search embeddings by similarity. ↗ A vector database stores embeddings — the numeric representations of meaning — and is optimized to quickly find the ones most similar to a given query. It is what lets a system search millions of items by meaning rather than exact match.
Vector databases are central to retrieval systems and RAG, giving AI a fast way to pull the most relevant information before generating an answer.
Vibe Coding Building software by describing what you want and letting AI write the code. ↗ Vibe coding is an informal term for creating software mainly by telling an AI what you want in plain language and letting it generate the code, rather than writing every line yourself. You steer by intent and refine through conversation.
It lowers the barrier to building things, though the results still need review and testing. It reflects how AI is shifting the role of the person from typing code to directing it.
Weights The learned values inside a model that determine how it responds. ↗ Weights are the trained numerical values within a neural network that encode what the model has learned. They are essentially the same idea as parameters — the internal settings that, taken together, produce the model's behavior.
When a model is released as "open weights," it means these values are made public so others can run and build on the model directly.
Explainable AI (XAI) Methods that make an AI's decisions understandable to people. ↗ Explainable AI refers to techniques that help people understand why a model produced a given output. Many advanced models are effectively black boxes, and XAI aims to open them up — showing what factors drove a decision.
It matters most where trust and accountability are essential, such as healthcare, finance, or legal contexts, where "the AI said so" is not a good enough reason on its own.
YAML / Config Human-readable files used to set up and control AI tools and workflows. ↗ Configuration files, often written in a readable format like YAML, are how many AI tools and systems are set up — defining settings, connections, and behavior without changing the underlying code. They let you adjust how a system runs in a clear, editable way.
For anyone working hands-on with AI tools, basic comfort reading these files goes a long way, since so much setup happens through them.
Zero-Shot Learning Asking a model to do a task with no examples, just a clear instruction. ↗ Zero-shot learning means giving a model a task it was not specifically trained for and no examples — just a description — and having it perform anyway, drawing on its broad training. "Write a haiku about rain" with no samples is a zero-shot request.
The ability of large models to handle zero-shot tasks reasonably well is part of what makes them so flexible and easy to use.
Go Deeper
Help Center
Knowing the words is the start. Our Help Center walks through where to begin, how to write prompts that work, common tasks, and how to use these tools safely — organized so you can find an answer fast.
Visit the Help Center →
Explore
Cybnex Hub Home
Head back to the Hub for the latest guidance, articles, and resources. The glossary gives you the vocabulary; the Hub puts it to work with practical, real-world guidance you can apply.
Go to Hub Home →
