LEARN
Learn
Explainers, primers and context for the AI topics in the news: how models are built, trained, run and measured, written for readers who want the mechanics behind the headlines.
Learn
All categories
Foundations
4- ComparisonTraining vs. Inference↗Training changes an AI model's learned settings so it can perform a task better; inference uses a selected state of that model to produce an output without changing those learned settings. Training creates or adapts the model, while inference puts that model to work.
- ConceptWhat Is Artificial Intelligence?↗Artificial intelligence (AI) is the field of designing computer systems that produce goal-directed outputs—such as predictions, content, recommendations, or decisions—using methods associated with abilities such as learning, reasoning, perception, and language. AI also refers to the systems created by that field.
- ConceptWhat Is Generative AI?↗Generative AI is artificial intelligence that learns patterns in existing data and uses those patterns to produce new content, such as text, images, audio, video, or code. It usually generates that content in response to an instruction or other input.
- ConceptWhat Is Machine Learning?↗Machine learning is a way to build software that learns patterns from examples or feedback and uses those patterns to make useful predictions, decisions, or outputs on cases it has not seen before. It is one approach within artificial intelligence, not a synonym for all AI.
Tokens, Context & Representation
8- ComparisonContext Window vs. Knowledge Cutoff↗A context window limits how much information an AI model can use in one request, while a knowledge cutoff describes how recent its built-in knowledge is. One is a capacity measured in tokens; the other is a date associated with training.
- ComparisonInput vs. Output vs. Reasoning Tokens↗Input tokens are what a model receives, output tokens are what it generates, and reasoning tokens are internal generated tokens used to work toward an answer. The catch is that reasoning tokens are often part of output usage even when an API displays them as a separate detail.
- ConceptWhat Are Input Tokens?↗Input tokens are the tokens a model receives and processes before it starts generating the next response. They come from the complete model-facing request, not only the latest words you typed.
- ConceptWhat Are Output Tokens?↗Output tokens are the units a model generates as it produces a response. They usually make up the text, code, structured data, or tool instructions you receive, although a provider may also count hidden reasoning as output usage.
- ConceptWhat Are Reasoning Tokens?↗Reasoning tokens are tokens an AI model generates while working toward an answer, separate from the answer text you see. They represent extra generated steps such as planning, checking, or trying an approach, and they can count toward usage and limits even when the service hides them.
- ConceptWhat Is a Context Window?↗A context window is the maximum amount of tokenized material an AI model can use while producing one response. In many language-model APIs, that budget covers the input and the generated output, although providers may also impose separate input or output limits.
- ConceptWhat Is a Token in AI?↗A token in AI is one item from the vocabulary an LLM uses to represent and generate text. Depending on the model's tokenizer, a token can correspond to a word, part of a word, punctuation, whitespace, or even part of an encoded character.
- ConceptWhat Is an AI Knowledge Cutoff?↗An AI knowledge cutoff is the date up to which a model's built-in knowledge is expected to reflect its training data. It is a rough boundary, not a guarantee that the model knows every earlier fact or that an AI product cannot fetch later information.
Models & Architecture
7- ComparisonModel Parameters vs. Model Weights↗Model parameters are all the fitted values that determine a model’s behavior; model weights are either one type of parameter or an informal name for the model’s learned values as a whole. In a layer equation, “weight” usually has the narrower meaning. In a phrase such as “download the model weights,” it usually has the broader one.
- ConceptWhat Are Model Parameters?↗Model parameters are the stored numerical values inside a fitted model that determine how it transforms an input into an output. In a neural network, they usually include weights and biases whose values are chosen during training and reused when the model makes predictions.
- ConceptWhat Are Model Weights?↗Model weights are learned numbers that determine how strongly a model uses each input or intermediate signal when computing an output. They are usually stored in large groups called matrices or tensors, and training changes them so the model's outputs better fit its objective.
- ConceptWhat Is a Foundation Model?↗A foundation model is a machine-learning model trained on broad data so it can be adapted for many different tasks, rather than built for just one. It is a reusable starting point, not a finished application.
- ConceptWhat Is a Neural Network?↗A neural network is a machine-learning model that learns to turn input numbers into useful output numbers by passing them through connected layers of simple calculations. Those layers form one trainable mathematical function, not a literal electronic brain.
- ConceptWhat Is an AI Model?↗An AI model is the part of an AI system that transforms an input into an output, such as a prediction, classification, recommendation, action, or generated result. In most current usage, it means a machine-learning model whose behavior was shaped by training data.
- ConceptWhat Is an LLM?↗A large language model (LLM) is an AI model trained on large collections of language data to process or generate likely sequences of text. Most LLMs used for text generation take a prompt and produce a response one small piece at a time.
Training & Adaptation
1Inference & Serving
6- ComparisonLatency vs. TTFT vs. Tokens per Second vs. Throughput↗Latency is how long one inference request takes, TTFT is how long a streamed request takes to begin answering, tokens per second is a rate that may describe one stream or a whole system, and throughput is how much total work a system completes over time. The sharpest distinction is scope: latency and TTFT time a request, while throughput counts work across a measurement window.
- ConceptWhat Are Tokens per Second?↗Tokens per second (TPS or tok/s) measures how many tokens an AI system produces in one second. The number may describe the output speed of one response or the combined throughput of an entire serving system, so the label alone is incomplete.
- ConceptWhat Is Inference Latency?↗Inference latency is the elapsed time between an AI system receiving an input and producing a specified output. A latency number is meaningful only when you know where the timer starts, where it stops, and what kind of output counts as complete.
- ConceptWhat Is Inference Throughput?↗Inference throughput is the amount of model work an inference system completes per unit of time across all the requests it is serving. It may be measured in completed requests, inferences, samples, or tokens per second.
- ConceptWhat Is Inference in AI?↗Inference in AI is the process of running a trained model on an input to produce an output, without updating what the model learned. It is the use phase of a model: a photo goes in and a label comes out, or a prompt goes in and generated text comes out.
- ConceptWhat Is Time to First Token?↗Time to first token, or TTFT, is the elapsed time between starting a language-model request and getting the first token of its response. It measures the initial wait, not how fast the rest of the response arrives.
AI Systems, Harnesses & Agents
5- ComparisonAI Model vs. Chatbot vs. Harness vs. Agent↗An AI model is the component that produces outputs; a chatbot is a conversational application; a harness is the software around a model that manages context and execution; and an agent is a goal-directed system that uses those pieces to choose and carry out steps. They are overlapping layers, not four competing kinds of AI.
- ConceptWhat Is Tool Use in LLMs?↗Tool use in LLMs is a mechanism that lets a language model request an external operation—such as searching a database, running a calculation, or sending a message—and use the returned result in its response. The model usually chooses the tool and supplies proposed inputs; software around the model decides whether to execute the call.
- ConceptWhat Is an AI Agent?↗An AI agent is a software system that pursues a goal by using an AI model to choose its next actions, observe the results, and continue until it finishes, needs help, or reaches a limit. Unlike a model that only produces an answer, an agent can control a multi-step process and act through tools within set permissions.
- ConceptWhat Is an AI Chatbot?↗An AI chatbot is a software application that uses artificial intelligence to hold a conversation with a person through text or voice. It receives messages, uses an AI model to interpret or answer them, and manages the context, rules, data, and actions needed to continue the exchange.
- ConceptWhat Is an AI Harness?↗An AI harness is the software around an AI model that controls what the model receives, what happens to its output, and how the model connects to the rest of an application. It can be as small as code that manages a chat history or as large as a runtime that manages tools, memory, permissions, retries, and long-running work.