What the model receives, learns, and produces.
Artificial intelligenceA broad family of methods and systems that use inputs to generate outputs for tasks such as recognition, prediction, language, recommendations, and action.What artificial intelligence means ↗
AI system and modelA model is one component in an AI system, which can also include data, software, hardware, tools, people, and operating procedures.What artificial intelligence means ↗Intelligence as a relationship ↗
Artificial general intelligence (AGI)A research term for broad, adaptable capability across tasks. Breadth, performance, and autonomy need separate definitions and evidence.What researchers mean by AGI ↗
Machine learningA method for fitting models from examples and feedback so they can make predictions or produce other outputs for new inputs.How models learn from data ↗
AI for scientific discoveryMachine-learning methods can forecast physical systems, infer structures, or search candidate spaces. Each output is a hypothesis whose value depends on independent measurements and experiments.From AI prediction to scientific evidence ↗How AI models predict the weather ↗
Deep learningA machine-learning family that uses multiple composed layers to learn representations from data. Depth alone does not guarantee accuracy.How layered neural networks learn ↗
Reinforcement learningA learner improves a policy through actions and feedback in an environment. The reward is a designed signal, so a high score needs to be checked against the real goal.How learning from feedback works ↗
Large language modelAn autoregressive language model produces a sequence by repeatedly estimating which token may come next from the context it receives.How language models generate text ↗
Foundation modelA model trained on broad data and adaptable to many downstream tasks. It can support many applications, but its name alone does not establish capability, reliability, or legal classification.Understand the shared base and the system around it ↗
Small language modelA compact model chosen for a narrower compute and memory budget. Its quality, response time, and privacy still depend on the task, runtime, device, and surrounding application.Compare compact models and device limits ↗
Transformer and attentionA neural-network architecture that processes sequences through stacked attention and feed-forward layers. Attention combines learned representations; it is not human focus.How Transformers process sequences ↗
WeightsWeights are learned numerical parameters inside a model. Making them downloadable does not by itself make the full system open source.Open source vs. open weights ↗
TokensA tokenizer converts text into model-specific token IDs. Tokens may be words, parts of words, spaces, punctuation, or other units.How text becomes token IDs ↗
Context windowThe finite input and output span available to a model call. It is not a promise that every detail will be used or remembered later.How token budgets and context work ↗Test what a model can use at long lengths ↗Shape context for an agent’s next decision ↗Design memory across tasks ↗
InferenceInference is using a trained model to produce an output. Prompt processing, token generation, memory, and the serving system shape how that output is delivered.What happens when a model answers ↗Measure the whole system cost ↗
KV cacheTemporary key and value tensors preserve attention work across generated tokens. Their memory use depends on sequence length, attention layout, data type, and concurrent requests.Estimate the memory behind token generation ↗
Prompt cachingA serving system can reuse computation for an identical prompt prefix across eligible requests. It does not store a person’s conversation history or skip processing the changing part of a request.How prompt-prefix reuse works ↗
Speculative decodingA draft process proposes several next tokens for a target model to verify together. Real speed depends on draft cost, accepted tokens, verification, and workload.Understand draft-and-verify generation ↗
Reasoning modelA model designed to spend additional computation on harder tasks at inference time. The method varies; more computation can raise latency and cost without guaranteeing a correct answer.How reasoning models use extra compute ↗
Local inferenceA model generates on a device or machine under your control. The app, its history, its network connections, and its tools still determine where other data can go.Map the whole data path ↗
Image generationText-to-image systems turn learned visual patterns into new images. Many use diffusion, while other architectures generate image tokens in sequence.How image generators work ↗
Multimodal AIA system that works with more than one kind of information, such as text, images, or audio. Input and output capabilities can differ by model.How multimodal systems connect inputs ↗
Voice AIA spoken interface may connect audio capture, speech recognition, dialogue, tools, and generated speech. Each handoff affects quality, timing, and data handling.How voice AI moves from sound to response ↗
Music and sound generationGenerators condition learned audio representations on descriptions, melodies, or reference recordings. Listen for structure, timing, and signal artifacts across the complete result.How music generators turn prompts into sound ↗
Video understandingVideo systems connect sampled frames, motion, sound, and timestamps to answer task-specific questions about a clip.How video models connect frames over time ↗
Video generationText-to-video and image-to-video models create a new sequence. Review the whole clip for motion, continuity, and how well it follows the brief.How video generators create moving sequences ↗
Computer visionMethods for classifying images, locating objects, reading visible text, or marking regions. Each output needs a task-specific evaluation.How AI works with images ↗