A trained model is one component. Development, evaluation, inference, integration, operation, monitoring, and human use create the behavior and consequences people encounter.
Plan for this page
Understand the idea, test it actively, then explain it in your own words.
You will leave with
distinguish training from inference and adaptation separate a model from the product and institution around it locate lifecycle controls before, during, and after deployment
Time
25 min
Before you begin
AI from Zero is useful preparation, but it is not a locked prerequisite. Each unit states the concepts it uses.
Do this now
Read the outcomes and useful terms before the first section.
distinguish training from inference and adaptation
separate a model from the product and institution around it
locate lifecycle controls before, during, and after deployment
Terms you will use
These definitions prepare you for the reading; you do not need to memorize them.Model
A parameterized computational structure that transforms inputs into outputs according to patterns shaped during development.
Training
The process of adjusting model parameters toward an objective using examples, feedback, optimization, and evaluation.
Inference
Running a trained model on a new input to produce a prediction, score, representation, or generated output.
01
Building and running are different operations
Training changes parameters. It may involve pretraining, fine-tuning, human or automated feedback, evaluation, and repeated design decisions. Inference uses the resulting model for a new input; it does not normally retrain the entire model from each conversation.
Some products adapt through retrieval, temporary context, user settings, memory stores, tool results, or later model updates. These mechanisms can change output without meaning that the base model learned permanently from that exact interaction.
02
The product is larger than the model
A deployed service may add input filtering, system instructions, document retrieval, tool calls, databases, business rules, ranking, safety controls, user interfaces, logging, human review, and fallback behavior.
When two products use related models but behave differently, the surrounding system may explain much of the difference. Model documentation alone cannot establish the actual permissions, data practices, or workflow of a product.
03
Operation continues after launch
Real systems face changing inputs, model updates, adversarial behavior, new uses, incidents, and organizational pressure. Monitoring should connect technical signals with user outcomes and routes for correction or recovery.
Lifecycle thinking asks who can change the system, approve a deployment, observe failures, stop an action, investigate an incident, notify affected people, and repair harm. These responsibilities cannot be delegated to model parameters.
Interactive model
Five layers from model to consequence
Selected layerDevelopment
Task framing, data, model choice, training, evaluation, documentation, and release decision.
Read the complete text alternative
DevelopmentTask framing, data, model choice, training, evaluation, documentation, and release decision.
InferenceNew input, context assembly, model execution, decoding or thresholding, and raw output.
ApplicationInterface, retrieval, tools, business rules, permissions, storage, and user experience.
OperationMonitoring, updates, incidents, abuse response, human review, rollback, and support.
InstitutionPurpose, accountability, incentives, law, labor, affected people, and remedy.
Misconception to correct
“The chatbot remembers everything because the model is retrained after each message.”
A product may send conversation context or retrieve stored memory during inference. That is distinct from retraining the base model, and retention practices must be checked for the specific service.
Active checks
Decide, then compare the reasoning
No grade, score, or streak. Progress records completed learning objects, not your worth or ability.01
You upload a document to an existing service and request a summary. What is most directly happening?
02
Two assistants using the same model family produce different results. Which explanation is plausible?
Explain back and transfer
Draw the lifecycle of one AI feature
Choose a feature and place what you know into development, inference, application, operation, and institution layers. Mark controls and unknowns.
Answer both checks before completing.
Source context
What supports this unit and where it stops
Evidence status: established
Machine Learning GlossaryGoogle for Developers
Reference vocabulary for model, training, inference, evaluation and related concepts; provider terminology may not cover every architecture or discipline.
inAi