Generative AI

Generative AI focuses on producing content based on patterns learned from training data. A model can generate a support response, summarize an incident report, create an image from a description, or suggest code. Its output is conditioned by inputs such as instructions, examples, documents, images, or conversation history. Generation does not guarantee originality, factual accuracy, or suitability for a particular task.

Artificial intelligence is the broad field of building systems that perform tasks associated with intelligence, including reasoning, perception, planning, and language processing. Machine learning is a subset of AI in which systems learn patterns from data instead of relying exclusively on explicitly programmed rules. Deep learning is a subset of machine learning that uses neural networks with multiple layers. Modern generative AI commonly uses deep learning, but these terms are not interchangeable: a deep neural network that classifies images is not necessarily generative, and AI also includes rule-based systems that do not learn from data.

Generative and predictive applications often share training infrastructure, datasets, evaluation processes, and deployment patterns. Their primary objectives differ. A classifier might label a support ticket as a billing issue, while a generative model might draft a response to that ticket. A numerical forecasting model can predict demand without generating language or images. A generative model can also perform classification when prompted, so the task and system behavior matter more than a simplistic distinction between model names.

Foundation models are trained on broad datasets and can be adapted to multiple downstream tasks. Large language models are models focused on language-related tasks; they are not synonyms for all generative AI. Image generation models and audio generation models also belong to this field. Prompting supplies instructions and context during inference. Fine-tuning updates model parameters using additional training data, whereas retrieval-augmented generation retrieves relevant information and includes it in the model's context without inherently changing its parameters.

In the cloud, engineers can invoke managed model APIs or deploy models on managed machine learning infrastructure. For an internal documentation assistant, an application can authenticate the user, retrieve documents that the user is authorized to access, and send relevant passages with the question to a model. The application can return an answer with references and decline requests when supporting evidence is insufficient. Retrieval can improve grounding, but it does not eliminate hallucinations or replace access controls.

Agentic AI describes systems that pursue goals through steps such as planning, selecting tools, taking actions, and evaluating results. Many agents use generative models to interpret requests and select actions, but generation alone does not make a system agentic. Drafting a refund explanation is generative behavior; checking an order through an API and initiating an approved refund adds agentic behavior. Tool permissions, approval gates, execution limits, and audit logs become especially important when outputs can cause real changes.

Engineering trade-offs include output quality, latency, context limits, data protection, and cost. Larger models or longer prompts can increase resource consumption without reliably improving every task. Lower generation temperature generally reduces variation but does not guarantee truth or identical results. Evaluate representative inputs for factuality, task completion, harmful content, and security failures. Treat retrieved documents and user inputs as untrusted, test for prompt injection, validate generated code and structured outputs, and require human review for consequential decisions.