Overviews

AI

AI, or Artificial Intelligence, is the broad field of building systems that perform tasks associated with perception, language, learning, reasoning, planning, generation, or decision-making. Modern AI products are complete sociotechnical systems: models are only one component alongside data, software, infrastructure, evaluation, governance, and people.

AI Machine Learning Generative AI AGI

The landscape

AI is broader than machine learning

Perception

Computer vision, speech, audio, and sensor interpretation turn raw inputs into useful representations.

Language

NLP and language models analyze, transform, retrieve, and generate human language.

Learning

Machine learning derives behavior from examples, feedback, objectives, and data rather than only explicit rules.

Reasoning and planning

Systems search, plan, use tools, optimize objectives, or combine symbolic and learned methods.

Generation

Generative AI produces text, images, audio, video, code, and structured outputs.

Action

Agents select and execute actions in software or physical environments while observing results.

Narrow AI is built for bounded tasks. AGI is a hypothetical system with broad, transferable capability across many domains. Most deployed systems today are narrow systems, even when they expose broad interfaces.

Models and learning

Different techniques solve different problems

Classical machine learning

Supervised, unsupervised, and reinforcement learning support prediction, classification, clustering, ranking, control, and optimization.

Deep learning

Neural networks learn layered representations. Transformers use attention and underpin many modern language and multimodal models.

Foundation models

Large pretrained models can be adapted through prompting, retrieval, fine-tuning, or task-specific components.

Supervised Learning Unsupervised Learning Reinforcement Learning Transformer Foundation Model

AI systems

A model is not a product

A production AI system may include a model gateway, prompt templates, retrieval, vector or conventional search, tools, policy checks, memory, orchestration, observability, feedback, and human review.

RAG

Retrieval-Augmented Generation supplies relevant external information at request time, improving grounding and freshness when retrieval is effective.

Agents

AI agents observe context, choose actions, invoke tools, and iterate toward a goal. Autonomy increases both capability and operational risk.

Multimodal systems

Multimodal AI combines text, image, audio, video, or sensor data in one workflow or model.

RAG AI Agent Multimodal AI LLM

Engineering lifecycle

Build evidence from problem framing through operations

01
Problem framing
Define the user need, decision, acceptable failure modes, and whether AI is justified.
02
Data
Collect, label, govern, protect, and evaluate representative data.
03
Model and system design
Select models, retrieval, tools, prompts, workflows, and surrounding software architecture.
04
Evaluation
Measure quality, safety, robustness, latency, cost, bias, and task-specific outcomes.
05
Deployment
Package the model and supporting system with versioning, access controls, rollout, and rollback.
06
Monitoring
Watch quality, drift, abuse, cost, latency, incidents, and changes in user behavior.

Evaluation

Quality is multidimensional and task-specific

Task quality

Accuracy, precision, recall, ranking quality, groundedness, usefulness, or human preference.

Robustness

Behavior under unusual inputs, distribution shifts, adversarial use, and dependency failures.

Operations

Latency, throughput, availability, cost, token or compute use, and capacity.

Impact

Fairness, privacy, security, accessibility, explainability, and downstream consequences.

Use offline benchmarks, curated test sets, red-team cases, human evaluation, controlled experiments, and production monitoring. A single aggregate benchmark is rarely sufficient.

Connecting AI to tools and data

MCP standardizes external capabilities for AI applications

MCP—Model Context Protocol—is an open standard for connecting AI applications to external systems. Instead of building a different proprietary adapter for every model, tool, repository, database, or workflow, an application can use a common protocol to discover and invoke capabilities.

Hosts and clients

An MCP host is the AI application. It manages user consent, policy, model interaction, and one or more MCP clients. A client communicates with a specific server and negotiates the capabilities that can be used.

Servers

An MCP server exposes a bounded integration to a local or remote system. Servers can provide tools for actions, resources for contextual data, and prompts or other protocol capabilities supported by the active specification.

Standard exchange

MCP defines structured messages and capability discovery so clients and servers can interoperate. Implementations still need authentication, authorization, validation, timeouts, observability, and safe handling of untrusted content.

Core primitives

Tools

Executable operations such as searching a catalog, creating a ticket, querying a database, or starting a workflow. Tool descriptions and schemas help the model or application choose and call them correctly.

Resources

Addressable context such as files, records, documentation, repository content, or application state that a client can read and provide to the model.

Prompts

Reusable interaction templates exposed by a server. They can package domain instructions, expected inputs, and workflow context without embedding every prompt in the host application.

Architecture and security boundaries

MCP commonly follows a host–client–server architecture. The host remains responsible for user experience, permission decisions, context aggregation, and model use. Each server should expose the smallest practical capability surface. Treat tool outputs, resource content, and server metadata as untrusted input.

Require explicit authorization for sensitive data and consequential actions.
Apply least privilege to credentials, scopes, files, databases, and network access.
Validate tool arguments and results instead of trusting model-generated calls.
Separate read-only context access from state-changing operations.
Record tool execution, failures, latency, and user approval decisions.
Version integrations and follow the current MCP specification because protocol capabilities evolve.

Where MCP fits

MCP is an integration protocol, not an AI model, agent framework, or guarantee of safe autonomy. It complements tool calling, RAG, agents, APIs, identity, and observability by standardizing how an AI application reaches external capabilities.

MCP Tool Calling AI Agent RAG Official MCP documentation

Safety and governance

Capability must be paired with controls

AI Safety

Reduce harmful failures through threat modeling, evaluation, safeguards, monitoring, incident response, and bounded autonomy.

Responsible AI

Establish accountability, transparency, fairness, privacy, compliance, documentation, and human oversight.

AI Alignment

Study whether system behavior reliably follows intended goals and constraints, including under novel or strategic conditions.

AI Safety Responsible AI AI Alignment

A useful boundary concept

AGI is hypothetical, not a synonym for a capable chatbot

Artificial General Intelligence generally refers to broad, adaptable cognitive capability across many domains rather than competence on a bounded task. There is no single agreed test or engineering definition. Strong language fluency, benchmark performance, tool use, or agentic behavior may be relevant evidence, but none alone establishes AGI.

For practical engineering, distinguish current system capabilities from speculative labels. Document what the system can do, where it fails, what resources and supervision it needs, and how its behavior is evaluated.