GOVIBETOOL KNOWLEDGE

Explore Artificial Intelligence

Explore modern artificial intelligence through free, clear, and accessible guides that explain complex AI concepts in simple, practical language.

Learning paths

From LLMs to AI agentsThe full build-up: how language models work, how to steer them, how they use tools, and how agents are assembled.
  1. What is a Large Language Model (LLM)?
  2. What are Tokens in AI?
  3. What is a Context Window?
  4. Prompt Engineering: How to Write Better AI Prompts
  5. Agent Tools: How AI Agents Take Actions
  6. Function Calling (Tool Calling) Explained
  7. What is MCP (Model Context Protocol)?
  8. Agent Memory: Short-Term, Long-Term and Beyond
  9. What is RAG (Retrieval-Augmented Generation)?
  10. What are AI Agents?
  11. Multi-Agent Systems Explained
  12. Agentic Workflows: Patterns for Reliable AI Automation
Retrieval and grounding (RAG)How to make a model answer from your own information, and how to tell whether it worked.
  1. What is RAG (Retrieval-Augmented Generation)?
  2. What are Embeddings in AI?
  3. What is a Vector Database?
  4. Chunking: How to Split Documents for RAG
  5. What is a Context Window?
  6. AI Hallucinations: Why Models Make Things Up
  7. How to Evaluate AI Systems and Outputs
Using AI safelyReliability, privacy and security for anyone putting AI into real work.
  1. AI Hallucinations: Why Models Make Things Up
  2. How to Reduce AI Hallucinations
  3. AI Privacy and Security: What to Know Before You Paste Data
  4. What is Prompt Injection?
  5. What is Local AI? Running Models on Your Own Device
  6. How to Evaluate AI Systems and Outputs
Build with PythonCall models, handle data and assemble a small RAG pipeline in Python.
  1. Python for AI: When and How to Use It
  2. Calling AI and Model APIs with Python
  3. JSON and CSV Handling for AI Work in Python
  4. Building RAG with Python
  5. Python AI Libraries: What to Use When
Web AI with JavaScriptKeep secrets on a Node server, type the client contract, and stream into a React chat UI.
  1. JavaScript for AI Applications
  2. Node.js for AI Applications
  3. TypeScript Types for an API Client
  4. Chatbot State and Conversation UI in React
  5. Streaming AI Responses with Node.js
Reasoning models and test-time computeHow models spend extra effort at answer time, how that is trained and verified, and what to show users.
  1. What is a Large Language Model (LLM)?
  2. Chain-of-Thought Prompting
  3. Reasoning Models
  4. Test-Time Compute (Inference-Time Scaling)
  5. Self-Consistency (Sampling and Voting)
  6. Best-of-N Sampling
  7. Process Reward Model (PRM)
  8. Outcome Reward Model (ORM)
  9. Reinforcement Learning for Reasoning (RLVR and GRPO)
  10. Thinking Budgets and Reasoning Effort
  11. Reasoning Transparency and Hidden Reasoning
Alignment: from RLHF to DPOHow assistants are tuned from human and AI preferences, and where reinforcement learning does and does not apply.
  1. Instruction Tuning
  2. Reinforcement Learning
  3. Proximal Policy Optimization (PPO)
  4. RLHF (Reinforcement Learning from Human Feedback)
  5. Preference Optimization (Aligning Models from Human or AI Preferences)
  6. Direct Preference Optimization (DPO)
  7. Reward Hacking and Over-Optimization
  8. Sycophancy in Language Models
  9. AI Alignment
  10. Red-Teaming AI Systems
Evaluating AI systemsFrom benchmarks and their pitfalls to task evals, judges, RAG and agent testing.
  1. How to Evaluate AI Systems and Outputs
  2. AI Benchmarks and Leaderboards
  3. Benchmark Contamination and Data Leakage
  4. Public Benchmarks vs Task-Specific Evals
  5. LLM-as-a-Judge
  6. RAG Evaluation (Retrieval and Answer Quality)
  7. Agent Evaluation
Structured output and constrained decodingGetting dependable machine-readable output from models.
  1. Sampling and Decoding (Temperature, Top-p, Beam Search)
  2. Structured Outputs from LLMs
  3. JSON Schema
  4. Constrained Decoding (Guided Generation)
  5. Grammar-Guided Generation (CFG and Regex Constraints)
  6. Structured Output Methods Compared
Running and serving modelsQuantisation, runtimes, serving and the hardware beneath them.
  1. Quantization of AI Models
  2. What is Local AI? Running Models on Your Own Device
  3. llama.cpp
  4. vLLM
  5. KV Cache
  6. Model Serving and Inference Infrastructure
  7. GPUs and AI Accelerators
  8. LLM Cost Optimization
Machine learning foundationsLearning paradigms, networks and architectures behind modern AI.
  1. Supervised Learning
  2. Neural Networks
  3. Backpropagation and Gradient Descent
  4. Overfitting and Regularization
  5. Transfer Learning
  6. Convolutional Neural Networks (CNNs)
  7. Vision Transformers (ViT)
  8. Diffusion Models
  9. State Space Models (Mamba and S4)
Governing AIFrameworks, standards and law for responsible AI use.
  1. AI Governance
  2. NIST AI Risk Management Framework
  3. ISO/IEC 42001 (AI Management System)
  4. EU AI Act
  5. Model Cards and Datasheets (AI Documentation)
  6. AI Bias and Fairness
  7. OWASP Top 10 for LLM Applications

Foundations

ConceptWhat is Artificial Intelligence?AI is the broad field of building software that performs tasks we normally associate with human judgement. Machine learning and generative AI are subsets of it.ConceptWhat is Generative AI?Generative AI models create new content — text, images, audio, video and code — by learning the patterns in very large training sets.ConceptWhat is a Large Language Model (LLM)?A large language model is a neural network trained on vast amounts of text to predict the next token, which lets it write, summarise, translate, reason about text and write code.ConceptWhat are Tokens in AI?Tokens are the chunks of text — whole words, word pieces or punctuation — that language models read and write. Limits and pricing are usually measured in them.ConceptWhat is a Context Window?The context window is the maximum amount of text, measured in tokens, that a model can consider in a single request. Anything outside it is invisible to the model.ConceptWhat is a Transformer? (Attention Explained)The transformer is the neural-network design behind most modern language models. Its key idea, attention, lets every token weigh how relevant every other token is.ConceptWhat is Multimodal AI?Multimodal AI models work with more than one type of data — for example text, images, audio and video — in the same system.

Prompting

ConceptPrompt Engineering: How to Write Better AI PromptsPrompt engineering is the practice of giving a model clear instructions, context, examples and format requirements so it produces useful, reliable output.ConceptWhat is a System Prompt?A system prompt is the standing set of instructions that defines how an AI model or assistant should behave, separate from what a user types in each message.ConceptCommon Prompting Mistakes (and How to Fix Them)Most poor AI output traces back to a handful of prompt problems: vague goals, missing context, no format, conflicting instructions and no testing.

Reliability

ConceptAI Hallucinations: Why Models Make Things UpA hallucination is output that sounds plausible but is false, unsupported or fabricated. It follows from how language models generate text.ConceptHow to Reduce AI HallucinationsYou cannot remove hallucinations completely, but you can lower them by grounding answers in sources, constraining the task, allowing "I don't know", and verifying output.ConceptHow to Evaluate AI Systems and OutputsAI evaluation means measuring quality on representative test cases — with automated checks, human review and, carefully, model-based judges — instead of trusting a few good demos.

Retrieval & Customisation

ConceptWhat are Embeddings in AI?An embedding is a list of numbers that represents the meaning of a piece of content, so similar meanings end up close together and can be found mathematically.ConceptWhat is a Vector Database?A vector database stores embeddings and quickly finds the ones nearest to a query vector, which is the retrieval step in semantic search and RAG.ConceptChunking: How to Split Documents for RAGChunking splits long documents into retrievable pieces. Chunk size and boundaries strongly affect whether retrieval finds the right information.ConceptWhat is RAG (Retrieval-Augmented Generation)?Retrieval-Augmented Generation fetches relevant information from your own data and gives it to a language model so answers are grounded in that material.ConceptWhat is Fine-Tuning an AI Model?Fine-tuning continues training a pre-trained model on your own examples so it adopts a consistent behaviour, format or specialised skill.

Safety & Privacy

ConceptWhat is Local AI? Running Models on Your Own DeviceLocal AI means running models on hardware you control — a laptop, workstation, server or even in the browser — instead of sending data to a hosted service.ConceptAI Privacy and Security: What to Know Before You Paste DataUsing AI safely means controlling what data leaves your environment, understanding provider terms, protecting credentials, and defending applications against manipulation.ConceptWhat is Prompt Injection?Prompt injection is an attack in which text supplied to a model — directly or hidden in documents, web pages or tool results — tries to override the developer's instructions.

Agents & Tools

ConceptWhat are AI Agents?An AI agent is a system in which a language model decides what to do next — calling tools, observing results and repeating — to pursue a goal with limited human direction.ConceptAgent Tools: How AI Agents Take ActionsTools are functions or services a model can ask to be run — search, databases, calculators, APIs, file access — which turn a text generator into something that can act.ConceptFunction Calling (Tool Calling) ExplainedFunction calling lets a model return a structured request to run a named function with arguments, so applications can connect models to code and data reliably.ConceptAgent Memory: Short-Term, Long-Term and BeyondModels have no built-in memory between requests. Agent memory is whatever your system stores and re-inserts into the context: recent messages, summaries, facts and retrieved knowledge.ConceptMulti-Agent Systems ExplainedA multi-agent system uses several model-driven agents with distinct roles or contexts that coordinate — for example a planner, specialists and a reviewer — to complete a larger task.ConceptAgentic Workflows: Patterns for Reliable AI AutomationAn agentic workflow orchestrates model calls and tools along a designed path — chaining, routing, parallelising, evaluating — giving more predictability than a fully autonomous agent.ConceptWhat is MCP (Model Context Protocol)?MCP is an open protocol that standardises how AI applications connect to external tools, data sources and prompts, so one integration can work across many AI clients.ConceptMCP Servers and Clients: Roles, Transports and SetupIn MCP, a host application runs clients that each connect to a server; servers expose tools, resources and prompts over a transport such as stdio or HTTP.

Python

ConceptPython for AI: When and How to Use ItPython is the most common language for calling models, preparing data, running small RAG pipelines and evaluating answers. Use it when you want clear scripts and a rich library ecosystem; keep secrets and long-running work on a machine you control.ConceptBuilding RAG with PythonBuild retrieval-augmented generation in Python by loading documents, chunking them, embedding chunks, storing vectors, retrieving neighbours for a question, stuffing them into a prompt, and checking the answer.ConceptCalling AI and Model APIs with PythonCall a model from Python with an HTTP request: read the API key from the environment, set a timeout, send a JSON body, handle errors, and pull the text out of the response.ConceptJSON and CSV Handling for AI Work in PythonRead JSON and CSV files in Python, validate the shape you expect, send clean records to a model or embedding API, and save outputs in a format the next step can load.ConceptPython AI Libraries: What to Use WhenChoose Python libraries by role: standard library and an HTTP client for API calls, a data library for tables, an embeddings or vector helper for retrieval, and an agent framework only when orchestration outgrows a simple loop.

JavaScript

ConceptJavaScript for AI ApplicationsUse JavaScript for AI UIs in the browser and for server-side calls in Node. Never put model API keys in front-end code; call your own backend, which holds the secret and talks to the model.ConceptTypeScript for AI ApplicationsTypeScript adds types to JavaScript so request and response shapes for model APIs and your own backend stay consistent, catching many integration bugs before runtime.ConceptTypeScript Types for an API ClientDefine TypeScript types for the request body, success response and error response of each API call, then implement a client function that returns a typed result and handles non-2xx cases.ConceptCalling AI Services from JavaScript and TypeScriptFrom Node (or another server), call the model vendor with fetch, an API key from the environment, timeouts and careful JSON parsing. Expose only your own authenticated route to browsers.ConceptStreaming AI Responses for Chat UIsStreaming delivers model tokens as they are generated (often SSE or chunked HTTP) so the UI can show text immediately. The client reads the byte stream, parses events, and appends text until the stream ends.

Interfaces

ConceptBuilding AI Interfaces with ReactBuild a React chat UI with a message list, an input box, loading and error states, and calls to your backend — never with a model API key in the client bundle.ConceptChatbot State and Conversation UI in ReactModel chatbot state as a messages array with roles, a pending flag, and optional streaming buffers. Keep the API key off the client and treat each turn as append-only history you send to your backend.ConceptNode.js for AI ApplicationsUse Node.js as the server that holds API keys, authenticates users, calls the model vendor, and exposes a thin route your React (or other) client can call.ConceptStreaming AI Responses with Node.jsIn Node.js, call the vendor streaming endpoint, then forward chunks to the browser as SSE or chunked HTTP while handling backpressure, aborts and errors.

Frameworks

ConceptWhat is an AI Framework?An AI framework is a library that helps you orchestrate prompts, tools, memory and model calls. It sits above the raw HTTP API and below your product logic — unlike the model itself.ConceptHow to Choose an Agent FrameworkChoose an agent approach by category: direct API loops, prompt/orchestration helpers, tool-calling loops, multi-agent graphs, or vendor SDKs. Named libraries illustrate categories; pick based on needs, and know when a framework is unnecessary.

Data

ConceptDatabases for AI ApplicationsAI apps still need ordinary databases for users, conversations, documents and evaluation rows. Vectors are an extra index for similarity — not a replacement for relational data.ConceptUsing PostgreSQL in AI ApplicationsUse PostgreSQL to store users, chats and documents with parameterised queries; call the model from your app server; add pgvector later only if you need similarity search in the same database.ConceptSQLSQL is the language of relational databases: you declare tables and ask questions with SELECT, JOIN, and WHERE. ORMs can generate SQL, but you still need to read it to understand performance and correctness.ConceptPostgreSQL, the Relational DatabasePostgreSQL (Postgres, often shortened to PG) is an open-source relational database. It stores tables with constraints, speaks SQL, and can add extensions such as pgvector. It is a strong default system of record for applications.ConceptMySQLMySQL is a widely used open-source relational database, common in existing web stacks and managed as Aurora or Cloud SQL. It speaks SQL with its own dialect and defaults. Pick it to extend a MySQL estate; pick PostgreSQL for many new apps unless your platform is already MySQL.ConceptSQLiteSQLite is a relational database stored in one file inside your process. It is the right default for local tools, tests, mobile apps and modest servers. It is the wrong default for many writers across many application servers.ConceptMongoDBMongoDB is a document database: you store JSON-like documents in collections and query them by fields. It fits data you mostly load as a whole document. It is a poor replacement for relational constraints you actually need.ConceptRedisRedis is an in-memory data store for keys with a few rich values: strings, lists, sets, sorted sets, and expirations. Applications use it for cache, rate limits, short-lived sessions and queues, not as the only copy of data they cannot lose.ConceptORMs and PrismaAn ORM maps database rows to objects in your language. Prisma is a popular TypeScript ORM that generates a client from a schema file. ORMs speed ordinary queries and can hide the SQL that explains a slow page.ConceptChoosing a Vector StoreChoose a vector store by how you already run data: pgvector inside PostgreSQL, a dedicated vector database, or an in-process library for small sets. The retrieval quality depends more on chunking and embeddings than on the brand.

Integrations

ConceptConnecting AI Agents to External ApplicationsAn agent reaches Gmail, calendars or CRMs only through tools you define and run. The model requests a tool call; your server performs the API or MCP action with proper credentials and permissions.ConceptGmail and Google Integration for AI AgentsAn agent accesses Gmail through your server: the user completes Google OAuth, you store a refresh token, and a tool lists or drafts mail via the Gmail API. GoVibeTool has no Gmail connector; treat message text as untrusted.ConceptOAuth for AI AgentsOAuth lets a user grant your agent app limited access to a provider without sharing their password. Your server holds tokens; the model only triggers tools that use those tokens under your policy.ConceptPermissions and Security for AI IntegrationsLimit OAuth scopes, enforce authorisation in code, require human approval before send or delete, and treat tool outputs such as email bodies as untrusted to reduce prompt injection and data leaks.ConceptGmail API Scopes and VerificationGmail API scopes fall into non-sensitive, sensitive and restricted classes. Restricted scopes (including gmail.readonly) require OAuth app verification, a security assessment if you store the data on servers, and Google's Limited Use rules, which bar using that data to train generalised AI models.

Building

ConceptJSON Fundamentals for API WorkJSON is a text format for structured data — objects, arrays, strings, numbers, booleans and null — used by almost every modern HTTP API and by model tool arguments.ConceptHow to Validate JSONValidate JSON by parsing it safely, then checking required fields, types and allowed values — ideally with a schema. Reject unexpected shapes before they reach your model or database.ConceptWhat is an API?An API is how one program asks another program to do work or return data — commonly over HTTP with JSON. Model providers expose APIs; your app exposes APIs to your own UI.ConceptREST APIs for Application DevelopmentREST APIs expose resources over HTTP using methods like GET, POST, PUT and DELETE, usually with JSON bodies and status codes that tell you whether the call worked.ConceptAPI Authentication: Keys, Sessions and OAuthChoose API keys for server-to-server access, session cookies for browser users, and OAuth when acting on a user's behalf at another provider. Always keep secrets on the server.

Microsoft 365

ConceptWhat is SharePoint?SharePoint is Microsoft's platform for team sites, document libraries, pages and intranet content inside Microsoft 365. Custom UI on modern pages is built with the SharePoint Framework, not classic farm solutions.ConceptSharePoint Framework (SPFx)SPFx is the SharePoint Framework: Microsoft's TypeScript toolchain for client-side SharePoint web parts and extensions. The default UI library is React, and projects are TypeScript-first.ConceptHow to Build an SPFx Web PartBuild an SPFx web part by scaffolding a React and TypeScript project with the SharePoint generator, coding against the web part context, serving the workbench, then packaging an .sppkg for the app catalog.ConceptMicrosoft 365 for DevelopersMicrosoft 365 (M365) is the cloud suite that includes SharePoint, Exchange, Teams and the identity boundary in Entra ID. Developers extend it with Graph, SPFx and Teams apps rather than scraping the web UI.ConceptMicrosoft Graph APIMicrosoft Graph is the REST API for Microsoft 365: users, groups, mail, files, sites, calendars and Teams. You call https://graph.microsoft.com with an Entra ID access token and the scopes an admin has consented.ConceptPower Platform BasicsPower Platform is Microsoft's low-code family: Power Apps for UI, Power Automate for workflows, Power BI for reports, and Dataverse for data, sitting beside Microsoft 365 and Azure rather than replacing them.ConceptMicrosoft Entra IDMicrosoft Entra ID is Microsoft's cloud identity service, formerly Azure Active Directory. It signs users in, issues tokens, and holds the app registrations that Microsoft Graph, Azure and your own apps trust.ConceptMicrosoft Teams DevelopmentTeams apps are packages of tabs, bots, message extensions and connectors that run inside Microsoft Teams and usually call Microsoft Graph. They are not SPFx web parts, though a tab can host a web app or an SPFx part.

Cloud

ConceptAzure Fundamentals for Application DevelopersAzure is Microsoft's cloud for compute, data, identity integration and AI hosting. Application teams usually start with App Service or containers, a database, storage, Key Vault and Entra ID, not with every service in the catalog.ConceptAWS Fundamentals for Application DevelopersAmazon Web Services is a cloud platform: compute, storage, databases and managed APIs in regions and accounts. Most new apps need a place to run, a database, object storage, identity and logs, not the full catalog.ConceptGoogle Cloud Fundamentals for Application DevelopersGoogle Cloud (GCP) is a public cloud organised into projects. Application teams usually need Cloud Run or another compute option, a database, Cloud Storage, IAM, and logs. BigQuery and Vertex are optional, not the starting point.

AI ecosystem

ConceptHugging Face for BuildersHugging Face is the common hub for open model weights, datasets and libraries such as transformers and the Hub. Teams use it to find models, download weights, and sometimes to host inference, then decide what must run in their own environment.ConceptLangChain, PracticallyLangChain is a Python and JavaScript framework that wires prompts, models, tools and retrieval into chains or agents. It speeds prototypes and adds abstraction; simple apps are often clearer calling a model API directly.ConceptLlamaIndex for RetrievalLlamaIndex is a framework focused on loading data, building indexes and retrieving context for language models. It overlaps LangChain on RAG, with more emphasis on document indices and less on being a general agent toolkit.ConceptModel APIs for ApplicationsA model API is an HTTP service that accepts messages and returns generated text or tool calls. You choose a model, send a prompt with authentication, handle tokens and errors, and keep the vendor SDK or raw HTTP behind a small interface you own.ConceptLocal LLM RuntimesLocal LLM runtimes such as Ollama, llama.cpp and vLLM load open weights on hardware you control and often expose an HTTP API. They trade operational work and hardware limits for privacy and a predictable machine.ConceptRAG FrameworksA RAG framework packages document loading, chunking, embeddings, retrieval and prompting. LlamaIndex and LangChain are the usual examples. Use one for speed; drop to plain code when you need to see every step.ConceptAI SDKsAn AI SDK is a library that calls a model API for you: authentication, types, streaming and retries. The Vercel AI SDK, official provider SDKs and similar toolkits are convenience layers. Your architecture is still the HTTP contract and your own wrapper.

Languages

ConceptPythonPython is a readable general-purpose language with a huge library ecosystem. It dominates data work, scripting and many AI prototypes, and it is a solid choice for backends when raw latency is not the main constraint.ConceptJavaScriptJavaScript is the language of web browsers and, with Node.js, a server language too. It is dynamically typed, event-driven, and the base that TypeScript compiles to. JS is the everyday abbreviation.ConceptTypeScriptTypeScript (TS) is JavaScript with a static type system. It compiles to plain JavaScript, catches a class of bugs in the editor, and is the default for SPFx, large React apps and many Node services.ConceptJavaJava is a statically typed, object-oriented language that runs on the JVM. Enterprises use it for long-lived services, Android's older stacks, and anywhere a stable toolchain and explicit types matter more than a short script.ConceptC#C# is Microsoft's typed language on .NET. Teams use it for Windows services, cross-platform web APIs, Unity games, and Azure backends. It is a strong default when the stack is already Microsoft.ConceptThe Go Programming LanguageGo (Golang) is a small, statically typed language designed for simple concurrent services and single-binary deployments. Teams pick it for APIs, CLIs and infrastructure tools when they want fast builds and boring code.ConceptRustRust is a systems language with ownership and borrowing instead of a garbage collector. It prevents many memory bugs at compile time. Teams use it for performance-critical tools, WASM, and services where correctness matters more than a short compile.

Web

ConceptHTML and CSSHTML is the structure of a web page. CSS is how that structure looks. JavaScript changes behaviour. You need the first two even when a framework such as React writes them for you.ConceptReactReact is a JavaScript library for building user interfaces out of components that render when state changes. It is the usual UI layer in SPFx and in many AI chat panels, but it is not a backend or a complete framework.ConceptNext.jsNext.js is a React framework that adds routing, server rendering, and server-side code next to the UI. Use it for full web apps. Use plain React when you only need a widget inside someone else's page.ConceptNode.jsNode.js is a JavaScript runtime outside the browser, used for servers, tools and scripts. It is built on V8 and an event loop, and it is the usual host for Express, Next.js servers, and SPFx build tooling.ConceptExpressExpress is a small Node.js library for HTTP servers: routes, middleware, and request and response objects. It is enough for many APIs. It does not include a database, auth, or a front end.ConceptFrontend and BackendThe frontend is the interface running on the user's device. The backend is the server you control: data, secrets, and business rules. Browsers are not a safe place for secrets, which is why AI keys and database passwords stay on the backend.ConceptGraphQLGraphQL is an API style where the client sends a query that names the fields it wants, and the server resolves them against a schema. It reduces over-fetching and adds a different set of caching, auth and performance problems than REST.ConceptWebSocketsWebSockets are a persistent two-way connection between a client and a server, used for live updates. They are more machinery than HTTP request and response. Many AI chat UIs only need server-sent events, not a full socket.

Developer infrastructure

ConceptGitGit is a distributed version-control system. It records snapshots of a project so you can branch, review, and return to a known state. GitHub is a host for Git repositories, not the tool itself.ConceptGitHubGitHub is a website and service that hosts Git repositories and adds pull requests, issues, actions and access control. Git is the tool. GitHub is one place teams push Git history and review it.ConceptDockerDocker is a tool for building and running containers: a packaged process with its filesystem. Use it to make local and production environments match. Do not use it to hide a broken build, and do not run everything as root.ConceptContainersA container is a process, or a small group of processes, isolated with its own filesystem view, using the host kernel. A virtual machine emulates a whole computer and boots its own kernel. Containers are the usual way to ship a service.ConceptCI and CDContinuous integration automatically builds and tests every change. Continuous delivery prepares a deployable artifact; continuous deployment ships it. The point is a repeatable pipeline, not a YAML file that nobody trusts.ConceptEnvironment VariablesEnvironment variables are named strings a process inherits from whoever started it. They are the standard way to configure an app per environment. Secrets can live there at runtime, but they must not live in git, images, or client bundles.ConceptPackage ManagersA package manager installs libraries and records versions. npm, pip, Cargo and NuGet are examples. A lockfile pins the tree so tomorrow's install matches today's. Installing global packages at random is how machines drift.

Security

ConceptOAuth 2.0OAuth 2.0 is a protocol for granting a client limited access to an API without sharing the user's password. It is how sign-in buttons and Microsoft Graph access tokens usually start. It is not itself a login profile; OpenID Connect adds identity on top.ConceptOpenID ConnectOpenID Connect (OIDC) is a thin identity layer on OAuth 2.0. It gives your app an ID token that says who signed in. OAuth alone gives an access token for an API. Most sign-in-with buttons are OIDC.ConceptJSON Web TokensA JWT is a signed JSON payload used as a token. Anyone can read the payload. Only a valid signature, the right issuer and the right audience make it trustworthy. Decoding is not validation.ConceptAPI KeysAn API key is a secret string that identifies a caller to an API. It is simple and blunt: anyone who has it is the caller. Keep keys on the server, scope them, rotate them, and do not ship them in a browser or a mobile binary.ConceptAuthentication versus AuthorizationAuthentication answers who is calling. Authorization answers what they may do. A valid login is not permission to read every row. Check both, on the server, for every action that matters.ConceptCORSCORS is a browser rule for cross-origin requests. The server must opt in with response headers. It is not authentication, and it is not a barrier to non-browser clients. Fix it by allowing the origins you intend, not by reflecting every origin with credentials.ConceptWebhooksA webhook is an HTTP callback: another system POSTs to your URL when something happens. You must authenticate the sender, respond quickly, and process the event idempotently. A public URL with no signature check is an unauthenticated API.

Reasoning & Test-Time Compute

ConceptTest-Time Compute (Inference-Time Scaling)Test-time compute means spending more computation when a model answers — longer reasoning, several attempts, search or verification — instead of only making the model bigger.ConceptReasoning ModelsReasoning models are language models trained or configured to work through a problem in intermediate steps before answering, generally improving multi-step tasks at the cost of time and tokens.ConceptChain-of-Thought PromptingChain-of-thought prompting asks a model to lay out intermediate steps before answering, which can improve multi-step reasoning on some tasks.ConceptReasoning Transparency and Hidden ReasoningSome AI providers keep a model's raw reasoning private and show only a summary or the final answer. This entry explains what that means for builders, users and safety, without trying to extract anything hidden.ConceptSelf-Consistency (Sampling and Voting)Self-consistency samples several independent reasoning attempts and chooses the answer that most of them agree on.ConceptBest-of-N SamplingBest-of-N draws several candidate answers and returns the one a scorer — a reward model, verifier or test — rates highest.ConceptSearch over Reasoning (Tree of Thoughts, MCTS, Beam Search)Search-based reasoning explores several partial solution paths, scores them, and expands the promising ones instead of committing to a single chain.ConceptProcess Reward Model (PRM)A process reward model scores each intermediate step of a reasoning chain, giving step-level feedback rather than only judging the final answer.ConceptOutcome Reward Model (ORM)An outcome reward model scores a completed answer, judging only whether the final result is right (or preferred), not the steps that led to it.ConceptReinforcement Learning for Reasoning (RLVR and GRPO)Reinforcement learning from verifiable rewards trains models on tasks with checkable answers — maths, code — so they learn longer, self-correcting reasoning. GRPO is one common algorithm.ConceptThinking Budgets and Reasoning EffortMany reasoning models let you cap or tune how much reasoning they do per request. Setting it well balances accuracy, latency and cost.ConceptOpenAI Reasoning Models: o-series, Effort, Reasoning Tokens and DeprecationsOpenAI's reasoning models think before answering. The API bills hidden reasoning tokens as output, lets you set a reasoning effort and request a summary (never the raw reasoning), and older o-series models are being retired on published dates.

Structured Generation

ConceptSampling and Decoding (Temperature, Top-p, Beam Search)Decoding turns a model's next-token probabilities into text. Settings such as temperature, top-p, top-k and beam search control how predictable or varied the output is.ConceptConstrained Decoding (Guided Generation)Constrained decoding restricts a model's next-token choices at each step so the output must follow a format — a JSON schema, regex or grammar — and cannot be malformed.ConceptStructured Outputs from LLMsStructured outputs make a model return data in a defined shape — usually JSON matching a schema — so software can use it without brittle parsing.ConceptGrammar-Guided Generation (CFG and Regex Constraints)Grammar-guided generation constrains a model with a formal grammar or regular expression so output belongs to a defined language — a config format, a small SQL subset, an ID pattern.ConceptGBNF Grammars (llama.cpp)GBNF is the grammar notation used by llama.cpp to constrain local-model output; it is a BNF-style language for rules over characters.ConceptJSON SchemaJSON Schema is a vocabulary for describing and validating the shape of JSON data. It underpins tool definitions, structured outputs and API contracts for AI applications.ConceptOutlines (Structured Generation Library)Outlines is an open-source Python library for guided generation: constraining model output with regexes, JSON schemas or grammars.

ML, Deep Learning & RL

ConceptSupervised LearningSupervised learning trains a model on labelled examples — inputs paired with the right answers — so it can predict answers for new inputs.ConceptUnsupervised LearningUnsupervised learning finds structure in unlabelled data — clusters, patterns, compact representations — without being told the right answers.ConceptSelf-Supervised LearningSelf-supervised learning creates its own training signal from raw data — predicting hidden words, masked patches or the next token — and is how most foundation models are pre-trained.ConceptNeural NetworksA neural network is a model built from layers of simple numeric units whose connection strengths are learned from data. It is the foundation of deep learning.ConceptDeep LearningDeep learning is machine learning with neural networks that have many layers, learning useful features directly from raw data like pixels, audio and text.ConceptBackpropagation and Gradient DescentGradient descent adjusts a model's parameters in the direction that reduces loss; backpropagation is the efficient way to compute those gradients through a neural network.ConceptOverfitting and RegularizationOverfitting is when a model memorises training data and fails on new data. Regularisation, more data, early stopping and simpler models reduce it.ConceptTransfer LearningTransfer learning reuses a model pre-trained on a large task as the starting point for a new, smaller one, saving data and compute.ConceptReinforcement LearningReinforcement learning trains an agent to choose actions in an environment to maximise cumulative reward through trial and feedback.ConceptMarkov Decision ProcessesA Markov decision process is the mathematical framework for sequential decision-making under uncertainty that underlies reinforcement learning.ConceptProximal Policy Optimization (PPO)PPO is a widely used policy-gradient reinforcement-learning algorithm that limits how far each update can move the policy, for stable training.ConceptDeep Q-Networks (DQN) and Q-LearningQ-learning learns the value of taking each action in each state; DQN uses a neural network to do this from raw inputs such as game pixels.

Model Architectures

ConceptConvolutional Neural Networks (CNNs)CNNs process grid-like data such as images by sliding small learned filters across it, capturing local patterns with far fewer parameters than fully connected nets.ConceptRecurrent Neural Networks and LSTMsRNNs process sequences one step at a time while carrying a hidden state; LSTMs and GRUs add gates to remember longer. Transformers have largely replaced them for language.ConceptVision Transformers (ViT)Vision Transformers split an image into patches and process them with transformer attention, matching or beating CNNs at scale.ConceptMixture of Experts (MoE)Mixture-of-experts models contain many specialist sub-networks and route each token to only a few, giving large capacity at lower per-token compute.ConceptState Space Models (Mamba and S4)State space models process sequences with a compact recurrent state and linear-time computation, offering an alternative to attention for long inputs.ConceptDiffusion ModelsDiffusion models generate images, audio and video by learning to reverse a gradual noising process, turning random noise into structured content step by step.ConceptGenerative Adversarial Networks (GANs)GANs train a generator and a discriminator against each other so the generator learns to produce realistic samples.ConceptVariational Autoencoders (VAEs)A VAE learns to compress data into a smooth probabilistic latent space and decode samples from it, enabling generation and representation learning.ConceptGraph Neural Networks (GNNs)GNNs learn from data shaped like graphs — molecules, social networks, road maps — by passing messages between connected nodes.ConceptKV CacheThe KV cache stores attention keys and values for past tokens so a model need not recompute them while generating, trading memory for speed.ConceptFlashAttentionFlashAttention is an exact attention algorithm that reorganises computation to cut memory traffic, speeding up transformers and enabling longer sequences.ConceptPositional Encoding (Sinusoidal, RoPE, ALiBi)Attention has no built-in sense of order, so transformers add positional information. Methods like RoPE and ALiBi affect how well models handle long contexts.ConceptAttention Sinks and StreamingLLMAttention sinks are the first few tokens of a sequence that transformers attend to heavily even when they carry little meaning. StreamingLLM keeps those tokens plus a recent window in the KV cache so a model can stream indefinitely without collapsing.ConceptFlashAttention-3 (and FlashAttention-4)FlashAttention-3 is an exact attention implementation redesigned for NVIDIA Hopper GPUs. It overlaps data movement and computation, interleaves matrix multiplication with softmax, and uses FP8 low-precision, reaching up to about 740 TFLOPs/s in FP16 on H100.

LLM Techniques

ConceptScaling LawsScaling laws describe how model loss tends to improve predictably with more parameters, data and compute — and guide how to split a training budget.ConceptInstruction TuningInstruction tuning fine-tunes a pre-trained language model on many instruction-and-response examples so it follows requests instead of merely continuing text.ConceptLoRA and Parameter-Efficient Fine-Tuning (PEFT)LoRA and other PEFT methods adapt a large model by training a small number of added parameters while the original weights stay frozen.ConceptQuantization of AI ModelsQuantisation stores and computes a model's numbers at lower precision — for example 8-bit or 4-bit — to cut memory and often speed up inference, at some quality cost.ConceptKnowledge DistillationDistillation trains a smaller student model to imitate a larger teacher, capturing much of its behaviour at lower cost.ConceptSpeculative DecodingSpeculative decoding speeds up generation by letting a small draft model propose several tokens that a large model verifies in one pass, without changing the output distribution.ConceptSmall Language Models (SLMs)Small language models are compact LLMs — typically up to a few billion parameters — that run cheaply, often on laptops or phones, and can match larger models on narrow tasks.ConceptOpen-Weights vs Closed ModelsOpen-weights models publish their trained parameters for download and use under a licence; closed models are available only through a provider's service. "Open weights" is not the same as "open source".

Multimodal AI

ConceptVision-Language Models (VLMs)Vision-language models take images (and often video) together with text and answer questions, describe scenes, read documents and follow visual instructions.ConceptContrastive Learning and CLIPCLIP learns a shared embedding space for images and text by training matching image–caption pairs to be close together, enabling zero-shot image classification and cross-modal search.ConceptSpeech AI: Speech Recognition and Text-to-SpeechSpeech AI turns audio into text (automatic speech recognition) and text into audio (text-to-speech); modern systems use large neural models and can power voice assistants.ConceptDocument Understanding AI (OCR, Layout, Extraction)Document understanding combines OCR, layout analysis and language models to extract structured information from PDFs, scans and forms.ConceptVideo Generation and Video Understanding ModelsVideo models generate or analyse moving images, typically by extending image diffusion or transformer methods across time. Quality, controllability and cost vary widely.ConceptSora (OpenAI Video Generation) and Its DiscontinuationSora was OpenAI's text-to-video model family. Sora 2 launched on 30 September 2025 with synchronised audio; OpenAI says the Sora product was discontinued on 26 April 2026 and scheduled the API's shutdown for 24 September 2026.

Agent Protocols, Frameworks & Runtimes

ConceptA2A (Agent2Agent) ProtocolA2A is an open protocol for communication between independent AI agents — discovering each other's capabilities, exchanging tasks and returning results — complementing MCP, which connects agents to tools.ConceptAgent Protocol Landscape (MCP, A2A and Others)Several protocols are emerging for agent interoperability. They operate at different layers: tools and context, agent-to-agent messaging, discovery and payments/identity.ConceptMCP Security Risks and MitigationsMCP servers can read data and take actions for an AI application, so they carry risks: malicious or compromised servers, tool-description poisoning, over-broad permissions and prompt injection through tool results.ConceptLangGraphLangGraph is a library for building stateful, multi-step agent workflows as graphs of nodes and edges, with persistence, human-in-the-loop checkpoints and streaming.ConceptDSPyDSPy is a framework for programming — rather than hand-prompting — language-model pipelines: you declare modules and metrics, and optimisers tune prompts or weights automatically.ConceptAutoGen (Microsoft)AutoGen is an open-source Microsoft Research framework for multi-agent LLM applications where agents converse and collaborate. Its repository now states it is in maintenance mode, with Microsoft Agent Framework as the recommended successor for new projects.ConceptCrewAICrewAI is a Python framework for orchestrating role-playing agent "crews" that carry out defined tasks, with a higher-level API than graph-style frameworks.ConceptSemantic KernelSemantic Kernel is Microsoft's open-source SDK (C#, Python, Java) for integrating LLMs into applications with plugins, planning and agent abstractions.ConceptOpenAI Agents SDKThe OpenAI Agents SDK is an open-source library for building agentic applications with agents, tools, handoffs and guardrails; it is provider-vendor maintained and can be used with other model providers via adapters.ConceptvLLMvLLM is an open-source LLM inference and serving engine known for PagedAttention, continuous batching and high throughput on GPUs.Conceptllama.cppllama.cpp is a C/C++ inference engine for running LLMs efficiently on CPUs and consumer GPUs, using the GGUF format and aggressive quantisation.ConceptOllamaOllama is a tool that downloads, manages and runs open models locally with a simple command line and a local HTTP API.ConceptONNX RuntimeONNX Runtime is a cross-platform engine for running machine-learning models exported to the ONNX format across CPUs, GPUs and edge devices.ConceptPyTorchPyTorch is the leading open-source deep-learning framework for research and increasingly production, built on tensors, automatic differentiation and eager execution.ConceptMicrosoft Agent FrameworkMicrosoft Agent Framework (MAF) is Microsoft's open-source framework for building agents and multi-agent workflows in .NET and Python, positioned by Microsoft as the successor to both AutoGen and Semantic Kernel.ConceptMCP Authorization (OAuth for MCP Servers)MCP authorization lets a client call a protected remote MCP server using OAuth 2.1: the server publishes protected-resource metadata, the client discovers the authorization server, requests a token bound to that server, and sends it as a bearer token on every request.

Agent & RAG Patterns

ConceptThe ReAct Pattern (Reason + Act)ReAct interleaves reasoning steps with actions: the model decides what to do, calls a tool, reads the observation and reasons again, forming the basic agent loop.ConceptAgent Planning and Task DecompositionPlanning agents break a goal into steps before or while acting, revising the plan as results arrive. Explicit plans help with long tasks and make oversight easier.ConceptComputer-Use and Browser AgentsComputer-use agents operate software through its user interface — viewing screens, clicking and typing — rather than through APIs. They are powerful and risky, and still error-prone.ConceptCode Execution and Sandboxing for AI AgentsAgents that write and run code need isolation: sandboxes limit filesystem, network, resources and credentials so a mistake or injected instruction cannot harm the host.ConceptHybrid Search and RerankingHybrid search combines keyword (lexical) and vector (semantic) retrieval; reranking then reorders the top candidates with a more precise model. Together they improve RAG retrieval quality.ConceptGraphRAG (Knowledge-Graph Retrieval)GraphRAG organises documents into a knowledge graph of entities and relationships, enabling retrieval and summarisation across connected information rather than isolated chunks.ConceptAgentic RAG (Adaptive and Self-Correcting Retrieval)Agentic RAG lets a model decide when and how to retrieve — rewriting queries, choosing sources, checking relevance and retrying — instead of a single fixed retrieve-then-generate pass.ConceptContext EngineeringContext engineering is deciding what goes into a model's context window — instructions, retrieved facts, tool results, memory, examples — and in what form, to make agents reliable.

Evaluation & Benchmarks

ConceptAI Benchmarks and LeaderboardsA benchmark is a fixed test set and scoring method used to compare models. Leaderboards rank models on them — useful for orientation, unreliable as a sole basis for choosing a model.ConceptBenchmark Contamination and Data LeakageContamination happens when benchmark questions or answers appear in a model's training data, inflating scores. It undermines claims of capability and is hard to rule out fully.ConceptLLM-as-a-JudgeLLM-as-a-judge uses a language model to grade other model outputs against a rubric or compare pairs. It scales evaluation but has biases and must be calibrated against humans.ConceptHuman Preference Evaluation (Arenas and Elo Rankings)Human-preference evaluation collects people's votes between model outputs, often turning pairwise wins into ratings (such as Elo or Bradley–Terry scores) as in public "arena" leaderboards.ConceptMMLU (Massive Multitask Language Understanding)MMLU is a multiple-choice benchmark spanning 57 subjects used to compare the breadth of knowledge in language models. It is widely reported but now largely saturated and contains known label issues.ConceptHumanEval and Code Generation BenchmarksHumanEval measures whether a model can write correct Python functions from docstrings, scored by running unit tests (pass@k). It is small and saturated; newer code benchmarks test harder and more realistic tasks.ConceptSWE-bench (Software Engineering Benchmark)SWE-bench evaluates whether AI systems can resolve real GitHub issues by editing a repository so that tests pass. It is a common measure of coding agents, with several variants and known caveats.ConceptGSM8K and Math BenchmarksGSM8K (grade-school maths word problems) and MATH (competition problems) are common reasoning benchmarks; GSM8K has been widely reported as close to saturated for strong models (a time-sensitive judgement), so harder maths sets are used to separate them.ConceptRAG Evaluation (Retrieval and Answer Quality)Evaluate a RAG system in two layers: did retrieval find the right passages, and did the generator answer faithfully and completely from them? Measure both separately.ConceptAgent EvaluationEvaluating agents means checking not just final answers but the path: tool choices, arguments, safety constraints, cost and whether the end state is correct.ConceptCommon Evaluation Metrics (Accuracy, F1, BLEU, ROUGE, Perplexity)Different tasks need different metrics: accuracy/precision/recall/F1 for classification, BLEU/ROUGE for overlap with references, perplexity for language modelling, and rubric- or judge-based scores for open-ended generation.ConceptInspect (UK AI Security Institute Evaluation Framework)Inspect is an open-source Python framework from the UK AI Security Institute for running LLM evaluations: you define a task from a dataset, a solver and a scorer, and Inspect runs it against any supported model and logs everything for review.

AI Infrastructure & MLOps

ConceptMLOps (Machine Learning Operations)MLOps applies software-engineering and operations practices — versioning, automation, testing, deployment, monitoring — to machine-learning systems so they stay reliable in production.ConceptModel Serving and Inference InfrastructureServing turns a trained model into a reliable service: loading it on suitable hardware, batching requests, managing memory, scaling, and meeting latency and cost targets.ConceptGPUs and AI AcceleratorsAI workloads run on hardware that does massive parallel matrix maths — GPUs, TPUs and other accelerators. Memory size and bandwidth often matter more than raw compute for LLMs.ConceptDistributed Training (Data, Tensor and Pipeline Parallelism)Large models are trained across many accelerators by splitting data, model layers or tensors, and sharding optimiser state. Each strategy trades memory, communication and complexity.ConceptLLM Observability (Tracing, Logging and Monitoring)LLM observability records prompts, retrieved context, tool calls, outputs, latency, cost and feedback so teams can debug, evaluate and improve AI applications — while protecting sensitive data in those logs.ConceptModel Drift and Production MonitoringOver time, the data a model sees and the world it models change, so performance can decay. Drift monitoring detects shifts in inputs, outputs and outcomes before users notice.ConceptPrompt CachingPrompt caching lets a provider or serving engine reuse the computation for a repeated prompt prefix — such as a long system prompt or document — cutting latency and often cost on later requests.ConceptLLM Cost OptimizationReduce the spend of LLM applications by sending fewer tokens, using the right-sized model, caching, batching, routing and limiting retries — while measuring quality so savings do not erode it.ConceptMLflowMLflow is an open-source platform for managing the machine-learning lifecycle: experiment tracking, model packaging, a model registry and, increasingly, GenAI tracing and evaluation.ConceptRay (Distributed Computing for AI)Ray is an open-source framework for scaling Python and AI workloads across a cluster, with libraries for data processing, training, tuning and serving.Conceptpgvector (Vector Search in PostgreSQL)pgvector is a PostgreSQL extension that adds a vector column type and similarity search, letting you keep embeddings alongside relational data.

Safety, Alignment & Governance

ConceptAI AlignmentAI alignment is the effort to make AI systems pursue the goals and values their developers and users actually intend — helpful, honest and safe — including when instructions are imperfect or situations are new.ConceptPreference Optimization (Aligning Models from Human or AI Preferences)Preference optimisation is the family of methods that tune a model using comparisons of better versus worse responses — RLHF, DPO and related algorithms. It is an alignment technique family, distinct from ordinary reinforcement learning.ConceptDirect Preference Optimization (DPO)DPO aligns a language model directly from preference pairs using a simple classification-style loss against a reference model — without training a separate reward model or running an online RL loop.ConceptRLHF (Reinforcement Learning from Human Feedback)RLHF trains a language model to behave in ways people prefer: gather human comparisons of outputs, fit a reward model, then optimise the model against that reward with reinforcement learning under a KL constraint.ConceptConstitutional AI and RLAIF (AI Feedback)Constitutional AI trains a model to critique and revise its own outputs against a written set of principles, and uses AI-generated preference labels (RLAIF) instead of or alongside human ones.ConceptReward Hacking and Over-OptimizationReward hacking happens when an AI system exploits flaws in its reward signal — scoring high without doing what was intended — including over-optimising a learned reward model.ConceptSycophancy in Language ModelsSycophancy is a model's tendency to tell users what they want to hear — agreeing with mistaken claims or flattering them — rather than what is accurate.ConceptRed-Teaming AI SystemsRed-teaming is structured adversarial testing of an AI system — humans or models trying to make it fail, misbehave or leak — to find weaknesses before real attackers or users do.ConceptAI GuardrailsGuardrails are runtime controls around an AI system — input and output filters, policy checks, permission limits, validation and human approvals — that reduce harmful or unwanted behaviour.ConceptMechanistic InterpretabilityMechanistic interpretability tries to reverse-engineer what is happening inside neural networks — identifying features and circuits — rather than only observing inputs and outputs.ConceptAI Bias and FairnessAI systems can reproduce or amplify biases in their data and design, producing systematically different outcomes for different groups. Fairness work measures, mitigates and documents this.ConceptAI GovernanceAI governance is the set of policies, roles, processes and legal controls an organisation or government uses to ensure AI is developed and used responsibly, safely and lawfully.ConceptEU AI ActThe EU AI Act (Regulation (EU) 2024/1689) is a risk-based law regulating AI systems placed on or used in the EU market, with prohibited practices, obligations for high-risk systems, transparency duties and rules for general-purpose AI models.ConceptNIST AI Risk Management FrameworkThe NIST AI RMF is a voluntary US framework for managing AI risks, organised around four functions — Govern, Map, Measure and Manage — with a companion profile for generative AI.ConceptISO/IEC 42001 (AI Management System)ISO/IEC 42001:2023 specifies requirements for an AI management system — a structured way for organisations to govern AI development and use — and can be certified by accredited bodies.ConceptModel Cards and Datasheets (AI Documentation)Model cards and datasheets are structured documents describing what a model or dataset is, how it was built and evaluated, intended uses and limitations — enabling informed and accountable use.ConceptC2PA and Content Provenance (Content Credentials)C2PA is an open technical standard for attaching tamper-evident provenance information — who made or edited content, and with what tools, including AI — to images, video and audio.ConceptOWASP Top 10 for LLM ApplicationsThe OWASP Top 10 for LLM Applications is a community-maintained list of the most critical security risks in LLM-based applications, such as prompt injection, sensitive information disclosure and excessive agency.

AI for Science

ConceptAI for ScienceAI for science applies machine learning to scientific problems — predicting protein structures, accelerating simulations, forecasting weather, proposing materials and molecules — complementing experiment and theory.ConceptAlphaFold (Protein Structure Prediction)AlphaFold is a family of deep-learning models from Google DeepMind that predict protein 3-D structure from amino-acid sequence with accuracy competitive with experiments for many proteins.ConceptPhysics-Informed Neural Networks (PINNs)PINNs train neural networks to satisfy the governing equations of a physical system (as differential-equation residuals) in addition to fitting data, to solve or learn from scientific problems with limited data.ConceptAI Weather and Climate Forecasting ModelsMachine-learning weather models learn from decades of reanalysis data to produce forecasts far faster than traditional numerical models, with accuracy that is competitive on many standard metrics.ConceptAI in Drug DiscoveryAI helps drug discovery by predicting molecular properties, designing candidate molecules, modelling protein–ligand interactions and prioritising experiments — but clinical validation and safety testing still decide whether a drug works.ConceptAI for Materials DiscoveryMachine learning accelerates materials science by predicting properties of candidate crystals and molecules and guiding which to synthesise — with experimental validation still required.

Robotics & Embodied AI

ConceptEmbodied AIEmbodied AI is AI that perceives and acts in the physical (or simulated) world through a body — robots, drones, vehicles — rather than only processing text or images.ConceptVision-Language-Action (VLA) ModelsVLA models take camera images and a language instruction and output robot actions, extending vision-language models so one network can follow commands like "put the cup in the sink".ConceptWorld ModelsA world model is a learned model of how an environment evolves in response to actions, used to imagine outcomes for planning, to train agents cheaply, or to generate interactive simulations.ConceptImitation Learning (Behaviour Cloning and DAgger)Imitation learning trains a policy from expert demonstrations — for example human teleoperation of a robot — instead of from a hand-designed reward.ConceptSim-to-Real TransferSim-to-real transfer trains robot policies in simulation and deploys them on real hardware, using techniques such as domain randomisation to bridge the gap between simulated and real physics and sensing.ConceptNVIDIA Cosmos (World Foundation Models for Physical AI)NVIDIA Cosmos is an open platform of world foundation models for physical AI — robots, autonomous vehicles and smart spaces. Cosmos 3 combines a reasoner and a generator in one omnimodal family released in three sizes.

Comparisons

ComparisonRAG vs Fine-Tuning: Which Should You Use?RAG gives a model access to information at question time; fine-tuning changes the model's behaviour through additional training. They solve different problems and are often combined.ComparisonAI Agent vs Chatbot vs Assistant: What's the Difference?The terms overlap and are used loosely. The practical difference is how much the system can do on its own: converse, help with tasks on request, or pursue goals using tools in a loop.ComparisonMCP vs API: What's the Difference?An API is an interface to one service for programmers. MCP is a standard protocol for AI applications to discover and use capabilities, and MCP servers often wrap APIs.ComparisonVector Database vs Traditional DatabaseTraditional databases answer exact, structured questions; vector databases answer "most similar" questions over embeddings. Many systems need both, and some databases do both.ComparisonLocal AI vs Cloud AI: Privacy, Cost, CapabilityLocal AI keeps data on your device and avoids per-use fees but is limited by your hardware; cloud AI gives access to the most capable models with no setup but means sending data to a provider.ComparisonFunction Calling vs MCPFunction calling is a model-API feature for requesting structured calls to functions you define. MCP is a protocol for packaging, sharing and discovering tools across AI applications. They work together.ComparisonAI Framework vs Direct Model APICall the model API directly when the flow is simple and you want full visibility; use a framework when shared orchestration, tools or retrieval abstractions clearly reduce risk and duplication.ComparisonREST versus GraphQLREST exposes resources over HTTP verbs and status codes. GraphQL exposes a schema and lets the client name fields. REST is the default for most services. GraphQL pays off when many clients need different shapes of the same data.ComparisonSQL versus NoSQLSQL databases store related rows with constraints. NoSQL is a loose label for documents, key-value stores and others. Choose based on access patterns and the invariants you need, not on which word sounds newer.ComparisonPRM vs ORM: Process vs Outcome Reward ModelsProcess reward models score each reasoning step; outcome reward models score only the final answer. The right choice depends on label cost, verifiability and how much error localisation you need.ComparisonReasoning Models vs Standard LLMsReasoning models deliberate before answering; standard models answer in one fast pass. Choose by task difficulty, latency tolerance and cost.ComparisonStructured Output Methods ComparedPrompt-only JSON, JSON mode, schema-enforced outputs, function calling and constrained decoding give different reliability, flexibility and portability. Choose by how bad a parse failure is.ComparisonCNN vs Vision TransformerCNNs bake in image structure and are data- and compute-efficient; Vision Transformers learn it and scale well with data. Pre-trained availability often decides.ComparisonTransformers vs State Space ModelsTransformers keep full access to the token history through attention; state space models keep a compressed state. The trade-off is exact recall versus linear-time efficiency.ComparisonEncoder-Only, Decoder-Only and Encoder-Decoder ModelsTransformer models come in three layouts: encoders for understanding, decoders for generating, and encoder-decoders for transforming one sequence into another.ComparisonLoRA vs Full Fine-TuningLoRA trains small adapters on a frozen model; full fine-tuning updates all weights. LoRA is cheaper and easier to manage; full fine-tuning can fit large shifts better.ComparisonA2A vs MCPMCP connects an AI application to tools, resources and prompts; A2A connects independent agents to one another. They are complementary layers, not rivals.ComparisonAI Agent Frameworks ComparedAgent frameworks differ in abstraction level, control, language and ecosystem. Compare by what you need — state control, retrieval, multi-agent, enterprise language — not by popularity.ComparisonLocal and Self-Hosted LLM Runtimes Comparedllama.cpp, Ollama, vLLM, ONNX Runtime and Hugging Face Transformers serve different needs: personal local use, multi-user GPU serving, portable deployment or general experimentation.ComparisonPublic Benchmarks vs Task-Specific EvalsPublic benchmarks give comparable numbers across models; task-specific evaluations measure fitness for your actual use. You need both, in that order of decision weight.ComparisonDPO vs RLHFRLHF learns a reward model and optimises the policy with reinforcement learning; DPO optimises the policy directly on preference pairs. DPO is simpler; RLHF is more flexible and on-policy.ComparisonMigrating from AutoGen or Semantic Kernel to Microsoft Agent FrameworkA practical map for moving AutoGen or Semantic Kernel code to Microsoft Agent Framework: what stays the same, what changes (teams become workflows, FunctionTool becomes @tool), and how to migrate in safe steps.

Reference

GlossaryAI Glossary A–Z143 terms defined in plain English.