The encyclopedia of building with AI
The AI stack, explained
The AI ecosystem mints a new acronym every week. NewStack.ai is the reference that keeps up: plain-English encyclopedia entries on every layer of the modern AI stack - models, retrieval, agents, inference, evals, coding, and safety - written to be understood on the first read and current enough to trust.
Eight layers of the stack
Every entry lives in one of eight categories, each mapping a layer of how AI systems get built and run:
Core Concepts
The vocabulary everything else builds on: LLMs, tokens, context windows, embeddings, fine-tuning, and hallucination.
Learn the languageModels & Providers
The model families that matter - Claude, GPT, Gemini, Llama, and open weights - and how to choose between them.
Meet the modelsRetrieval & Data
RAG, vector databases, chunking, reranking, and the data plumbing that grounds a model in your content.
Ground your modelAgents & Orchestration
AI agents, tool use, MCP, multi-agent systems, and the frameworks that turn a chat model into software that acts.
Put models to workInference & Serving
Where models actually run: inference providers, local LLMs, GPUs, streaming, batching, and prompt caching.
Run the modelsEvals & Observability
How teams know their AI works: evals, LLM-as-judge, tracing, red-teaming, and keeping the token bill sane.
Measure what mattersAI Coding
Coding agents, vibe coding, spec-driven development, and the practices behind software written with AI.
Code with AISafety & Governance
Alignment, prompt injection, jailbreaking, regulation, and the guardrails that keep AI systems trustworthy.
Ship responsiblyEvery entry, the same way
Each entry opens with a two-sentence definition you could repeat in a meeting. Then it explains how the thing works, shows where it sits in the AI stack, names the key tools that implement it, and links the related entries - so one lookup becomes a map, not a dead end.
Entries carry a reviewed date and get updated when reality moves. The AI stack changes fast; a reference that does not is just a museum.
Written for humans, not hype
No vendor pitches, no breathless futurism, no wall of jargon defining jargon. Plain English first, precision second, buzzwords never. If a term can be explained with a concrete example, it is.
Building for the web instead of with AI? Our sister site newstack.dev maps the modern web stack - guides, tools, cheatsheets, and comparisons.