How AI actually works
Long-form tutorials for concepts that need more than a definition. Diagrams, worked examples, working code, and links to primary sources throughout.
Long-form tutorials for concepts that need more than a definition. A term page tells you what something is. A concept page walks you through how it actually works — with diagrams, worked examples, code, and links to primary sources.
Every concept tutorial is written for readers who want depth: you finish with a working mental model, not just vocabulary. Perfect for developers implementing a technique, executives who need to understand what their team is doing, and students moving from "I recognize the word" to "I can explain it."
Deep-dive concept tutorials
Long-form, diagram-heavy, code-included. Each takes 15-30 minutes to read.
How Transformers Work
The full walkthrough — embeddings, positional encoding, self-attention, multi-head attention, feed-forward networks, layer norms, residual connections. With diagrams and code.
RAGHow RAG Works (End-to-End)
Chunking strategies, embedding choice, vector store selection, retrieval, reranking, prompt construction, and evaluation. Real code you can adapt.
AgentsHow Agentic AI Works
The agentic loop explained: tool selection, execution, observation, planning. Compare ReAct, planner-executor, and multi-agent patterns.
TrainingHow Fine-Tuning Works
From SFT through RLHF, DPO, and GRPO. When to use each, what data you need, what results to expect.
LLMsUnderstanding Context Windows
Why longer isn't always better. "Lost in the middle" phenomenon, effective context length vs. max, and how to structure prompts for long contexts.
Agents · MCPHow MCP Works
Servers, clients, transports, and resources. The protocol behind the AI plumbing standard, with examples of building and using an MCP server.
Browse concept tutorials by topic
Architectures
Transformers, MoE, State Space Models, Mamba, RWKV.
Training & fine-tuning
Pretraining, SFT, RLHF, DPO, GRPO, LoRA, QLoRA.
RAG & retrieval
Chunking, embeddings, hybrid search, reranking, eval.
Agents & tools
Tool use, MCP, ReAct, multi-agent, computer use.
Prompt engineering
CoT, few-shot, DSPy, prompt injection defenses.
Evaluation
Benchmarks, LLM-as-judge, human eval, eval harnesses.
Inference & serving
vLLM, PagedAttention, KV cache, prompt caching.
Multimodal
Vision-language models, diffusion, video generation.
Safety & alignment
Constitutional AI, red teaming, interpretability.
Concept tutorials by depth
Start here
Questions about Concepts
A term page defines something in about 1,000 words with examples. A Concept page is a 3,000-5,000-word tutorial with diagrams, working code, and step-by-step walkthroughs. Terms are for reference. Concepts are for learning.
No — they are teaching-optimized. Production systems need error handling, retries, security, and evaluation that our examples do not include. See our disclaimer on code examples.
Yes. Every technical tutorial is reviewed by an AI engineer with hands-on experience in that area before publication. Reviewer names appear on the byline for each tutorial.
For non-commercial classroom use with proper attribution, yes. For commercial course material or embedded licensing, contact us.
Full re-review every 6 months. Substantial updates whenever the underlying technique changes materially (new fine-tuning methods, updated APIs, etc.). See our update cadence.