Fundamentals01
What Is a Large Language Model?
Tokens, the transformer, attention and the training pipeline from pretraining to RLHF, explained without hand-waving.
Knowledge base · updated November 2026
Understand how modern AI actually works
Clear, engineering-grade explanations of large language models, the specialised models around them, and the tools used to build real systems on top.
- in-depth articles
- 10
- minutes of reading
- 92
- glossary terms
- 47
Explore by topic
Three ways into the field
Start with the fundamentals, meet the specialised model families, then learn how production systems are put together.
Fundamentals
How language models work under the hood: tokens, transformers, training and reasoning.
Browse 2 articlesModel types
Beyond chat: speech, document reading and decision models, and when to reach for each.
- Speech AI: How STT and TTS Models Work
- OCR and Document AI: From Tesseract to Vision-Language Models
- Decision Models: AI That Answers With Probabilities
Building with AI
Retrieval, agents, harnesses, automation and running models on your own hardware.
- Embeddings, Vector Databases and RAG
- AI Agents, Tool Calling and MCP
- Harnesses: Why the Model Is Not the Product
- Modern Automation with LLMs: Workflows, Tools and Human-in-the-Loop
- Running Models Locally: Quantization, vLLM, Ollama and LoRA
Library
Filter by topic and level
All articles
Each article takes 8–12 minutes and comes with diagrams, comparison tables and working examples.
Fundamentals02
Reasoning Models: How LLMs Think Before They Answer
Chain of thought, test-time compute, RL with verifiable rewards and thinking budgets: when a slower answer is a better one.
Model types03
Speech AI: How STT and TTS Models Work
Spectrograms, Whisper, word error rate, diarization, vocoders, codec-based TTS and the latency budget of a voice agent.
Model types04
OCR and Document AI: From Tesseract to Vision-Language Models
Classic OCR pipelines, Tesseract, layout analysis, Donut and VLMs, schema-based extraction and how to measure accuracy.
Model types05
Decision Models: AI That Answers With Probabilities
Jev, d1-ultra and the new class of models that return calibrated yes/no, choice and score answers instead of text.
Building with AI06
Embeddings, Vector Databases and RAG
Embeddings, chunking, vector search, hybrid retrieval and reranking: how RAG grounds answers in your own documents.
Building with AI07
AI Agents, Tool Calling and MCP
Tool calling, the agent loop, workflows versus agents, the Model Context Protocol, tool design and agent security.
Building with AI08
Harnesses: Why the Model Is Not the Product
The loop, tools, context handling and permissions around a model decide how well an agent works. How harnesses are built and judged.
Building with AI09
Modern Automation with LLMs: Workflows, Tools and Human-in-the-Loop
Patterns, architecture, structured outputs, human approval and reliability for LLM-powered workflows, with a support-triage example.
Building with AI10
Running Models Locally: Quantization, vLLM, Ollama and LoRA
Memory maths, quantization, llama.cpp, Ollama, vLLM and LoRA: how to run and adapt open-weight models on your own hardware.
Learning path
New to LLMs? Read in this order
Five articles that take you from "what is a token" to how coding agents are engineered.
- What Is a Large Language Model?Tokens, transformers and how training works.
- Reasoning Models: How LLMs Think Before They AnswerWhy models now think before they answer.
- Embeddings, Vector Databases and RAGGrounding answers in your own documents.
- AI Agents, Tool Calling and MCPLetting models act through tools.
- Harnesses: Why the Model Is Not the ProductThe software that turns a model into a product.