OpenSpec, the spec-driven workflow it imposes on an AI coding agent, what each artifact file contains, and how it fits into a production harness.

The production-engineering scaffolding around a hosted model: retries, structured outputs, tool dispatch, caching, observability, evals, and prompt-injection defense.

How to build a Plan-Execute-Reflect agent in Go, with a SQLite-backed memory store, a reflector that catches confident-hallucination failures, and the rough edges that show up at hundreds of steps.

A working ReAct agent in Go: the loop, concurrent tool execution, and a streaming event channel for plugging it into a CLI or HTTP server.

When fine-tuning is worth the effort over prompts and RAG, the practical flavors (LoRA, QLoRA, RLHF), and a working recipe in code.

Why asking a model to think step by step measurably improves accuracy on reasoning tasks, and how that intuition led to today's reasoning models.

The patterns inside a prompt that change model behaviour the most, plus an interactive demo of how a few words shift the next-token distribution.

Retrieval-augmented generation past the basic pattern: chunking, hybrid retrieval, re-ranking, HyDE, and the failure modes that show up in production.

How a model sees its context window, why position inside the prompt affects what gets used, and tactics for laying information out so the model can use it.

Why language models confidently make things up, what is happening at the token level when they do, and the prevention techniques worth using.

What the billions of numbers inside a language model are, how training writes them, and what they can hold versus what has to come from context.

The optimisation algorithm that trains every neural network, from the math up, and the modern variants (SGD, Adam) that show up in production training.

The matrix math that runs inside a language model on each forward pass: embeddings, attention, and the softmax over the vocabulary.