A series in 20 parts
The World of AI
How large language models actually work, and how they are used in real products. Written for engineers, founders and IT managers who know software well but have no machine learning background. No maths beyond arithmetic, and one running example from start to finish: an AI agent that fixes a login bug and opens a pull request.
Part 1
How an LLM works
01What an LLM actually is
3 Nov 2025A machine that predicts the next piece of text, and why that turns out to be enough to fix a bug.
02Tokens and tokenization
21 Dec 2025How text gets chopped into pieces before a model sees it, and why those pieces decide what you pay.
03Embeddings
9 Jan 2026How a model turns tokens into lists of numbers, and why meaning becomes distance.
04Weights and neurons
15 Jan 2026One neuron calculated by hand, and what it really means when a model has 70 billion of them working together.
05The transformer and attention
17 Feb 2026How every token gets to look at every other token, and why what you put in the context decides what the model can do.
06Training
18 Mar 2026How billions of random numbers become a model that knows how to fix a login bug.
07From raw model to assistant
14 Apr 2026The extra training that turns a text predictor into something that follows instructions.
08Generation and sampling
26 Apr 2026How the model turns a list of probabilities into one token, and why your answer arrives word by word.
Part 2
Running models
09GPUs explained
4 May 2026Why graphics chips run AI, and why memory speed matters more than raw calculating power.
10Inference at scale
6 May 2026How one GPU serves hundreds of users at once, and why speed for you and speed for everyone are different numbers.
11Proprietary vs open-weight models
8 May 2026Rent a model through an API, or download one and run it yourself. What you gain, what you take on, and what the licence actually lets you do.
12Running models locally
17 May 2026How much memory a model really needs, how quantization shrinks it, and when a laptop tool stops being enough.
Part 3
Making models useful
13Context windows and prompt engineering
24 May 2026What fits in the model's view, how to arrange it so the model uses it well, and how caching stops you paying for the same text twice.
14RAG from first principles
30 May 2026How to give a model the right pages to read, instead of hoping it already knows them.
15Fine-tuning
11 Jul 2026Changing how a model behaves by training it further on your own examples, and when that is the wrong tool.
16Tool use, agents and MCP
30 Jul 2026How a model that only writes text ends up searching code, running tests and opening pull requests.
Part 4
Running AI in production
17The LLM gateway
6 Oct 2026One front door for every model call: routing, fallbacks, caching, budgets, logging and privacy in one place.
18Evals and observability
7 Oct 2026How to know whether your AI system is any good, and how to see what it actually did.
19Security and failure modes
8 Oct 2026Hallucination, prompt injection, data leakage, and how to box in an agent so mistakes stay small.
20Compliance and strategy
9 Oct 2026GDPR, the EU AI Act, build versus buy, and what an AI system really costs.