How AI Actually Works

The maths. The intuition. And why every generative model is the same idea wearing different clothes.

PLAIN ENGLISH TECHNICAL BOOK

How a Transformer Produces One Token

Follow one prompt through tokenisation, attention, inference, model formats, and a working provider implementation.

The Transformer

GPT, Claude, Gemini, LLaMA — all built on this.

Attention Variants, LM Heads & Tokenisation

MQA, GQA, FlashAttention, weight tying, BPE vs SentencePiece — the implementation details.

Dense vs. Sparse vs. Mixture-of-Experts

Why some models only wake up a fraction of their parameters per token.

Diffusion Models

Stable Diffusion, DALL-E, Midjourney — the other architecture.

Determinism, Seeds & Sampling

Why the same prompt gives different outputs.

Beyond Transformers & Diffusion

GANs, video models, audio models — the rest of the landscape.

Finetuning, LoRA & PEFT

How you adapt a model without training from scratch.

The Wavefunction Metaphor

They're all the same idea. Here's why.

Inference: The Vocabulary of Running AI

Tensors, layers, quantisation, and why the numbers matter.