Skip to content
How AI works

How modern AI works, one moving part at a time

Short, animated explainers of the ideas inside today's AI systems, from a single neuron to transformers, vector search and agents that call your tools. 48 topics, each with the original papers.

The big picture

The journey of one prompt

Everything below is a part of this loop. Watch a request travel through it, then open any stage to see how it works.

1

Your prompt

“Summarise this contract in 3 bullets”

Tokenisation
2

Tokens

Sum9538mar3876ise1082 this420 contract5376 in304 3220 bullets34810
Tokenisation
3

Vectors + position

Sum
mar
ise
this
contract

+ position signal

Embeddings
4

Attention layers

Self-attention
5

Next-token odds

The
46%
This
22%
Key
17%
In
9%
Sampling
6

Pick and repeat

Answer so far: The

Append, then run again for the next token, reusing cached attention.

KV cache
token ids and numbers are illustrative
Chapters

Pick a chapter, or read in order

Each topic is a few minutes: an interactive visual, a plain explanation, why it matters for a product, and the primary sources.

01

Neural network basics

Neurons, layers, loss and how a network learns from its mistakes.

02

Language as numbers

How text becomes tokens, tokens become vectors, and meaning becomes distance.

05

Speed and cost

Experts, quantisation and serving tricks that make large models affordable.

06

Training and alignment

Pre-training, fine-tuning and teaching a model what people actually want.

07

Retrieval and search

Keyword search, vector search, HNSW graphs and grounding answers in your data.

08

Agents and protocols

Models that plan, call tools and talk to your systems through MCP.

09

Observability

Traces, evals and monitoring: how you know a system works, in the lab and in production.

10

Interpretability and safety

Looking inside the model to see what it is doing, and why that matters.

basics intermediate deep dive
Two more ways in

A roadmap and a map

The roadmap turns the chapters into tracks with modules and a progress tick per topic. The topography draws every chapter, topic and term as one connected map.

Want these ideas working on your data?

We build the retrieval, agents, voice and model systems these explainers describe. Start with a free discovery call.