ai-explained

v56.0 · 57 demos · 13 narrated walkthroughs

How modern AI actually works.

Interactive walk-throughs of the math behind transformers, agents, and everything else. Click, drag, type — the goal is that you walk away understanding one thing you didn't before.

live · 8×8 causal attention try it →
thecatsatonthemat—•
softmax(QKᵀ/√d) per row · lower triangle = causal mask
113
articles
57
interactive demos
13
animated walkthroughs
7h 45m
of reading

where do you want to start?

Pick your route in.

Four tracks, four reasons to be here. The whole site is structured around your starting point — not ours.

1
"I want to understand how machines learn at all."
start here →
2
"I know ML basics. How does a transformer actually work?"
start here →
3
"I want to build with LLMs. Show me what works."
start here →
4
"I'm shipping AI. How do I keep it shipped?"
start here →

Or browse the full curriculum · articles · demos.

if you read three things on this site

Editor's picks.

The pieces that earn their length. Hand-picked from the 113-article backlog — not auto-ranked.

01
The Transformer Block
The keystone. If you read one thing on this site, read this — it composes self-attention, residual streams, and the MLP into the unit every modern model is built from.
4 min · 3 demos
02
RAG Fundamentals
Retrieval-augmented generation, end to end. The single most-deployed pattern in production AI right now.
5 min · 1 demo
03
Scaling Laws
Why GPT-3 was undertrained and LLaMA-3 trains 70B on 15T tokens. Chinchilla in plain language.
4 min · 1 demo

shipped recently

What's new on the site.

The site is still being written. Three additions worth a look if you've been here before.

new section
/case-studies →
Four full products built on the /ship stack — docs assistant, code-review agent, research assistant, support bot. Architecture, eval numbers, retrospectives.
/ship complete
Production release →
Steps 12–16 are live: observability with Phoenix, prod evals, cost/latency tuning, deploy targets, what's next. The whole 17-step curriculum is now shippable.
/articles polish
Companion-track links →
Every theory article now points at the /build, /ship, or /case-studies that turn it into running code. Click any RAG or agents article to see it.

when you click a demo, here's what runs

Three tiers of "real."

Every demo is labelled with one of three tiers, so you know what's actually computing what you're seeing.

tier A
runs a real model
The same model or algorithm a production system would run, live in your browser. Embeddings, attention, sampling, tokenizers — actually computing.
tier B
real algorithm, curated data
The algorithm is real; the data is hand-picked so the mechanism is visible without a 60-second wait. Most demos on the site are this tier.
tier C
illustrative scaffolding
A drawing, not a computation — used when the real thing is too slow or too large to run in a browser. Always clearly labelled.

Animated demos play through a curated story with voice narration; interactive demos hand you the dials. Both can be any tier.