How this site is built
What do the interface, storage, language model, and sandbox do? Two concrete examples show how this site works.
Understand AI through an example
Imagine reading about RAG and wondering: why does it need embeddings? This site uses drawings and concrete examples to make questions like that easier to understand.
An interface you can use
Next.js builds the interface: posts, the video player, and the box for your questions. You can watch a video, for example, and ask about the example it shows.
Where does the video come from?
Cloudflare runs the site. The video file lives in R2, its media storage. When you press play, the file loads from there. The interface controls the player; the storage supplies the file.
How AI uses the posts
Your question goes through OpenRouter to a language model. The assistant can use tools to search for posts and read them. For the question about embeddings, it can use this site’s explanation as context for its answer.
Actually running the code
When you ask for a code example, the AI can run Python in an E2B sandbox. That is a separate test environment. This code, for example, adds one, two, and three. The AI gets the actual result back: six.
Giving drawings a voice
A video needs pictures and sound. Remotion turns the drawings into a timed animation. ElevenLabs generates the narration using a clone of my voice. Both come together in a film, which we then review.
Understand by trying
You can watch an explanation, ask a specific question, and test an example. The goal is to turn an abstract term into something you really understand.
Production: cost, time, and footprint
API cost estimate
≈ $0.16
Recorded job time
8.9 min
Local CO₂e estimate
≈ 1.9 g
Figures cover recorded automated production steps for this language, including both video themes. Job durations are added together; parallel jobs do not equal elapsed working time. Text development, programming, human work, hosting, and unrecorded attempts are excluded.
Speech: published characters × reference rate ($0.08 / 1,000 · ElevenLabs). 18 recorded review requests. CO₂e covers local rendering only: 4.7 Wh · 35 W/job · 400 g/kWh. Assumed power and grid intensity; cloud model inference and hardware manufacturing are not included. Method.
My notes
Notes and highlights for this post
Stored only in this browser. Nothing is sent.
Your notes and highlights appear here.
Select a passage and choose “Highlight” or “Note”. Or write a note about the whole post.