WDEV Lab

Context

The idea is to learn AI by experimenting rather than reading: no Python to install, no environment to set up, and every concept paired with something that runs immediately. Each lesson should reveal something the learner could only guess at before.

Problems & solutions

1. 01

   Measuring the "Vietnamese tax" with a real tokenizer

   The first lesson tokenises Vietnamese text with the o200k_base tokenizer right in the browser. The sample Vietnamese sentence comes to 13 tokens for 40 characters; the English sentence with the same meaning comes to 8 tokens for 45 — 1.63× more expensive despite being shorter. The page states plainly that this is OpenAI's tokenizer, because Anthropic doesn't publish Claude's.

2. 02

   Calling a real API with no server

   Lessons that call a model use the learner's own API key: it's stored in localStorage and goes straight from the browser to Anthropic through the official SDK. With no server in between, infrastructure costs nothing and there's nothing to abuse.

3. 03

   A 2.5 MB token table that doesn't slow the page down

   The tokenizer's vocabulary is imported dynamically, only when the learner opens lesson 01, with a clear loading state.

4. 04

   A cost calculator that reflects how APIs actually bill

   Each turn has to resend the whole conversation history, so cost grows faster than the number of turns. The calculator models that, shows how much prompt caching saves, and applies the Vietnamese multiplier from lesson 01.