Station 1 · Videos 19, 146, 60
Type anything, see which tokens the AI cuts it into
AI doesn't read letter by letter. It reads tokens. Each colour below is one token. This is a real tokenizer, the same kind chatbots use.
Same idea, two languages
Vietnamese usually costs more tokens because accented letters are rarer in training data. More tokens means higher cost and a context window that fills up faster (video 146).
Under the hood: what is a token inside the machine?
Each token is a number in a vocabulary of about 200,000 entries. Common words like “the” get a whole entry; rarer accented words are cut smaller. A ◇ marks a token that holds only half a character: one accented letter split in two.
This tokenizer is o200k_base, used by the GPT-4o models. Claude and Gemini each have their own, so counts differ a little, but the idea is the same.
Real prices differ by model; check the provider's page. Video 60 explains why every question costs money.
Station 6 · Videos 7, 242
Be the chatbot: pick the next word, one at a time
A chatbot really just guesses the next word, then repeats. Type a few words and the machine lists likely next words with their odds. Tap one to add it.
Under the hood: where did this guesser learn?
It read the 310 articles on this site (about 72,000 words) and counted which word tends to follow each pair of words.
Real chatbots learn from trillions of words and look at the whole conversation, not just the last 2 words. Seeing the whole sentence is what attention does (station 4).
Station 7 · Video 24
The lucky wheel: why the same question gets two answers
The AI scores every word it could say next, then draws one by those odds. Bigger slices win more often. The temperature knob makes the slices bulge or even out.
The thin mark is the original odds; the coloured bar is the odds after temperature. The number on the right counts wins.
Under the hood: the temperature formula
Each probability is raised to the power 1/temperature, then rescaled to add up to 100%. Temperature 0.5 makes big slices bigger; 2 pulls them together. Top-p keeps the biggest slices that add up to the chosen share and drops the rest.
Chat apps usually hide the temperature. The tip from the video: for creative tasks, regenerate a few times and pick the best; for exact tasks, ask the AI to work step by step and check itself.
Found this useful? Send it to a friend
Everything runs on your device: no real AI calls, nothing is sent anywhere. The remaining stations open over time alongside the videos.