Applied AI — Session 3: The language of machines

✂️ The live tokenizer

Type text: watch the machine cut it into tokens , each with its own number. This is exactly the first step when you talk to an AI.

The cat eats a pizza. 🐱🍕 Unconstitutionally! Invented word 🧪 With an emoji 😀

0characters

0words

0tokens

—characters / token

💡 Observe: common words often fit into 1 token, long or rare words are cut into several pieces. Invented words explode into little pieces! Each number displayed under the token is what the model “sees” for you.

⚠ This tokenizer is a educational simulation : real AI models each have their own division and their own numbers, learned automatically. The principle is the same.

🗺️ The embeddings explorer

Each word becomes a point on a map of meaning . Words close in meaning are neighbors. Grab the magnifying glass 🔍 and move it : it shows you the closest words. You can also move the words themselves!

Animals Food Royalty Emotions School

✨ The magic equation of words

On the map, we can make math with meaning ! Try the most famous: king − man + woman = ?

− +

⚠ This arithmetic comes from the first word models (word2vec, 2013). Modern AIs use much more complex maps — but the idea “direction = position” remains true.

🌡️ The temperature lab

The AI chooses each next token based on probabilities . The temperature regulates chance: 🧊 low = it almost always takes the most probable. 🔥 high = crazy options have their chance. Same question, different cursor → different answer!

1. Choose a question:

2. Set the temperature:

🧊 🔥

0.7

The answer will appear here, token by token…

🔬 In the head of the machine: a live choice

Here are the probabilities of last interesting choice of generation, after adjustment by temperature. The orange bar = the token that won the draw.

Generate a response to see a live choice.

💡 Experience to do: generate 3 times in a row at temperature 0 → the answers are (almost) identical. Then 3 times at temperature 2 → each answer is different… and sometimes completely absurd! This is exactly exercise 3 of the worksheet.

⚠ Educational simulation: a real AI chooses from tens of thousands of tokens at each step, dozens of times per sentence. Here we simplify to see the mechanism. In most consumer chatbots, the temperature is not adjustable — it is set by the service.