SerpkaiSERPKAI

SERPKAI LAB · EXPERIMENT 01

How an AI thinks, step by step

A language model does not understand sentences the way you do: it splits them into pieces, turns them into numbers and works out which word is most likely to come next. Pick a sentence and look inside.

← Serpkai Lab

01

You type a sentence

02

It is split into tokens

The model does not read letters or words: it reads tokens, frequent chunks of text. Common words stay whole; rare ones are split. Each token has a fixed number in the vocabulary.

03

Each token becomes a point

Each token is turned into a list of hundreds of numbers (an embedding). Here we flatten it to two dimensions: words with similar meanings end up close together.

04

Attention: who looks at whom

Before deciding, each token “looks” at the previous ones to understand the context. Hover over (or tap) a token: the thicker the line, the more it weighs.

05

The next word

In the end you get a list of candidates with their probability. Temperature sets how much risk it takes: low, almost always the favourite; high, the bars even out and stranger things come out.

0.8

Note: there is no real model inside this page. Tokens, the map, attention and probabilities are simulated to behave like a real model, but the specific numbers are illustrative.

Why it matters if you run a website

AI assistants do not “search” like a person: they complete text based on what they have seen and the sources they consult. That is why whether your brand shows up in their answers depends on how and where it is talked about. At Serpkai we measure it: see how in AI visibility.

Check whether AI mentions your brand

Track your Google rankings and your presence in AI answers from a single dashboard.

Try it free
ES · Español