The Lab

AIs don't read letters or whole words. They read tokens: chunks of text. Type anything and watch it split.

A rough split, to show the idea. Each real AI cuts text with its own dictionary, so its counts differ a little.

For the next word, an AI has a list of guesses, each with a chance. Temperature decides how adventurous its pick is.

Made-up chances, to show the idea.

An AI can only hold so much of a chat at once: its context window. Add messages and watch the oldest ones fall out.

A tiny window, to show the idea. Real windows hold far more, but they still fill up.

Some emails hide words you can't see, written for the AI that reads them. Switch views to see what your AI sees.

An AI always has an answer ready, even when it doesn't know. Ask it something, then let it look things up first.

Made-up questions, answers and shop, to show the idea.

Teams test an AI on real questions before it goes live, and again after every change. Run the test, update the AI, then run it again.

A made-up shop and test, to show the idea.

Some of these reviews were written by an AI, with no real buyer behind them. Tap the ones you think are fake, then check.

Made-up reviews and kettle, to show the idea.

A chat AI answers in sentences. A typed model, like Jev, fills in a form instead: only the answers you allow, each with how sure it is.

Made-up messages and chances, to show the idea.