How does a large language model work?

It continues a text by computing step by step which word piece fits next and choosing one. It learned the feel for this from a vast amount of text during training.

What does temperature mean for ChatGPT and other AI models?

Temperature sets how often less likely words are chosen. Low gives reliable, uniform answers; high gives more creative and sometimes nonsensical ones.

Why do language models sometimes invent facts?

They write what fits their patterns and do not check whether it is true. Asking for sources, providing your own material and double-checking what matters helps.

Does the app run a real language model?

No. The candidates and their weights are prepared by hand so the idea is clearly visible. No input is sent to an AI service.

All apps

The word oracle: how a language model writes

About this app

A sentence grows word by word: at its end the candidates for the next word float as spheres of light, the larger, the more likely. A courage dial (temperature) shows how the same beginning continues reliably or imaginatively, threads of light show attention to earlier words. An ocean of pages explains where the feel for language comes from.

With explain mode: an AI tutor walks you through the app, points out what matters in the scene and answers your questions.

How a language model writes

Large language models such as ChatGPT or Claude write their answers word by word, more precisely word piece by word piece. At each step they compute which continuation fits the text so far and pick one. Their feel comes from training on a vast amount of text. What remains are patterns: which words fit together and how sentences are built. Whether a statement is true is not checked while writing.

What the app shows

A sentence grows word by word. At its end the candidates for the next word float as spheres of light; the bigger and brighter, the more likely. Threads of light show which earlier words the model attends to, for example from "she" back to "doctor". An ocean of pages stands for the texts it read. The candidates are prepared by hand; no real model is running.

What you try out

Choose one of six sentence starts, draw word by word or tap a sphere yourself. With the courage dial (temperature) you see how the same start continues identically when careful and differently each time when bold. Using opening hours as an example, the app shows why sounding right is not the same as being true, and how a source you provide changes the answer.

Frequently asked questions

How does a large language model work?: It continues a text by computing step by step which word piece fits next and choosing one. It learned the feel for this from a vast amount of text during training.
What does temperature mean for ChatGPT and other AI models?: Temperature sets how often less likely words are chosen. Low gives reliable, uniform answers; high gives more creative and sometimes nonsensical ones.
Why do language models sometimes invent facts?: They write what fits their patterns and do not check whether it is true. Asking for sources, providing your own material and double-checking what matters helps.
Does the app run a real language model?: No. The candidates and their weights are prepared by hand so the idea is clearly visible. No input is sent to an AI service.

Subject: Understanding AI

More from Understanding AI

A VisuApp by heyprof: interactive, free in your browser, no sign-up. What is a VisuApp? · All apps