Skip to content
Welt am Draht
DE
Artificial intelligenceDemonstration article

World on a wire

How a language model picks the next word: a short technical approach to probabilities, temperature and the illusion of certainty.

By Editors (demo) · · 1 min read

Black-and-white photo: a long aisle between server cabinets under a grid ceiling, a silhouette at the end.
Computing on a grid: infrastructure that is rarely shown.Image: Original AI-generated concept image made for this prototype

Core claim & position

Core claim
What looks like knowledge is, first of all, a weighted choice among probabilities.
Position
Anyone evaluating model output should separate linguistic confidence from factual correctness.

At its core, a language model produces a number for every possible continuation. Those numbers are turned into probabilities, from which a choice is made.

01From raw scores to probabilities

The function below shows the principle in simplified form. A parameter called temperature controls how strongly the model concentrates on the most likely option.

src/sampling/softmax.ts, TS, 11 lines
export function softmax(logits: number[], temperature = 1): number[] {  // Scale, then subtract the max for numerical stability  const scaled = logits.map((l) => l / temperature);  const max = Math.max(...scaled);  const exps = scaled.map((s) => Math.exp(s - max));  const sum = exps.reduce((a, b) => a + b, 0);  return exps.map((e) => e / sum);} // Low temperature: more decisive. High temperature: more spread out.const probabilities = softmax([2.0, 1.0, 0.1], 0.5);
Simplified softmax with temperature (teaching example, not optimised)

02A filmic study

Globe study. Twelve seconds of a slow approach to a wire globe. Original silent video created from an AI concept image.

03Where uncertainty begins

A high probability for a word does not mean the statement is true.[1]

Notes

  1. [1]Editor’s note: a real publication would cite a verifiable source here. For this demonstration article, no source has been invented.

Further reading