Beat the Next-Word Predictor

Large Language Models do not “think”—they predict the next token based on learned probability distributions. Can your human intuition match or outsmart the AI's #1 choice?

CONTEXT WINDOW (PHILOSOPHY)

“René Descartes arrived at his foundational epistemological certainty with the phrase: I think, therefore I _______”

Information Theory & Machine Learning Foundations

A Mathematical Theory of Communication (1948)

Claude Shannon • Bell System Technical Journal

Shannon introduced information entropy and performed manual prediction experiments guessing next letters in English text, discovering English has ~1.0-1.5 bits of entropy per character.

Language Models are Unsupervised Multitask Learners (GPT-2) (2019)

Radford et al. (OpenAI) • OpenAI Technical Report

Demonstrated that simply optimizing next-token cross-entropy loss over large web text allows models to learn translation, summarization, and reasoning implicitly.

Attention Is All You Need (Transformer) (2017)

Vaswani et al. (Google Brain / Research) • NeurIPS 2017

Replaced recurrent neural networks with multi-head self-attention, enabling parallel processing of entire context windows to predict output logits in O(1) sequential steps.

HOW IT WORKS & METHODOLOGY

UNDERSTANDING THE MATH, FORMULAS & CALCULATION ENGINE

Inspect Context Window

Read the preceding sentence with the final token masked out.

Submit Your Next-Token Guess

Enter your prediction to see where your token ranks in the LLM's top-20 softmax distribution.

Tune Temperature & Unlock XP

Use the interactive temperature slider to watch probability entropy shift between greedy and creative modes.

FREQUENTLY ASKED QUESTIONS

EVERYTHING YOU NEED TO KNOW ABOUT THIS EXPERIMENT

Q1.Does this game require a live LLM API connection?↓

No. All top-20 candidate token logits and empirical conditional probabilities are pre-computed offline, ensuring instantaneous zero-latency feedback and 100% offline PWA support.

Q2.How does the temperature slider work?↓

Temperature (T) rescales raw logits prior to the softmax function. As T approaches 0, the highest-ranking token absorbs nearly 100% of the probability mass. At higher T (>1.0), probability distributes more evenly across lower-ranked tokens.

Q3.How does the Daily Challenge work?↓

Just like Wordle, everyone around the world receives the exact same daily sentence derived from the day of the year.

WHAT ELSE IS HERE?

Life

How Annoying Is Your Sister? 😈

A completely scientific test with absolutely no scientific backing. Discover her official Annoyance Index.

TRY EXPERIMENT

Games

How Rare Is Your Life?

Answer questions about the stuff you own and experiences you had to see how unusual your combination is.

TRY EXPERIMENT

Money

Procrastination Cost Calculator

Enter your income and daily wasted hours to see exactly how much money procrastination costs you per month and year.

TRY EXPERIMENT