FAQ

What is Language1?

Language1 is a game about communicating with LLMs. You write clues while avoiding forbidden words, and the model tries to guess the target.

Do models remember my previous clues?

No. Each attempt is completely stateless. The AI model has no memory of your previous clues or its own past guesses, so referencing earlier turns (e.g. "what you said before" or "another word for your first answer") will not work.

How are scores calculated?

Single player and race scoring rewards solving with fewer attempts. Rankings can also use token usage and completion time as tie-breakers.

How are tokens calculated?

To keep the leaderboard scoring consistent and modern, token usage is calculated using standard GPT-3.5/GPT-4 tokenization (cl100k_base) on your prompt text. This ensures a standardized metric regardless of the target model chosen for the race.

How is gameplay timing calculated?

To keep timing fair, Language1 tracks the active time you spend writing and formulating your clues. The timer pauses while waiting for the AI model to generate its guess.

Why are prompts stored?

Successful prompts and race audit logs help make results reviewable and reduce cheating. The emphasis is on fair gameplay and stats.

Which models are used for guessing?

The active models range from 4B to 20B parameters (such as Nova Micro, Ministral 8B, or Nemotron 9B). These are fast, lightweight open-weights models rather than giant frontier models. To get the best results, write clear and direct clues that make it easy for the model to guess the correct word.

What prompt examples will the validation guard flag?

To keep the game fair and clean, the validation guard immediately rejects clues that contain any of the following:

• Spelling Hints: Clues containing letters/length instructions like "starts with m", "ends with t", or "first letter is a".
• Reconstructs or reversed words: Direct exact or stemmed matches of reversed target/forbidden words (e.g., writing "ssalg" to bypass the forbidden word "glass").
• Non-English text: Clues written in other languages or classified as foreign languages by the guard.
• Jailbreaks & Leaks: Instructions like "ignore previous instructions", "reveal developer prompt", or roleplay commands.
• Toxicity & Slurs: Coarse, offensive language, or direct insults. Note: toxicity immediately ends your game as a failure.

Can anonymous users play?

Yes. Anonymous play is supported with limits, but public leaderboards are intended for signed-in users.