This platform consolidates several mini games inspired by recent projects. Quick overview below:

SUDOKUBE

Sodokube stems from a project exploring the possibility of sudokus in dimensions greater than 2. 4x4x4 grids are easy to conceptualize and computing all solutions isn’t too compute intensive, which led to this game.

The version of the game published last year was a bit too easy (solvable in <5 mins) so we used the opportunity of the port to games.pitti.io to remove the easier levels. To make things more challenging, assisted navigation in the 3D space is disabled at the highest level.

SATIRE

Satire stems from a recent project on classification of ideological and political biases in LLMs. It appeared that AI struggles to judge satirical content. But the task is also surprisingly hard for humans.

The game is an adaptation of a tool we built to explore and annotate the original Speechmap dataset. To avoid being influenced by the model name or the LLM judges classifications, these informations are hidden until a feedback is given by the annotator.

WORDLE BATTLES

Wordle Battles is derived from experimentations with Reinforcement Learning reward functions to train models to play Wordle.

The function that calculates the information content of a guess (Shannon entropy) can be used as a backbone for an algorithmic Wordle solver. The objective of the Wordle Battles game is to solve the puzzle in fewer guesses than the algorithm

Technical details: Typescript app with Node backend (better-sqlite database). Monorepo hosted on Hetzner