Show HN: TrickAI – Uncovering AI Blindspots (trytrickai.com)

During a hackathon last week, my team built TrickAI - a project exploring AI's limitations in problem-solving compared to humans.

The core question is this - using puzzles / context clues, could humans send information to other humans without ChatGPT learning the information?

You can try to solve puzzles created by other humans that ChatGPT failed to solve. You can also create a puzzle, and if it fools ChatGPT, it'll become available for humans to try.

The leaderboard shows which puzzles (1) fooled ChatGPT, and (2) did not fool humans.

Our project aims to compile a dataset of AI blindspots. This dataset will be freely available to anyone working on AGI alignment, helping to enhance the understanding of AI limitations and improve alignment efforts.

Would love to hear any feedback/advice anyone has and see if I should invest a bit more in this. Thanks!