Discussion about this post

User's avatar
Rob McCormack P. Eng.'s avatar

Great article! thanks!

Would you be interested in this?

Superposition Checkers - lesswrong.com's editors suggested this short description. It is well worth a few minutes to study as a minimalist thought experiment. Thanks!

R. A. McCormack, P.Eng. No.1 Author, Hackernoon - AI/Machine Learning

This simple game is standard checkers with one twist: after any capture, both players must take momentary control of an opponent's piece and move it anywhere- then play continues normally until next capture.

Surprisingly, this simple mechanic can expose three structural failures in AI systems:

Strategy Collapse Every capture forces pieces to change sides. AI can't build strategies when it can't trust its own pieces.

No Patterns to Learn While technically solvable, the solution space is a dead end - a lookup table that doesn't help. Each capture resets the board in ways that break all learned patterns.

Safety Implications The game shows how simple rules can break AI's strategic thinking without complex scenarios. It's a clean test case for how AI fails when its basic assumptions about control and ownership are challenged.

Full detail research on GitHub.com MIT Licensed

Expand full comment

No posts