A Hidden Smiley Face Lurking Inside Neural Networks

willdepue · x · 2026-07-14

The author ran a fascinating experiment: training a neural network with standard initialization, sorting its weights by value to form a "smiley face," and then restoring the original weights post-training. Surprisingly, the smiley face persisted.

This suggests that neural networks might inherently contain a "will depue / smiley" structure, proving that internal representations can be far weirder and more interesting than our intuition suggests.

Original post →

More from Fun

Fun channel →