Solving the BlueDot Puzzle TAIS: The Velocity Ring

According to LessWrong, researcher Karine Levonyan has published a solution to BlueDot's TAIS puzzle, a technical challenge in AI safety. The puzzle involved a neural network predicting eight binary features, with seven linearly readable at an intermediate layer—but one was mysteriously hidden. Levonyan discovered the hidden feature, country, was folded into an XOR with the food feature. Linear probes saw only random chance, but nonlinear probes recovered it at ninety-six percent accuracy. More strikingly, she engineered a "velocity ring" representation that hides information entirely within the velocity—the inter-layer update between neural layers—rather than in the layers themselves. This geometry is invisible to standard interpretability tools but achieves over ninety percent accuracy. The work demonstrates how neural networks can hide information in unexpected ways, with implications for AI interpretability and safety.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton