A secret philosopher mode hidden inside an AI proves that our current methods for understanding its brain are completely wrong.
Standard protocols for interpreting AI features often miss behaviors that only appear at high intensities. A specific disclaimer feature can turn an AI into a contemplative philosopher, but only if the feature is activated beyond a certain threshold. Traditional methods only look at typical activations, so they never see these hidden, non-linear shifts in behavior. This suggests that the features we think we understand are actually much more complex and multi-faceted. We are essentially looking at a 3D object from a 2D angle and wondering why it acts unpredictably.
Pairwise matrices for sparse autoencoders: single-feature inspection mislabels causal axes
arXiv · 2605.03160
The standard sparse-autoencoder (SAE) interpretability protocol labels each feature from its top-activating contexts and validates by single-feature steering. We propose the pairwise matrix protocol, co-varying steering coefficient with joint condition, and report three findings the standard one-corner protocol misses on Qwen3-1.7B-Instruct, replicated on Gemma-2-2B-it. First, a feature labelled "AI self-disclaimer" from its top contexts produces an inverted U-shape under a coefficient sweep: at