Researchers Demonstrate 'Military Dog' Identity Confusion from Adding Feature Vectors

Interpretability researcher @Sauers showed that adding a 'catdog neuron' and a 'soldier neuron' produces a 'military dog' response, prompting @voooooogel to share similar 'identity arithmetic' behavior observed in earlier steering vectors research.

2026-10-24 ~ 2026-10-24 · 2 related posts