If LLMs Have Human-Like Attributes, Then So Does Age of Empires II

The paper's sharpest contribution isn't just skepticism about anthropomorphization in general; it's the specific argument that behavioral signatures used to infer understanding or moral reasoning in LLMs are substrate-agnostic, meaning they can emerge from any sufficiently complex system trained on sequential data, regardless of what that system was built to do.

This connects obliquely to the sparse autoencoder work covered the same day ('On the Relationship Between Activation Outliers and Feature Death in Sparse Autoencoders'). That paper is trying to make mechanistic interpretability tools more reliable by diagnosing why learned features fail to activate. Both papers are, at bottom, about the same underlying problem: our current methods for reading meaning into neural network internals and outputs are fragile and poorly calibrated. If SAEs misfire at initialization and behavioral signatures appear in Age of Empires bots, the interpretability toolkit researchers rely on to make safety claims is shakier than most public discourse acknowledges. This is largely disconnected from recent funding or product activity, sitting instead within a quieter methodological debate inside the research community.

Watch whether any of the major interpretability groups (Anthropic's or DeepMind's) respond by testing whether their specific SAE-derived feature claims survive the same substrate-agnosticism critique applied here. A non-response within the next two conference cycles would itself be informative.