The case for shared hard tasks making AI's inner workings more brain-like
A post connects this argument to NeuroAI's “Contravariance Principle” and predicts stronger alignment as models train on more multimodal real-world data and learn to perform the same tasks as humans.
TLDR
A post argues that tackling the same difficult tasks as humans across domains pressures AI models' internal workings to align with brains. It connects that claim to NeuroAI's “Contravariance Principle” and predicts that alignment will increase as models train on more multimodal real-world data and learn to carry out the same tasks in those settings. A follow-up links the paper “Contravariance Theory: Strong Alignment for Minimal Solutions to Hard Tasks” for the mathematical details.