Researchers have developed an AI model named Centaur, claiming it can simulate the human mind by training on a data set called Psych-101, which aggregates data from 160 psychology experiments, encompassing over 60,000 participants’ decisions. Originally published in Nature, Centaur purportedly predicts human behavior in experiments articulated in natural language. Centaur, trained on Meta’s Llama language model, reportedly outperforms traditional cognitive models in mimicking human decision-making, even extending to modified tasks not included in its training data. This suggests potential for in silico experiment development and new behavioral theories. However, the claim of Centaur mimicking human cognition faces skepticism. Critics like Blake Richards and Marcel Binz argue it doesn’t genuinely reflect human cognitive processes. Jeffrey Bowers highlights its limitations, noting its ability to recall 256 digits and respond in milliseconds—feats beyond human capability, indicating a lack of faith in its generalizability. Federico Adolfi and others emphasize the dataset’s narrow scope compared to the vastness of human cognition. Despite criticisms of Centaur, some see value in the Psych-101 dataset for future model testing. Katherine Storrs and others remain cautiously optimistic about Centaur’s potential, acknowledging significant efforts invested that might yield scientific benefits eventually. To read more, click here. [Source: Science Magazine, July 2, 2025]
First seen: 2025-07-07 00:25
Last seen: 2025-07-07 04:25