Scientists Trained AI On Human Cognition

A team of international cognitive scientists has developed an AI system, Centaur, that mimics the human brain, including logical thought, decision-making, and behavioral psychology.

The Big Thought: Historically, psychological experiments have focused on a single aspect of research. The team behind Centaur aims to study how the entire brain functions to understand the connections between thoughts and actions on a broader scale.

Between the Gray Matter: Marcel Binz, a cognitive scientist at the German research center Helmholtz Munich, wrote about the innovations of Centaur in Nature.

• The “ChatGPT-like system” was trained on Meta’s Llama system using 160 experiments that included “over 10 million responses from more than 60,000 volunteers,” per NYT.
• The goal was to train the system to act as a volunteer in studies and respond and behave exactly as a human would.
• And that’s precisely what happened — it answered questions as humans did, strategized on games as they did, and even got stumped by the same questions as they did.
• Additionally, Centaur was able to predict, to some extent, how participants responded to other questions that it wasn’t trained on, based on what it learned about human psychology.

Final Idea: Unlike most cognitive studies, Dr. Binz admits that Centaur was built without testing a specific theory of the mind — he and his team just wanted to see if they could pull it off. The team plans to train Centaur on five times as many studies before doing another round of experiments.

Prediction: Centaur could be a paradigm shift for research of the mind, enabling researchers to validate their findings with an AI trained on a larger group of humans than they’ll likely ever have access to.